Let’s get into it.

Pricing American Options Beyond Heston

A fast spectral method for pricing American options now works across the stochastic-volatility models that actually fit equity markets, not just Heston.

Pricing American options under stochastic volatility is usually a slow finite-difference grind, and the fast Fourier-based shortcuts have historically only worked for the Heston model because it has a clean closed-form characteristic function. This paper extends the damped-sinc (DSINC) method to the stochastic-volatility models that empirical studies actually prefer, including the 3/2 model, SABR, GARCH diffusion, and lognormal volatility. The core finding is that the spectral collapse identity (the trick that reduces the early-exercise integral to one Fourier integral) still holds without the affine structure, provided damping sits in a well-defined strip.

For correlated SABR, a change of volatility coordinate avoids the usual coupling headache. Accuracy lands within one basis point of fine finite-difference benchmarks at comparable or lower cost. For practitioners, this means desks modeling realized variance, VIX products, or FX smiles can price American exercise using the models that fit the data, not just the one that fits the math.

Itkin, Andrey, DSINC Beyond Heston: The 3/2 and Other Non-affine Stochastic-Volatility Models (October 20, 2026). Available at SSRN: https://ssrn.com/abstract=7653118 or http://dx.doi.org/10.2139/ssrn.7653118

Turning SHAP Explanations Into Better Portfolios

Feature attribution in machine-learning portfolios can do more than explain, it can actually improve out-of-sample net Sharpe ratios without retraining the model.

Most machine-learning explainability tools tell you which features drove a model’s output, but that’s not the same as knowing which features actually improved your portfolio. This paper builds a framework that traces a feature’s influence all the way from allocation weights through turnover, volatility, and finally the net Sharpe ratio. The key insight is that a feature can be critical for setting positions yet destructive for performance once transaction costs enter the picture, and separating those channels matters.

Using a statistical arbitrage setup on U.S. equities from 2006 to 2024, the authors show that feeding attribution signals back into feature exposure (without retraining the model) lifts the FFN net Sharpe from 0.58 to 0.83, mostly by cutting turnover roughly in half. For XGB, averaging attributions across models raises the net Sharpe from 0.62 to 0.76 via better gross performance. For investors, this reframes explainability as a usable alpha-preservation tool rather than a diagnostic afterthought.

Liang, Yunqi and Zohren, Stefan and Cucuringu, Mihai, From Economic Explanation to Performance Improvement in Machine-Learning Portfolios (September 01, 2026). Available at SSRN: https://ssrn.com/abstract=7671079 or http://dx.doi.org/10.2139/ssrn.7671079

When Do Earnings Prediction Markets Actually Beat the Analysts?

Earnings prediction markets add forecasting value in combination with analyst estimates about five days before announcements, but only beat them outright in the final day.

Polymarket’s earnings contracts, which pay a dollar if a firm beats consensus EPS, have been hailed as more accurate than Wall Street analysts. Zhang pushes back on that framing by asking a sharper question: when do these prices actually add information beyond what we already get from a firm’s beat history and the analyst consensus? Evaluated prospectively (fitting weights in one earnings season, testing in the next), contract prices start adding value in combination with existing forecasts about five days before the announcement, but only outperform those benchmarks on their own in the final day of trading.

The market’s edge seems to come from incorporating public news (like peer firms’ earnings surprises) that stale analyst estimates have not yet absorbed, not from a hidden group of informed traders. For investors, the practical takeaway is that these contracts complement traditional forecasts rather than replace them, and the stock market has largely priced in their signal before the bell rings.

Zhang, Tengfei, What do Earnings Prediction Markets Know? (October 24, 2026). Available at SSRN: https://ssrn.com/abstract=7667338 or http://dx.doi.org/10.2139/ssrn.7667338

Where Are the Dealers? Inside China’s Corporate Bond Market

China’s $5 trillion corporate bond market runs without the dealer core that defines Western OTC markets, and the reasons trace back to bank incentives.

Using a Chinese counterpart to TRACE, the authors deliver the first trade-level look at the world’s second-largest corporate bond market, and what they find breaks the standard OTC playbook. There is no dealer-dominated core-periphery network. Designated dealers (mostly banks) supply only about a quarter of liquidity, and customers trade directly with each other for most of the market’s volume.

When customers step in as backup liquidity providers, there is a trade-off: long-term holders like insurers provide persistent but expensive liquidity, while trading-oriented players offer tighter execution but disappear quickly. Two frictions explain the dealer gap. Smaller banks are hemmed in by capital constraints, while the Big Six state-owned banks concentrate on central SOE bonds and pull back from market making when lending margins rise, because safe SOE loans are simply more profitable. For investors, this means Chinese bond liquidity is thinner, costlier, and more fragile than headline market size suggests.

Du, Linlin and Hu, Yue and Song, Yang and Wu, Wenfeng, China’s Corporate Bond Market: A Transaction-level Analysis (September 01, 2025). Available at SSRN: https://ssrn.com/abstract=5963114 or http://dx.doi.org/10.2139/ssrn.5963114

This week for paid subscribers

Paid subscribers are getting a full deep dive of the covered-call options strategy, reproduced on its original historical window and extended through August 2025 on the authors’ exact specification, against survival thresholds fixed before out-of-sample tests were observed. The post traces how the classic volatility risk premium leg in S&P 500 and Nasdaq-100 ran hot into the end of the paper sample and turned into a post-paper point estimate near zero, and tests whether duration approximation noise, execution drag, or a structural break in volatility spikes is actually behind it. Python notebook and data included.

Disclaimer: The content provided in this newsletter, “Alpha in Academia,” is for informational and educational purposes only. It should not be construed as financial advice, investment recommendations, or an offer or solicitation to buy or sell any securities or financial instruments. Past performance is not indicative of future results. The financial markets involve risks, and readers should conduct their own research and consult with qualified financial advisors before making any investment decisions.

The interpretations, opinions, and analyses presented herein are those of the author and do not necessarily reflect the views of the original researchers, their institutions, or the full implications of the cited academic papers. While every effort is made to accurately represent the research discussed, readers should be aware that the summaries and interpretations may not capture the full scope or nuances of the original studies. The information contained in this newsletter is believed to be accurate and reliable at the time of publication, but accuracy and completeness cannot be guaranteed. The author and publisher accept no liability for any loss or damage resulting from reliance on the information provided.

This newsletter may contain links to external websites or resources. The author is not responsible for the content, accuracy, or reliability of these external sources.

By subscribing to or reading this newsletter, you acknowledge that you have read and understood this disclaimer and agree to hold the author and publisher harmless from any liability that may arise from your use of the information contained herein.

Today we are looking at covered calls, the strategy behind the income ETFs that have become a staple of retail portfolios. Israelov and Nielsen showed in 2015 that a covered call is really three trades: a position in the index, a bet that the market will move less than option prices imply, and a timing bet that comes attached to the option. Rebuilding their split on thirty years of option data, the timing bet carries a third of the risk and is not paid, and the volatility leg, the one part sellers are actually compensated for, has mostly stopped paying since the paper was published.

Let’s dive right in.

Introduction

The covered call is the first option strategy most people learn. You own the index, sell a call against it every month, and keep the premium. The usual pitch is that you trade some upside for income and a smoother ride. Funds like XYLD and QYLD package exactly this: each month they sell an at-the-money one-month call on the S&P 500 or the Nasdaq-100 and pay most of the premium out as a distribution. Yields of 7% to 12% a year are the whole appeal.

The standard way to think about the trade is as one position with a capped upside. Israelov and Nielsen made the case in the Financial Analysts Journal in 2015 that this hides three separate exposures.

The first is passive equity. On average you hold about half the index, because a short at-the-money call offsets about half of your stock.

The second is short volatility. Strip out the call’s stock exposure and what is left is a bet that the index will move less than option prices imply. Option sellers have historically been paid for taking that bet, which is the volatility risk premium.

The third is equity timing, and it is the one nobody chose. A call’s delta moves with the market. After a rally the call you sold is deep in the money, it moves almost one for one with the index, and your net stock exposure shrinks toward zero. After a sell-off the call is nearly worthless and you are back to holding almost the full index. The position mechanically sells after gains and buys after losses, which is a bet that moves reverse.

Israelov and Nielsen found that the short volatility leg earned a Sharpe ratio of nearly 1.0 while contributing about 10% of the risk, and that the timing leg contributed about 25% of the risk for very little return. Israelov, Klein and Tummala later ran the same split on eleven global indexes and found the timing returns statistically insignificant in every one.

The original sample ends in 2014. Since then covered-call ETFs have gone from a niche to a fixture of income portfolios. So the questions here are simple: Does the split still hold on another decade of data? And what does it say about the funds people actually hold?

Data and Methodology

Option prices and Greeks come from OptionMetrics via WRDS: daily quotes, implied volatilities and deltas for S&P 500 (SPX) and Nasdaq-100 (NDX) index options, January 1996 to 29 August 2025. The Nasdaq-100 sample starts in 2001, because before then OptionMetrics has no index opens for NDX and the option chain was thin enough that the nearest strike above spot was sometimes 8% out of the money. Dividends come from the S&P 500 total return index and, for the Nasdaq-100, from QQQ’s distributions. The cash rate is the 13-week T-bill yield. Cboe’s published BXM and BXN histories serve as the benchmark, and fund prices are from Yahoo.

The construction follows Cboe’s buy-write rules. On each monthly expiration (the third Friday), sell the next month’s call at the smallest listed strike at or above the index, and hold it to expiry. That gives 355 monthly rolls on the S&P 500 and 295 on the Nasdaq-100, with the written strike a median 0.15% above spot.

Three construction choices are worth stating plainly:

1. Cboe writes the new call at the average traded price between 11:00 and 11:30 on roll day, and settles the old one at a special opening quotation. I only have end-of-day data, so I write at the close and settle at the index open.

2. The call is sold at the bid and marked at the mid afterwards. The gap between the two is booked as its own line, execution, so that the volatility leg measures the premium rather than the cost of collecting it. Checking the rebuild against Cboe shows why the bid is the right assumption.

3. OptionMetrics leaves out deltas when it cannot fit an implied volatility, which happens almost entirely on the last trading day before expiry. On the S&P 500 that is 514 of 7,442 days. I fill them with a Black-Scholes delta from the call’s last good implied volatility; on days where both exist, the two differ by a median of 0.012.

The decomposition itself is an accounting identity. Each day, with S the index, C the call, Δ its delta and dS the index’s total-return move in points:

The equity weight w is the share of the portfolio effectively invested in the index. Its average is 47% on both indexes. Passive equity is that average weight times the index return; timing is the daily deviation from it. Everything is measured in excess of T-bills, and the pieces add back to the total to the sixteenth decimal place.

Before splitting anything, the rebuild has to match the indexes it claims to be. Monthly returns correlate 0.98 with Cboe’s BXM and 0.99 with BXN. Writing at the bid rather than the mid is what makes the levels line up: BXN then runs within 0.03 percentage points a year of Cboe after 2012, and BXM within 0.47. Regressing each roll’s remaining gap on its likely causes says why. The coefficient on half the bid-ask spread is 3.1 (t=2.09), so the trades Cboe records between 11:00 and 11:30 print well below the closing mid. Settling at the open rather than at Cboe’s opening quotation adds a small upward bias when the old call expires in the money (t=3.59). Neither touches the daily deltas or marks the decomposition is built from, so I take the rebuild as it stands and disclose the gap.

Figure 1: The rebuild against Cboe’s published indexes, written at the bid. The shapes are the same index; the remaining BXM level gap is concentrated before 2012, when option spreads were several times wider than they are today.

Read more

Let’s get into it.

Lever Up Your IRA, Says the Retirement Model

A leveraged ETF held for years doesn’t decay the way its warning label says, and a standard retirement model actually wants one in your IRA.

A 2x leveraged ETF promises double the S&P 500’s move every day, and since 2009 regulators have warned that the daily reset makes it rot over long holds, the famous “volatility drag.” Lai and Reilly model nearly every US equity leveraged ETF and find that volatility alone doesn’t lower the average return at all. Two things do cost you. The first is rebalancing: To stay at 2x, the fund adds exposure after up days and trims it after down days, and because indexes tend to partly reverse the next day, it ends up buying high and selling low. The second is rent on the leverage, which funds get from banks through swaps at about 1.3 percentage points over the Fed’s overnight rate.

Figure 1: The optimal retirement portfolio in the authors’ model. The household holds only a 2x S&P 500 fund (dark) until its early thirties, then gradually adds international stocks (light), but never drops below roughly a third in the leveraged fund, even after retiring at 65 (dashed line). Bonds and bills never make the cut.

The myth survives because leverage makes results lopsided, a few huge years and many losing ones, so anyone eyeballing past charts mostly sees losers. Put a 2x S&P fund into a standard retirement model and the optimal household holds it for life, and holds nothing else while young. So the product built for day traders suits the most boring investor imaginable.

The math seems reasonable, but the fragile part is that the model assumes you never panic sell in a crash. If you already hold only stocks in an IRA, where margin is banned, a slice of a 2x index fund is defensible. Judge it by its swap costs and how closely it tracks its target, not by its volatility tables. If you still hold bonds, raise your stock share first.

Reilly, Christopher and Lai, Qihao, Lever Up Your IRA: Understanding the Long Run Performance of Leveraged ETFs (October 19, 2026). Available at SSRN: https://ssrn.com/abstract=7638818 or http://dx.doi.org/10.2139/ssrn.7638818

Funds That Inflate Themselves

When a fund that owns thinly traded stocks grows fast, its own buying pumps up its returns, investors chase those returns with more money, and the loop tends to end badly a few years later.

If you buy a lot of a stock that barely trades, its price goes up. Now picture a fund holding a basket of thinly traded names that gets new money every day. It invests that money in its existing proportions, so it keeps buying the same small stocks, which lifts their prices, which shows up as a nice return. Investors see the return and send more money, in effect applauding their own purchases.

Figure 2: Each day, ETFs are sorted by how much new money they took in. For funds concentrated in thinly traded stocks, returns rise in lockstep with inflows. For funds holding easily traded stocks, new money barely moves returns. Returns are measured relative to the overall market.

The authors measure this with “fund illiquidity,” roughly how much of its holdings’ daily trading volume a fund has to buy when new money arrives, beyond what its most liquid possible portfolio would require.

What was surprising is that nobody leans against it. Managers barely shift toward easier stocks as they grow (bigger funds mean bigger fees). Investors chase this price pressure as eagerly as real stock picking. And shorting illiquid small caps hurts when the crash takes years to arrive. It does arrive. Stocks where these inflated funds owned over 5% of shares lost about 30% over the next three years, which the authors call “a dynamic akin to fire sales, but in the positive direction.”

The long run evidence rests on a fairly small number of mutual fund episodes, so treat the size of that drop as rough. A broad S&P 500 fund is untouched by this. If you hold a thematic ETF that ballooned after a hot year, check how big a slice of each holding it owns before reading its record as skill.

van der Beck, Philippe and Bouchaud, Jean-Philippe and Villamaina, Dario, Self-Inflated Funds (May 20, 2024). Harvard Business School Finance Working Paper No. 26-066, Harvard Business School Working Paper No. 26-066, Available at SSRN: https://ssrn.com/abstract=4834390 or http://dx.doi.org/10.2139/ssrn.4834390

Rating Downgrades and the Funds That Buy Them

When a bond gets downgraded, actively managed bond funds tend to buy it, and the ones they pick usually recover.

When a rating agency cuts a company’s credit grade, lots of bondholders sell, and most of them haven’t learned anything new. Insurers face capital rules tied to ratings, especially when a bond drops from investment grade to junk, and retail investors tend to treat the rating itself as the verdict. Active bond funds face neither constraint, so they take the other side. They buy more when the downgrade was predictable and when other signals point up.

Figure 3: Downgraded bonds lose roughly 7% over the year before the cut. Afterward, the ones active funds bought recover faster than the ones they passed on. Prices are shown relative to four quarters before the downgrade.

The bonds they buy beat the downgraded bonds they skip by about 0.84 percentage points over the next year, roughly a typical active bond fund’s annual fee, and they default less often. The authors call these funds “informed liquidity providers that process forward-looking credit information.”

The gap survives because corporate bonds trade rarely, and dealers have had less room to hold them since the Volcker Rule, which also shrank the funds’ buying. The catch is that the data are quarterly snapshots, so nobody sees the prices funds actually paid or whether the edge survives trading costs. If you own a bond index ETF, you are on the selling side whenever a bond falls out of the index, which makes this one of the few places an active bond fund’s fee looks earned.

Andreani, Martina and Juliani, Elsa Maria and Vasvari, Florin P., Corporate Bond Ownership and Rating Downgrades (September 28, 2026). Available at SSRN: https://ssrn.com/abstract=7609643 or http://dx.doi.org/10.2139/ssrn.7609643

Digitalizing a State Monopoly, Repricing County Debt

When Pennsylvania legalized online lottery sales, bond investors started charging the counties most dependent on lottery-funded senior programs a little less to borrow, months before anyone bought a ticket online.

Pennsylvania’s lottery profits pay for programs for older residents, and some counties rely on those programs far more than others relative to their own budgets. When the state let its lottery sell online, did that create new gambling or just move it off convenience store counters? Relative to New York, which stayed offline, total lottery play rose about 14 percent, and store sales rose too.

The bond story is subtler. Counties borrow by issuing municipal bonds, and the spread is the extra yield investors demand over Treasuries for the risk of lending to that county. A bigger state lottery pot lowers the odds that senior programs get cut in a lean year and pushed onto county budgets, so investors in the most dependent counties trimmed spreads by about 5 basis points when the law was signed, then did nothing at launch.

Figure 4: How much more cheaply Pennsylvania’s most lottery‑dependent counties could borrow, relative to less dependent counties, over time. Below zero means investors demanded less extra yield. The drop comes when the online lottery law was signed in October 2017, not when online sales began in May 2018.

Nobody arbitrages this away, because county bonds rarely trade and the move is small. The sample is a few dozen counties in one state, and the extra program money never visibly showed up, so I treat this as suggestive. If you hold a muni bond fund, there is nothing to trade here, but it is a clean example of prices moving when laws pass, not when money arrives.

Soylemezgil, Sevgi, Digitalizing a State Monopoly: Lottery Demand and Municipal Finance (May 10, 2026). Available at SSRN: https://ssrn.com/abstract=7657378 or http://dx.doi.org/10.2139/ssrn.7657378

Disclaimer: The content provided in this newsletter, "Alpha in Academia," is for informational and educational purposes only. It should not be construed as financial advice, investment recommendations, or an offer or solicitation to buy or sell any securities or financial instruments. Past performance is not indicative of future results. The financial markets involve risks, and readers should conduct their own research and consult with qualified financial advisors before making any investment decisions.

The interpretations, opinions, and analyses presented herein are those of the author and do not necessarily reflect the views of the original researchers, their institutions, or the full implications of the cited academic papers. While every effort is made to accurately represent the research discussed, readers should be aware that the summaries and interpretations may not capture the full scope or nuances of the original studies. The information contained in this newsletter is believed to be accurate and reliable at the time of publication, but accuracy and completeness cannot be guaranteed. The author and publisher accept no liability for any loss or damage resulting from reliance on the information provided.

This newsletter may contain links to external websites or resources. The author is not responsible for the content, accuracy, or reliability of these external sources.

By subscribing to or reading this newsletter, you acknowledge that you have read and understood this disclaimer and agree to hold the author and publisher harmless from any liability that may arise from your use of the information contained herein.

Let’s get into it.

Rank-and-File Workers Are Hidden Drivers of Reporting Quality

Corporate reporting quality depends less on executives and more on the rank-and-file workers who produce, verify, and respond to it.

This review reframes accounting research by tracing how non-executive employees shape, and are shaped by, corporate disclosure across a four-stage Labor Life Cycle (education, job search, employment, and turnover). The authors synthesize over 130 studies showing that workforce quality is a first-order driver of reporting outcomes. Firms paying above-market wages to accountants and auditors produce fewer restatements and stronger internal controls, consistent with efficiency wage theory. Stricter CPA licensing (the 150-hour rule) shrinks the talent pipeline, especially among minority candidates, without measurably improving quality, pointing to supply restriction rather than skill-building. On the user side, earnings announcements trigger job searches by incumbent employees, pay ratio disclosures shift satisfaction and turnover, and Glassdoor reviews predict stock returns.

For investors, the takeaway is that human capital metrics (turnover rates, staffing levels, pay dispersion) carry signal about future reporting reliability and firm performance, yet remain poorly standardized in current disclosures.

Barrios, John Manuel and Choi, Jung Ho and Deller, Carolyn and Pacelli, Joseph and Packard, Heidi A., Labor and the Corporate Information Environment: A Review. (June 18, 2026). Harvard Business School Working Paper No. 26-060, Harvard Business School Accounting & Management Unit Working Paper No. 26-060, The Wharton School Research Paper , HKU Jockey Club Enterprise Sustainability Global Research Institute Paper No. 2026/037, Available at SSRN: https://ssrn.com/abstract=6390718 or http://dx.doi.org/10.2139/ssrn.6390718

A Faster Way to Price American Commodity Options

An explicit finite difference method called ADE prices American commodity options under the Schwartz two-factor model hundreds of times faster than Least Squares Monte Carlo at matched accuracy.

Pricing American options on commodities is harder than on stocks because you need a second state variable (the convenience yield, the implicit benefit of holding the physical commodity) and because the holder can exercise any day. The standard workhorse, Longstaff-Schwartz Monte Carlo, handles this but is slow. The author implements the Alternating Direction Explicit (ADE) scheme, which keeps the simplicity of an explicit method while remaining stable, on the Schwartz two-factor PDE for WTI crude. Validated against the closed-form European price and benchmarked against LSMC with 250,000 paths, the ADE solver matches the Monte Carlo price to within 0.03% on a six-month at-the-money put, and reaches a given accuracy roughly 200 to 400 times faster.

Extensions include exercising on a futures price (how listed ag options actually trade), seasonal parameters, and two-factor Greeks that produce a hedge in two futures maturities. For anyone pricing or risk-managing commodity options, the practical read is that PDE methods are not obsolete, they can be meaningfully cheaper than simulation when the dimensionality stays small.

Prabhu, Sanjith and Duffy, Daniel, Pricing American Commodity Options Under the Schwartz Two-Factor Model Using the Alternating Direction Explicit (ADE) Method (October 10, 2026). Available at SSRN: https://ssrn.com/abstract=7604358 or http://dx.doi.org/10.2139/ssrn.7604358

Bond Indices Quietly Tax Their Own Trackers Every Month-End

Government bond futures earn roughly seventeen times their normal daily return on the last trading day of the month, driven entirely by mechanical index rebalancing.

Fixed-income indices rebalance at month-end, and because bonds age while new issuance tends to come at longer maturities, the index duration systematically extends by about 0.08 years each month. Benchmark-tracking funds (ETFs, pensions, insurers) must buy that duration near the index pricing snapshot to avoid tracking error, and this synchronized, price-insensitive demand pushes prices up. The authors find duration-scaled futures returns average 1.19 basis points on month-end days versus 0.07 on other days, with a single day accounting for 47% of average monthly returns. The effect peaks right before the Bloomberg snapshot and partially reverses after, a textbook signature of non-informational price pressure.

When Bloomberg shifted its US snapshot time by one hour in 2021, trading volume and returns migrated with it, while Germany showed no shift. For benchmark-constrained managers this is a recurring hidden cost; for anyone willing to supply liquidity into the snapshot, it is a predictable, non-risk-based source of return.

Soebhag, Amar and Verwijmeren, Patrick and Martinetti, Fabio, Buying Duration: Index Rebalancing and Price Pressure in Government Bond Markets (October 12, 2026). Available at SSRN: https://ssrn.com/abstract=7601879 or http://dx.doi.org/10.2139/ssrn.7601879

Gold Still Beats Silver When You Score Them Side by Side

Across a battery of financial tests from 1990 to 2024, gold beats silver as the better investment, mainly because silver is far more volatile for only a modest return premium.

The author runs both metals through five tests (macro integration, predictability, loss rate, volatility, and returns across short, medium, and long horizons) and scores them with a Pugh decision matrix. Gold shows strong long-run integration with the US stock market (NASDAQ, S&P 500, Dow Jones) and oil, qualifying it as a genuine safe haven, while silver only tracks oil meaningfully. Both metals post losses in roughly 45% of months, but silver’s drawdowns are deeper, with return variance running about three times gold’s.

Silver does edge gold on medium and long-term annualized returns (12.2% and 8.7% versus 10.5% and 6.2%), but gold wins the short term and carries far less risk. The practical read for investors: gold works as the defensive long position when markets wobble, while silver behaves more like a speculative, higher-beta cousin whose extra return rarely compensates for the ride.

Vacio-Hernández, Iván Alejandro, Gold or Silver: An Alternative to Investing (March 12, 2025). Apuntes del Cenes, volume 44, issue 79, 2025[10.19053/uptc.01203053.v44.n79.2025.17833], Available at SSRN: https://ssrn.com/abstract=7620858

Volatility Scaling Tames Momentum Crashes, But Not Everywhere Anymore

Volatility-managed momentum dramatically cuts crash risk in developed markets, but the edge has quietly faded in emerging markets since 2010.

The authors revisit the classic Daniel-Moskowitz momentum crash story using international Fama-French factor data through April 2026. The pattern still shows up, past losers rebound violently after prolonged market weakness, crushing the short leg of a winners-minus-losers portfolio, but it is sharpest in developed markets and in the pre-2008 sample. Applying a strictly ex-ante volatility scaling rule (position size inversely proportional to recent realized variance) lifts the developed-market Sharpe ratio from 0.53 to 0.87 and shrinks the worst drawdown from -43.6% to -10.7%. Emerging markets see a full-sample Sharpe lift from 0.98 to 1.29.

Figure 2. Drawdowns of static developed and emerging momentum portfolios.

The important caveat: after 2010, static emerging momentum actually posts a higher Sharpe (1.36) than the managed version (1.29), suggesting the overlay works best as drawdown insurance rather than a return booster. For practitioners, the takeaway is to treat volatility scaling as risk governance, instead of a universal alpha engine, and to recalibrate as market structure evolves.

Feng, Xuan and Huang, Linghan and Su, Yu, Revisiting Momentum Crashes: Volatility Timing across Developed and Emerging Markets (July 01, 2026). Forthcoming in Applied Economics, Available at SSRN: https://ssrn.com/abstract=7612298 or http://dx.doi.org/10.2139/ssrn.7612298

This week for paid subscribers

Paid subscribers are getting a full deep dive of the corporate bond credit return anomaly, reproduced on its original historical window and extended through September 2026 on the authors’ exact specification, against survival thresholds fixed before out-of-sample tests were observed. The post traces how the classic first-five-days return concentration in high yield ran hot into the end of the paper sample and turned into a post-paper point estimate of zero, and tests whether duration approximation noise, regime shifts, or a structural break is actually behind it. Python notebook and data included.

Disclaimer: The content provided in this newsletter, “Alpha in Academia,” is for informational and educational purposes only. It should not be construed as financial advice, investment recommendations, or an offer or solicitation to buy or sell any securities or financial instruments. Past performance is not indicative of future results. The financial markets involve risks, and readers should conduct their own research and consult with qualified financial advisors before making any investment decisions.

The interpretations, opinions, and analyses presented herein are those of the author and do not necessarily reflect the views of the original researchers, their institutions, or the full implications of the cited academic papers. While every effort is made to accurately represent the research discussed, readers should be aware that the summaries and interpretations may not capture the full scope or nuances of the original studies. The information contained in this newsletter is believed to be accurate and reliable at the time of publication, but accuracy and completeness cannot be guaranteed. The author and publisher accept no liability for any loss or damage resulting from reliance on the information provided.

This newsletter may contain links to external websites or resources. The author is not responsible for the content, accuracy, or reliability of these external sources.

By subscribing to or reading this newsletter, you acknowledge that you have read and understood this disclaimer and agree to hold the author and publisher harmless from any liability that may arise from your use of the information contained herein.

Dickerson and Nozawa report that most of the credit return in US corporate bonds is earned in the first five trading days of each calendar month, with the rest of the month averaging close to zero. The claim is strong, the sample ends in January 2023, and the authors state that the mechanism behind it remains an open puzzle. This post reproduces their result from scratch, pushes it forward through September 2026, and reports what the extension can and cannot tell us.

Let’s dive right in.

What the paper claims

The paper documents that corporate bond credit returns are concentrated in a narrow window at the start of each month. Over the sample period from February 1997 to January 2023, the first five trading days of each month account for 73 percent of individual bonds’ credit returns, 83 percent of the US corporate bond market credit premium, and nearly 80 percent of the monthly credit premium in high-yield bonds. The average corporate bond earns 11.8 of its 16.1 basis points of monthly credit return in that window, with the balance of the month contributing a statistically insignificant remainder.

The authors measure credit returns as corporate bond returns in excess of duration-matched Treasury returns, following van Binsbergen, Nozawa, and Schwert (2025). The pattern is specific to that measurement: using corporate bond returns in excess of T-bill rates, the first-five-day premium is negative and insignificant. The premium arises because Treasury returns tend to underperform corporate bonds during the first five days of the month.

What makes the paper notable is less the pattern itself than the mechanism search that follows it. The authors test eleven candidate explanations drawn from the existing calendar-effects literature: month-end cash needs, passive reinvestment, customer buying pressure, index reconstitution, primary-market issuance, the coupon payment cycle, the turn-of-the-quarter effect, scheduled macroeconomic news announcements, the FOMC cycle, prior-month stock returns, and prior-month VIX. None fully explains the premium at the individual-bond, market-portfolio, and factor levels simultaneously. Customer buying pressure and prior-month VIX account for roughly half of the bond-level premium with wide confidence intervals; everything else fails. The paper closes the mechanism section by stating that why compensation for credit risk follows this schedule remains an open puzzle.

Their proposed framing is instead a bond-market version of the Savor and Wilson (2014) macroeconomic-announcement story: the bond CAPM explains the cross-section of average factor returns well, but only over the first five days of the month. The first-five-day market credit premium prices a factor zoo of 103 long-short bond portfolios with a cross-sectional R-squared of 76 percent, versus 6 percent using full-month returns. In that reading, there is one anomaly in the market portfolio that lives in a specific calendar window, and the rest of the factor zoo inherits it.

Read more

Let’s get into it.

The Activeness Hiding in Passive Investing

Japanese index funds built on niche themes charge higher fees, attract investors who stop paying attention to those fees, and then go on to underperform.

Japan has moved into passive investing faster than the US. As plain market trackers raced toward zero fees, fund managers found a way out: launch index funds on AI, robotics, semiconductors and other catchy themes. Looking at two decades of Japanese index funds and ETFs, Anan finds that many of these “passive” products deviate from the market about as much as a typical stock picker’s fund.

Figure 1: Niche index funds (teal) deviate from the market about as much as actively managed funds (pink), and in the high-activeness tail they often deviate more. Chart recreated from Figure 2 of Anan, “Activeness in Passive Investing: Evidence from Japanese Index Funds.”

That activeness comes with higher fees. The more surprising part is investor behavior. People buying these funds are noticeably less sensitive to fees and chase recent performance harder. It sadly doesn’t pay off though. Among niche funds, a typical step up in activeness costs roughly 1 to 1.5 percentage points of risk adjusted return a year, even when comparing funds within the same category, and the gap has grown over the last decade. The author warns that “niche index funds may function as a structural trap.” And this lesson applies well beyond Tokyo. Calling something an index fund doesn’t make it cheap, diversified or passive. What you’re really buying is the index designer’s bet.

Anan, Anju, Activeness in Passive Investing: Evidence from Japanese Index Funds. Available at SSRN: https://ssrn.com/abstract=7578303 or http://dx.doi.org/10.2139/ssrn.7578303

Graduated Trading Restrictions and the Expiry Day Rebound

Borsa Istanbul’s toughest volatility curbs don’t calm overheated stocks so much as pause them, and the trading comes back on the day the restrictions expire.

When a small Turkish stock rockets on a wave of retail buying, Borsa Istanbul can hit it with escalating restrictions. These start with bans on short selling and margin buying, move up to a rule that forces every trade to settle in full (so shares bought today can’t be sold the same day), and end with periodic auctions in place of continuous trading. The first step barely registers. The settlement rule does the real work, wiping out roughly two thirds of trading value while each lira that still trades moves the price far more.

The paper’s method for proving this is clever. Every measure comes with a preset end date, and when the exchange lengthened those windows from two weeks to about a month, the rebound in trading shifted with the calendar and landed right on expiry day.

Figure 2: Trading in stocks placed under gross settlement, relative to similar unrestricted stocks. Activity drops the day the rule starts and recovers on its scheduled end date, about 11 trading days later under the old two week rule and about 22 under the newer one month rule. Chart redrawn from Fig. 3a of Apaydın, “Graduated trading restrictions and market stability.”

Apaydın believes that the system “mainly postpones trading in the restricted stocks.” Flagged names do give back part of their run up in the following months, though that may just be ordinary mean reversion after a big rally. For anyone chasing momentum in emerging markets, a regulatory flag works like a liquidity freeze with a known end date. It says little about whether the stock is fairly priced.

Apaydın, Can, Graduated trading restrictions and market stability: Evidence from Borsa Istanbul's volatility-based measure system. Available at SSRN: https://ssrn.com/abstract=7577947 or http://dx.doi.org/10.2139/ssrn.7577947

When Hedging Changes the Payoff

When a large trader hedges an option, the hedge itself moves the price that decides what the option pays.

Classic option pricing assumes the hedger is invisible, buying and selling stock without nudging its price. Researchers at LSE and Oxford drop that assumption and find the problem becomes circular. A dealer hedging a call buys shares, which lifts the price, which raises the call’s payoff, which requires an even bigger hedge. They show this moving target can still be hedged perfectly for calls and puts, but always at a higher cost than the textbook price.

In effect, the strike drifts lower by an amount tied to the size of the hedge, so dealers end up holding more stock than standard delta (the usual hedge ratio) would suggest. Size matters a lot. As the authors note, “price impact introduces a nonlinear dependence on the number of contracts,” so a bigger position raises the cost of each contract, not just the total.

Figure 3: Call option price per contract around a $100 stock. The solid blue line is the classic price with no market impact. Each line above it adds friction, and hedging 10 contracts in a high impact market (dotted green) costs the most per contract.

In their simulations, about 1 in 100 options expired in the money only because of the hedger’s own buying. Same day expiry options now make up most S&P 500 index option volume, so this effect is practical. Large dealer hedges can move closing prices, and the option prices you pay already include that feedback.

Itkin, David and Sánchez-Betancourt, Leandro, When Hedging Changes the Payoff: Option Replication with Price Impact and Execution Costs (September 30, 2026). Available at SSRN: https://ssrn.com/abstract=7555822 or http://dx.doi.org/10.2139/ssrn.7555822

Agents Propose, Code Decides, the Market Disposes

AI agents can now run an entire stock factor search on their own, and the honest result is a model that kept its ranking skill on new data but lost most of its profit.

Sadeghi handed one of quant finance’s most error prone chores to a team of Claude agents: Test every published stock predictor and keep only the ones that add something new. The twist is that the agents decide what to test, while fixed code decides how each test is scored. The rules, the testing order and even the expectations for the final exam were all written down before any results existed. In three days the loop worked through all 212 predictors in a public catalogue, kept a small fraction, and sharply improved the model’s ability to rank stocks on historical data.

Figure 4: Average monthly return for each tenth of stocks, ranked from worst (D1) to best (D10) by the AI-built model. In the historical data, returns climb steadily with the ranking. On new data, the model still flags the stocks to avoid, but the other nine groups earn almost the same.

On the untouched recent data, that ranking skill mostly held up, but the PnL could not keep up. The strategy earned roughly half its expected risk adjusted return, and much of it came from a market hedge riding a rising market rather than from stock picking. For investors, the lesson is that ranking stocks well and profiting from it are different things, and writing expectations down first “makes a mixed result reportable rather than a disappointment to be explained away.”

Sadeghi, Erfan, Agents Propose, Code Decides: An Autonomous Research Loop Framework to Build an Equity Alpha Model (October 05, 2026). Available at SSRN: https://ssrn.com/abstract=7568638 or http://dx.doi.org/10.2139/ssrn.7568638

Disclaimer: The content provided in this newsletter, "Alpha in Academia," is for informational and educational purposes only. It should not be construed as financial advice, investment recommendations, or an offer or solicitation to buy or sell any securities or financial instruments. Past performance is not indicative of future results. The financial markets involve risks, and readers should conduct their own research and consult with qualified financial advisors before making any investment decisions.

The interpretations, opinions, and analyses presented herein are those of the author and do not necessarily reflect the views of the original researchers, their institutions, or the full implications of the cited academic papers. While every effort is made to accurately represent the research discussed, readers should be aware that the summaries and interpretations may not capture the full scope or nuances of the original studies. The information contained in this newsletter is believed to be accurate and reliable at the time of publication, but accuracy and completeness cannot be guaranteed. The author and publisher accept no liability for any loss or damage resulting from reliance on the information provided.

This newsletter may contain links to external websites or resources. The author is not responsible for the content, accuracy, or reliability of these external sources.

By subscribing to or reading this newsletter, you acknowledge that you have read and understood this disclaimer and agree to hold the author and publisher harmless from any liability that may arise from your use of the information contained herein.

Let’s get into it.

The Workers Behind the Numbers

Accounting’s information environment is shaped as much by rank-and-file employees as by the executives who sign the filings.

This review pulls together thirty years of research arguing that ordinary workers sit on both sides of corporate disclosure: they produce the numbers (as accountants, auditors, and operational staff whose skill and job stability determine reporting quality) and they consume them (as job seekers, negotiators, and insiders deciding whether to stay or leave). The authors organize the evidence around a Labor Life Cycle with four stages, from human capital development through turnover, and show that information frictions bite at each one.

One concrete thread: in states that adopted the inevitable disclosure doctrine, which restricts where employees can jump to, firms reported less upward earnings management, because workers with fewer outside options pushed back less on aggressive accounting. The broader claim is a feedback loop between disclosure and the labor market that remains thinly tested empirically. For investors, the practical takeaway is that workforce metrics, retention, mandated human capital disclosures, and even Glassdoor data are not soft ESG filler but inputs into reporting quality itself, and worth reading alongside the financial statements rather than after them.

Barrios, John Manuel and Choi, Jung Ho and Deller, Carolyn and Pacelli, Joseph and Packard, Heidi A., Labor and the Corporate Information Environment: A Review. (June 18, 2026). Harvard Business School Working Paper No. 26-060, Harvard Business School Accounting & Management Unit Working Paper No. 26-060, The Wharton School Research Paper , HKU Jockey Club Enterprise Sustainability Global Research Institute Paper No. 2026/037, Available at SSRN: https://ssrn.com/abstract=6390718 or http://dx.doi.org/10.2139/ssrn.6390718

The Hidden Cost of Hard Regime Switches

Switching regime models on and off creates estimator jumps that posterior averaging almost entirely removes, at little cost to accuracy.

Regime-aware risk models typically pick one market state (calm, turbulent, expansion, contraction) and estimate volatility, correlations, or tail risk only from historical weeks that match. The problem is that a small shift in the classifier’s belief can flip the selected state and swap the entire estimation sample overnight, which the author calls sample gating. On transition weeks, covariance estimates jump 15 to 30 times their usual size, forcing portfolio weights to lurch and triggering risk-review flags 5 to 6 times a year.

The fix is simple: instead of committing to one state, average the state-specific estimates using the classifier’s probabilities. Across S&P 500, EuroStoxx 50, and US sector ETF panels, this cuts transition jumps by 72 to 81 percent, lowers minimum-variance portfolio turnover by 23 to 30 percent, and improves risk-forecast loss in all nine panel-loss comparisons tested. For anyone running regime-based allocation or risk models, the takeaway is that the discrete on-off switch is doing real damage, and the smoother alternative is already sitting in the classifier’s own output.

Hydari, Syed Bashir, Sample Gating: Estimator Discontinuity under Latent-State Conditioning (September 28, 2026). Available at SSRN: https://ssrn.com/abstract=7538119 or http://dx.doi.org/10.2139/ssrn.7538119

The Credit Premium Shows Up for Work Five Days a Month

The entire credit risk premium in corporate bonds is earned in the first five trading days of each month.

Dickerson and Nozawa document something strange about how corporate bonds pay investors. Using every US corporate bond from 1997 to 2023, they find that roughly 73% of individual bond credit returns, 83% of the market credit premium, and 51% of the premia across 103 long-short bond factors are earned during the first five trading days of each month, a window that accounts for only about a quarter of trading days. On the remaining three weeks, credit returns are statistically indistinguishable from zero.

The pattern holds in both investment-grade and high-yield bonds, survives filters for illiquid or stale-quoted names, and shows up in credit default swaps as well. Even more striking, a simple one-factor bond CAPM explains 76% of the cross-sectional variation in factor returns during this window, versus 6% using full-month returns, which suggests the sprawling “factor zoo” is largely one credit premium wearing different costumes. For investors, the takeaway is that rebalancing timing is not neutral, and portfolios that reduce credit exposure early in the month may be systematically giving up the compensation they thought they were earning.

Dickerson, Alexander and Nozawa, Yoshio, Credit When it's Due: Corporate Bond Factors on a Schedule (September 25, 2026). Available at SSRN: https://ssrn.com/abstract=7524680 or http://dx.doi.org/10.2139/ssrn.7524680

When Futures Prices Break: Modeling the Oil Crash and Nickel Squeeze

A tractable options-pricing model explains how forced liquidation by delivery-constrained traders can push oil futures below zero and nickel futures to $100,000.

Zimbidis and Sircar tackle two of the strangest commodity-futures events in recent memory, the April 2020 WTI collapse to −$37.63 and the March 2022 LME nickel squeeze to $101,365, using a single framework. Their insight is to split traders into those who can actually take physical delivery (producers, refiners, warehouses) and those who cannot (most financial participants), and model the cost that unlicensed traders bear when they are forced to close positions into a thin pool of counterparties. That cost is a “roll option” whose payoff depends on the observed price it helps determine, which creates a feedback loop and a nonlinear pricing equation. Remarkably, the whole thing collapses into a Margrabe exchange-option formula with one scalar equation to solve.

Backing out the implied imbalance from market data gives roughly 87% for WTI at the April 20 low and 93% for nickel at the March 8 high. The practical takeaway is that extreme futures prints near expiry are not always “fundamentals breaking”, they can be the mechanical signature of who is allowed to stay in the contract and who is not.

Sircar, Ronnie and Zimbidis, Iosif, Negative Oil & Nickel Squeeze: A Feedback Model for Extreme Commodity Futures Prices (September 30, 2026). Available at SSRN: https://ssrn.com/abstract=7546838 or http://dx.doi.org/10.2139/ssrn.7546838

This week for paid subscribers

Paid subscribers are getting a full deep dive of the S&P 500 inclusion effect anomaly, reproduced on its original historical window and extended through 2024 on the author’s exact specification, against survival thresholds fixed before out-of-sample tests were observed. The post traces how the classic price-pressure spike decayed smoothly toward zero in the 2010s, and turns to the 2020s run-up to test whether the index demand mechanism or a false revival is actually behind it. Python notebook and data included.

Disclaimer: The content provided in this newsletter, “Alpha in Academia,” is for informational and educational purposes only. It should not be construed as financial advice, investment recommendations, or an offer or solicitation to buy or sell any securities or financial instruments. Past performance is not indicative of future results. The financial markets involve risks, and readers should conduct their own research and consult with qualified financial advisors before making any investment decisions.

The interpretations, opinions, and analyses presented herein are those of the author and do not necessarily reflect the views of the original researchers, their institutions, or the full implications of the cited academic papers. While every effort is made to accurately represent the research discussed, readers should be aware that the summaries and interpretations may not capture the full scope or nuances of the original studies. The information contained in this newsletter is believed to be accurate and reliable at the time of publication, but accuracy and completeness cannot be guaranteed. The author and publisher accept no liability for any loss or damage resulting from reliance on the information provided.

This newsletter may contain links to external websites or resources. The author is not responsible for the content, accuracy, or reliability of these external sources.

By subscribing to or reading this newsletter, you acknowledge that you have read and understood this disclaimer and agree to hold the author and publisher harmless from any liability that may arise from your use of the information contained herein.

Today we are looking at the famous demand-shock anomaly in equities. When S&P announces an addition to the 500, every fund tracking the index has to buy the stock by the effective date, and for decades that forced buying pushed the price up beforehand and let it fall back after. The question here is whether that still happens, and what to make of a run-up that has recently returned while the fall-back has not.

Let’s dive right in.

Introduction

The mechanism behind the index effect is price pressure. When S&P announces an addition, every fund tracking the index has to buy the stock, and has to finish buying by the effective date. That demand is large, mechanical, and indifferent to price. If the market cannot supply the shares at the prevailing price, the price rises, and once the buying stops it drifts back toward whatever fundamentals support.

Harris and Gurel documented this in 1986, finding a jump of roughly three percent around announcement that unwound within a fortnight. Shleifer published a companion result the same year and drew the sharper conclusion, that demand curves for stocks slope downward, which contradicts the textbook assumption that a stock’s price is pinned by its cash flows and that any quantity can be absorbed at that price.

The reversal is the half of the claim that does the work. A permanent repricing would mean inclusion changed something real about the company. A temporary one means index funds moved the price by leaning on it, and the price recovered when they stopped leaning. Without the reversal there is no evidence that demand did anything.

Greenwood and Sammon revisited the question and titled their paper The Disappearing Index Effect. Their explanation is migrations. A stock joining the S&P 500 today usually leaves the S&P MidCap 400 the same day, so funds tracking the 400 are forced sellers against the 500’s forced buyers, and the two flows partly cancel.

Data and Methodology

The sample is every addition to the S&P 500 recorded in the CRSP membership file between January 1995 and December 2024, which comes to 778 events. Everything is keyed on permno rather than ticker, so there is no ticker-mapping problem and no survivorship bias from companies that later delisted.

Abnormal returns are four-factor: market, size, value and momentum, with loadings estimated per event over days 300 through 121 before the effective date. The estimation window closes six months ahead of the event so that none of the pre-event drift leaks into the betas and gets subtracted as risk. I report the version that excludes the estimated alpha, since an alpha fitted over a window that may itself contain drift would quietly remove the thing I am trying to measure. The with-alpha figures are in the notebook and tell the same story at larger magnitudes.

CRSP records effective dates rather than announcement dates, and that distinction shapes everything downstream. S&P guarantees at least three business days of notice and announces quarterly changes roughly five trading days ahead, so the announcement lands somewhere in the days just before the effective date and moves around from event to event. I use days 10 through 1 before the effective date as the run-up window, days 30 through 11 before it as a background window that should contain no announcement at all, and the twenty trading days after it for the reversal.

Of the 778 additions, 642 have enough prior trading history to estimate betas. The 136 that do not are disproportionately recent listings, which matters and which I return to.

Results

The late 1990s run-up over the final ten trading days is 9.56% across 138 additions at t=12.84. The number worth sitting with is not that one. It is that 93.5% of those additions had a positive run-up. This was not a tendency you needed a regression to see. It was close to a rule.

The 2000s deliver 5.47% across 246 additions at t=8.53, with 76.0% of events positive. The decay has started and the effect is still unmistakable.

The 2010s deliver 0.57% across 188 additions, and a coin flip. The median run-up is 0.30% and 52.7% of events are positive.

Whether 0.57% is statistically distinguishable from zero depends on how you count. Treating each addition as an independent observation gives p=0.142. Averaging within effective dates first, which is the honest correction given that additions cluster on quarterly rebalance days, gives p=0.027. I would not build anything on either. An effect where barely half the events point the right way is not something you can put on.

From 2020 through December 2024, across the 70 additions with estimable betas, the run-up is 3.39% at t=3.39. That is the number this post exists to explain.

Table 1: Four-factor abnormal returns by era, excluding the estimated alpha. Share positive refers to the run-up window.

The decay through 2019 is replication. It lines up with Greenwood and Sammon closely enough that I take the pipeline as clean, which matters mostly because of what the last row does.

Read more

Let’s get into it.

Borrowed Money, Borrowed Momentum

Factor momentum gets much stronger right after investors borrow heavily to chase recent performance.

Momentum among asset pricing factors has always been awkward, because factors are supposed to pay you for bearing risk, not for their own past returns. Sun and Xia offer a concrete answer using an old and unglamorous data series, margin debt, the money brokerage customers borrow against their holdings. Across 201 equity factors, they find that after quarters when margin balances grow quickly, the shortest horizon version of factor momentum earns roughly 49 basis points a month more than usual, most of its typical risk adjusted return.

Figure 1: Factor momentum's risk-adjusted return, sorted by how fast margin debt grew the previous quarter. Chart recreated from Sun and Xia (2026), “Speculative Leverage and Factor Momentum,” Figure 2.

The effect fades as the momentum signal ages, concentrates in factors arbitrageurs find hardest to trade against, and shows up only in customers’ own borrowing, not in the cash sitting in their accounts or in hedge fund leverage.

Risk, uncertainty and credit conditions explain none of it. Once anomalies are published, their average returns decay but their momentum does not, which the authors read as “persistent speculative demand, financed with leverage and left uncorrected by arbitrageurs.” Leveraged performance chasing leaves a measurable footprint, and that footprint has timing value.

Sun, Weijian and Xia, Yu, Speculative Leverage and Factor Momentum. Available at SSRN: https://ssrn.com/abstract=7512099 or http://dx.doi.org/10.2139/ssrn.7512099

A Price the Market Declines to Adopt

India now settles its index derivatives on a closing auction price, and the futures market trades as though it only half believes it.

Since August 2026, India has priced the close of every derivative eligible stock through a twenty minute auction, and that price became the official settlement for index derivatives. The awkward part is that the published index is frozen by rule while the auction runs, so for roughly thirteen minutes a liquid futures market is valuing contracts against a number that is static by definition. When the auction price finally lands, futures shift only about a tenth of the way toward it, and nearly all of the remaining gap shows up in the futures basis instead.

Figure 2: Each dot is one trading session. The dashed line is where futures would sit if they fully accepted the auction's price. But they sit on the flat line instead.

Part of the move is anticipated during the freeze, but even measured generously the futures market prices in at most half of the revaluation. The author's broader point is that this auction passes every recognized benchmark integrity test and adoption still fails, because a settlement price can “still be a price the referencing market does not use.” The practical read is that option payoffs and futures marks can disagree on the same index at the same instant, and only the settlement contract closes that gap.

Balas, Vijay, A Settlement Price the Market Declines to Adopt: Evidence from India's Closing Auction (September 21, 2026). Available at SSRN: https://ssrn.com/abstract=7502458 or http://dx.doi.org/10.2139/ssrn.7502458

When the Benchmark Is Risky

Swap rates sit below Treasury yields because the U.S. government, unlike the benchmark rates banks swap against, can actually default.

Since 2008, the fixed rate on interest rate swaps has sat below the yield on Treasuries of the same maturity. That ordering should be impossible, since Treasuries are meant to be the safest and most liquid asset in the market.

The authors argue the puzzle dissolves once you allow the U.S. government to be, in a small way, risky. A swap is tied to a benchmark index like the federal funds rate, and an index cannot default because nobody can own it. But a Treasury bond can. That asymmetry makes the swap the safer leg and drags its rate below the Treasury yield. Credit default swap premiums on U.S. debt, now roughly 100 times their pre-crisis level, move in step with these spreads.

Figure 3: Model-implied OIS minus Treasury spreads at 30 years. The solid line includes U.S. default risk and turns sharply negative after 2008. The counterfactual without default risk (shaded) stays near zero. The gap between them is the sovereign risk premium.

Strip sovereign credit risk out of their model and spreads turn positive at every maturity. For investors, the risk-free rate is a modeling convenience rather than an observed price, and “the risk premium associated with it may be large.”

Augustin, Patrick and Chernov, Mikhail and Schmid, Lukas and Song, Dongho, Benchmark Interest Rates When the Government is Risky (November 2019). NBER Working Paper No. w26429, Available at SSRN: https://ssrn.com/abstract=3480299

One Institution, Many Voices

Mutual fund families are far less unified in how they vote than almost everyone assumed, and that internal splintering quietly weakens their grip on corporate boards.

The standard assumption, in academic work and in the current policy fight over proxy voting, is that when a giant like Vanguard or BlackRock casts a vote, the entire fund family speaks with one voice. This paper shows that assumption fails routinely, and has been failing since at least 2006. Internal disagreement looks trivial at roughly 4 percent of votes, but that figure is drowned out by routine proposals nobody argues about. On contested items, environmental and social proposals, and votes where proxy advisors recommend going against management, internal splits rise to the 15 to 20 percent range.

Figure 4: Same level of shareholder opposition, different outcome. When fund families vote as a bloc (blue), directors are far more likely to leave the board within three years. When families split internally (red), the pressure flattens out.

The consequence matters more than the measurement. When a family votes as a bloc, directors facing shareholder opposition are meaningfully more likely to leave the board. When the family splits, that pressure fades, a pattern the authors confirm using Vanguard's 2019 decentralization as a natural experiment. Funds that break ranks charge higher fees without better returns, yet institutional clients reward them with inflows. Pass through voting may give shareholders more choice and less actual influence.

Michaely, Roni and Ringgenberg, Matthew C. and Rubio, Silvina and Yi, Irene, Decentralized Voting in Mutual Fund Families (September 19, 2026). European Corporate Governance Institute – Finance Working Paper no. 1160/2026, Available at SSRN: https://ssrn.com/abstract=7491320 or http://dx.doi.org/10.2139/ssrn.7491320

This week for paid subscribers

Paid subscribers are getting a full replication of the FOMC cycle anomaly, reproduced on its original 1994 to 2013 window and extended through 2026 on the authors' exact specification, against survival thresholds fixed before the out-of-sample estimate was observed. The post traces how the even-week coefficient decays smoothly toward zero as newer data enters, and turns to the Treasury curve to test whether the Fed leakage mechanism was ever there to begin with. Python notebooks and data included.

Disclaimer: The content provided in this newsletter, "Alpha in Academia," is for informational and educational purposes only. It should not be construed as financial advice, investment recommendations, or an offer or solicitation to buy or sell any securities or financial instruments. Past performance is not indicative of future results. The financial markets involve risks, and readers should conduct their own research and consult with qualified financial advisors before making any investment decisions.

The interpretations, opinions, and analyses presented herein are those of the author and do not necessarily reflect the views of the original researchers, their institutions, or the full implications of the cited academic papers. While every effort is made to accurately represent the research discussed, readers should be aware that the summaries and interpretations may not capture the full scope or nuances of the original studies. The information contained in this newsletter is believed to be accurate and reliable at the time of publication, but accuracy and completeness cannot be guaranteed. The author and publisher accept no liability for any loss or damage resulting from reliance on the information provided.

This newsletter may contain links to external websites or resources. The author is not responsible for the content, accuracy, or reliability of these external sources.

By subscribing to or reading this newsletter, you acknowledge that you have read and understood this disclaimer and agree to hold the author and publisher harmless from any liability that may arise from your use of the information contained herein.

In 2019, Cieslak, Morse and Vissing-Jørgensen published a paper in the Journal of Finance arguing that the entire post-1994 US equity risk premium was earned during the “even” weeks of the Federal Reserve’s rate-setting cycle. The Wall Street Journal wrote it up. The Economist wrote it up. It became a fixture in trading-desk lore and in classroom arguments about market efficiency. This piece reproduces their finding on the original 1994 to 2013 sample, extends it through the end of 2026 on the exact specification they defined, and traces what happens to the pattern once we let fresh data into the estimation. The result is a case study in how a striking empirical claim moves through its own out-of-sample life, and in what a careful replication can teach even when the headline pattern does not survive.

Let’s dive right in.

What CMVJ found

Cieslak, Morse and Vissing-Jørgensen (CMVJ from here) made a specific claim about the timing of equity risk premia. Their construction is mechanical. Day 0 is the FOMC announcement day. Days minus one through three form week 0. Days four through eight form week 1. The counter runs out to week 6, at which point the next FOMC meeting resets it. Weeks 0, 2, 4 and 6 are even; weeks 1, 3 and 5 are odd. Their finding, from 1994 through the end of 2013, was that the market’s daily excess return averaged roughly twelve basis points on even-week days and essentially zero on odd-week days. Half the trading days, nearly all the return.

Figure 1 shows the image that made the paper famous. A hundred dollars invested in the US market only during even weeks tracks the buy-and-hold portfolio closely across two decades, with half the time in market.

Figure 1. Value of $100 invested in the US market, 1994 to 2013. The even-week strategy captures nearly all the return of continuous exposure with half the time in market.

Their proposed mechanism was informational. The Federal Reserve’s Board of Governors historically held internal briefings on a biweekly cadence, and CMVJ argued that information from those briefings leaked selectively to sophisticated market participants on an even-week rhythm. Sophisticated investors, holding stock through the even-week window when good news was more likely to be revealed, earned a premium for absorbing that risk. It was a story that fit the pattern, fit the institutional detail, and was specific enough to be testable.

Read more

Let’s get into it.

Fair-Weather Liquidity: When HFT Helps Corporate Bond Investors, and When It Runs

High-frequency traders tighten corporate bond spreads in calm markets and then vanish exactly when funds need them most.

Using two decades of TRACE, CRSP, and TAQ data covering 712 mutual funds and roughly 14 million bond trades, this paper maps how algorithmic trading and portfolio illiquidity jointly drive corporate bond fragility. The headline finding is a clean split by regime. In normal conditions, HFT activity compresses bid-ask spreads, improves price efficiency, and even buffers funds against flow-induced volatility. But during the 2008 crisis, the 2013 Taper Tantrum, and the March 2020 dash-for-cash, HFT participants withdraw sharply, and spreads widen fastest for the very bonds that had benefited most from their presence.

Figure 2. HFT Activity Index and Corporate Bond Bid-Ask Spread (2003–2023). Quarterly averages of the fund portfolio-weighted HFT intensity index (solid blue area) and bid-ask spread in basis points, inverted and scaled for comparability (dashed red area). HFT intensity is standardized. Bid-ask spread is the portfolio-weighted average of bond-level estimates from TRACE Enhanced.

Layered on top, funds holding more illiquid portfolios show much steeper flow-performance concavity, meaning investors redeem harder on bad returns, and this asymmetry is significantly stronger in low-rate environments where investors have fewer income alternatives. Funds with more illiquid holdings do earn higher alphas in calm periods (consistent with a liquidity premium) but suffer larger drawdowns when HFT retreats. For bond fund investors, the practical takeaway is that headline liquidity metrics look best precisely when they matter least.

Zhao, Yichang and Yang, Liu, Market Liquidity, High-Frequency Trading, and Corporate Bond Pricing: Evidence from U.S. Mutual Fund Flows and Volatility Dynamics. Available at SSRN: https://ssrn.com/abstract=7412644 or http://dx.doi.org/10.2139/ssrn.7412644

The Portfolio Optimization Problem Everyone Skips: Choosing the Information Itself

Treating your choice of inputs as a decision, not an assumption, exposes two silent errors that make backtests lie and diversification break.

Standard mean-variance optimization takes the inputs (expected returns, covariances, the signals feeding them) as given and just picks weights. This paper argues that’s where most of the damage happens, because two errors sneak in unnoticed: using data that wouldn’t actually have been available at decision time (look-ahead), and treating shared drivers of risk as if each asset moved on its own. The author reframes the problem as a two-stage decision, first selecting an admissible information set (chronologically valid, arbitrage-preserving, and statistically able to separate common from idiosyncratic risk), then running the classical optimizer inside it. Tested on twelve US equities against 127 macro and financial drivers, the framework meets its separation condition on individual stocks but fails on pre-sorted portfolios (industry, size, value), because the residual dependence there is the common factor itself.

Even when the condition fails, the resulting minimum-variance portfolios match Ledoit-Wolf shrinkage on volatility at roughly a third less turnover, though reported risk is understated by around 15 percent. The message for anyone building factor or multi-asset portfolios: the covariance matrix you trust is only as honest as the information set you never audited.

Rodriguez Dominguez, Alejandro, Admissible Portfolio Optimization: Information Constraints, Conditional Efficient Frontiers, and the Price of Causal Identification (September 15, 2026). Available at SSRN: https://ssrn.com/abstract=7468398 or http://dx.doi.org/10.2139/ssrn.7468398

A Reliability-Aware Approach to Crypto Portfolio Optimization

Weighting expert forecasts by how reliable they actually are can completely reshape which crypto assets end up in your portfolio, not just tweak the weights.

Most quantitative portfolio models treat every expert return forecast as equally credible, which is a strong assumption when the inputs come from analysts with wildly different track records. This paper builds a framework using Z-numbers (a construction from fuzzy set theory that pairs each forecast with a reliability score) and plugs it into standard credibilistic VaR and CVaR optimization. The authors test it on 46 cryptocurrencies across 27 scenarios spanning different return targets, portfolio sizes, and allocation constraints. The headline result is that reliability weighting is not a marginal adjustment. In one representative scenario, the standard model concentrates in BNB, BTC, and MNT, while the reliability-aware version holds CRO, NEAR, and SEI, sharing zero assets in common.

A sensitivity analysis shows the framework responds cleanly to shifts in reliability inputs, which also means sloppy reliability estimates will push you into a bad portfolio just as surely as ignoring reliability entirely. For anyone building crypto allocations off analyst forecasts or model outputs of varying quality, the takeaway is that treating information quality as a first-class input, not an afterthought, materially changes what you own.

Ghanbari, Hossein and Mohammadi, Emran and Sadjadi, Seyed Jafar and Kumar, Ronald and Stauvermann, Peter J., A Novel Framework for Considering the Experts’ Reliability in Portfolio Optimization under Partial Information: Incorporating Z-numbers into Credibilistic Quantile-Based Risk Measures(September 06, 2026). Available at SSRN: https://ssrn.com/abstract=7424938 or http://dx.doi.org/10.2139/ssrn.7424938

How SLR Constraints Broke the Treasury Market in March 2020

When leverage rules bite, primary dealers dump Treasuries at fire-sale prices, and the whole market feels it.

During the March 2020 dash-for-cash, primary dealers were supposed to absorb a tidal wave of Treasury selling, but a quiet regulatory constraint got in the way. This HKIMR study uses transaction-level data from the Fed’s emergency purchase program to show that bank-affiliated dealers running low on Supplementary Leverage Ratio headroom sold Treasuries to the Fed at meaningfully lower prices than their less constrained peers. For 30-year bonds, dealers with one percentage point less SLR headroom accepted roughly 46 basis points lower prices, a sizable haircut when bid-ask spreads were already blown out.

The authors then run a counterfactual, asking what would have happened if the Fed’s recently proposed SLR reform (which would have added about 120 bps of headroom on average) had been in place. Price dispersion, a clean proxy for illiquidity, drops by roughly 20 percent, though it still sits well above normal levels. The takeaway for investors is that capital rules quietly shape who can warehouse risk in a crisis, and dealer balance sheet constraints can turn a liquidity event into a price dislocation.

Institute for Monetary and Financial Research, Hong Kong, Assessing the Impact of Supplementary Leverage Ratio Requirements on the Market-Making Behaviours of US Primary Dealers in the US Treasury Market (September 14, 2026). Hong Kong Institute for Monetary and Financial Research (HKIMR) Research Paper No. 09/2026, Available at SSRN: https://ssrn.com/abstract=7458278 or http://dx.doi.org/10.2139/ssrn.7458278

This week for paid subscribers

Paid subscribers are getting a look at what happens when a massive anomaly database meets four decades of out-of-sample data. Across 210 published strategies, in-sample monthly returns average 0.671% but drop by roughly half outside that window. Pre-discovery and post-publication returns are statistically indistinguishable, proving the decay is driven by overfitting rather than academic arbitrage. Robustness checks confirm the shortfall is not a data backfilling artifact or a microcap distortion. Python backtest code included. [Part 1 of 2]

Disclaimer: The content provided in this newsletter, “Alpha in Academia,” is for informational and educational purposes only. It should not be construed as financial advice, investment recommendations, or an offer or solicitation to buy or sell any securities or financial instruments. Past performance is not indicative of future results. The financial markets involve risks, and readers should conduct their own research and consult with qualified financial advisors before making any investment decisions.

The interpretations, opinions, and analyses presented herein are those of the author and do not necessarily reflect the views of the original researchers, their institutions, or the full implications of the cited academic papers. While every effort is made to accurately represent the research discussed, readers should be aware that the summaries and interpretations may not capture the full scope or nuances of the original studies. The information contained in this newsletter is believed to be accurate and reliable at the time of publication, but accuracy and completeness cannot be guaranteed. The author and publisher accept no liability for any loss or damage resulting from reliance on the information provided.

This newsletter may contain links to external websites or resources. The author is not responsible for the content, accuracy, or reliability of these external sources.

By subscribing to or reading this newsletter, you acknowledge that you have read and understood this disclaimer and agree to hold the author and publisher harmless from any liability that may arise from your use of the information contained herein.

Part 2 of 2. Part 1 measured what 210 published anomalies earned inside and outside the window their own papers examined, and found returns run at roughly half outside it in both directions. This piece takes the one number that looked like an arbitrage effect and finds that the method producing it would have produced it anyway.

The standard design compares an anomaly’s returns in the window after its paper’s sample ends to its returns after the paper is published, and attributes the difference to investors learning from the literature. Run on 210 anomalies, that design gives a residual of 13.4% of the in-sample return. Run on the same anomalies with the real publication dates thrown out and fake ones substituted, it gives 13.4%.

Across 200 randomized runs the average gap was −0.0897. The one measured from genuine publication dates was −0.0903, sitting at the 49.5th percentile of the random distribution. The reason is that anomaly returns drift downward across their whole post-sample life, so splitting that history at any point produces a lower second half. The estimator is built to return a negative number, and calling that number arbitrage attributes a trend to an event.

Let’s dive right in.

Introduction

Part 1 cut each anomaly’s history into four regimes on its own clock: the decades before its paper’s sample begins, the paper’s own window, the gap between the end of that window and publication, and everything since. The headline was that returns outside the discovery window run at about half those inside it, and that pre-discovery and post-publication are statistically indistinguishable from each other.

One thing was deliberately left hanging. The decay is not flat across the two post-sample regimes. Returns fall 0.247 percentage points once the paper’s sample ends, and 0.338 once the paper is out. The difference, 0.090 percentage points or 13.4% of the in-sample return, arrives only after publication.

That gap is doing a lot of work in this literature. It is the quantity McLean and Pontiff built their headline on, attributing 32 points of it to investors learning about mispricing from academic research. The logic is clean and I find it persuasive on its face. Decay in the post-sample window cannot be publication-driven, because the paper is not out. Anything on top of that has to be something publication caused.

The logic has a hole in it, and the hole is not about these anomalies. It is about the estimator.

The argument assumes that if publication did nothing, the two windows would look alike. That assumption holds only if returns are flat across the post-sample period. If instead they drift downward throughout, then the later window is lower than the earlier one for reasons that have nothing to do with the paper, and the difference between them gets read as arbitrage regardless.

In plain terms. The published method works like this: an anomaly's returns fall after its paper comes out, and the portion of that fall which appears only after publication gets labelled arbitrage. That reasoning holds if returns would otherwise have stayed flat. If they were already sliding, then the second half of any comparison comes in below the first, and the method reports arbitrage for a decline that was happening regardless. The whole of this post is one question: Which of those two is going on here?

This post tests that directly. The cleanest way to find out whether a date is doing any work is to replace it with a date that cannot be, and see whether the answer changes.

Data and Methodology

Same dataset and construction as Part 1. The October 2025 release of Open Source Asset Pricing, 212 predictors with monthly long-short returns from January 1926 to December 2024, built the way each original paper built them, with 210 surviving a requirement of at least 24 months in each regime. Regressions are pooled with standard errors clustered by calendar month.

The companion notebook is standalone. It rebuilds the regimes from scratch rather than depending on Part 1’s, so it runs on its own once the data cache exists.

One measurement note that matters throughout. The natural way to express decay is as a share of the in-sample return, but that ratio explodes when an anomaly’s in-sample return sits near zero, and a handful of those dominate any average. The unwinsorized mean post-publication ratio is −0.797. Winsorized at the 5th and 95th percentiles it is −0.576, and the median is −0.617. So raw differences in percentage points are the headline here and ratios are the readable secondary view. Where they disagree, trust the raw difference.

Results

The regression is Part 1's. In-sample is 0.671% a month, the post-sample window comes in 0.247 lower at t=−3.77, and post-publication comes in 0.338 lower at t=−6.56. Both are significant, and neither is the quantity in question.

The test that matters is not either coefficient. It is the difference between them, which is the part publication could be responsible for:

post-publication minus post-sample
coefficient  −0.0903
std error     0.0700
z            −1.30
p             0.195
95% interval [−0.227, +0.046]

Not significant, and the interval contains zero comfortably. But a p-value of 0.195 on its own is a weak thing to argue around. Plenty of real effects fail to clear significance with 210 units and a noisy dependent variable, and failing to reject the null is not the same result as claiming there is no conclusion to be drawn.

The better test is to ask what this estimator returns when the date it depends on is meaningless. Assign each anomaly a fake publication year drawn from the same distribution of sample-end-to-publication gaps the real data shows, re-cut the regimes, and re-run. Two hundred times.

observed gap              −0.0903
randomized gaps, mean     −0.0897
randomized gaps, sd        0.0274
share at least as negative  0.495

The real publication dates produce a coefficient of −0.0903. Fake ones produce an average of −0.0897. The observed value sits at the 49.5th percentile of the randomized distribution, which is as close to the middle as this test can put it. The estimator returns the same answer either way.

Read more

Part 1 of 2. This piece sets up the measurement and asks what a published anomaly was doing in the decades before its own paper's sample begins. Part 2 takes the post-publication drop that falls out of it and tests whether that drop is an arbitrage effect at all.

Today we are looking at what happens to a published stock market anomaly across its life. Across 210 anomalies from the academic literature, the long-short return inside each paper’s own sample averages 0.671% a month. Outside that window it averages roughly half as much, and it does not much matter which direction you look. Before the anomaly was discovered, 0.402%. After it was published, 0.334%.

Those two out-of-sample figures are not distinguishable from one another. A formal test of the difference between them returns a z-statistic of 1.04 and a p-value of 0.299, which is nowhere near rejecting equality. The decay from in-sample is large and it is robust. What it is not is something that happened at publication, because the same shortfall is already sitting there in data from decades earlier.

Let’s dive right in.

Introduction

The standard account of anomaly decay runs through arbitrage. An academic finds a predictable pattern in returns, publishes it, practitioners read the paper and trade the signal, and the excess return gets competed away. McLean and Pontiff put hard numbers on this in 2016. Across 97 predictors, portfolio returns were 26% lower out-of-sample and 58% lower after publication, and they attributed the 32-point gap between those to investors learning about mispricing from the academic literature.

The mechanism is sensible and the decay is easy to verify. The part that gets less attention is the counterfactual. If you want to know what publication did to an anomaly, you need a period where the anomaly was measurable but had not been selected, and there are two such periods rather than one. There is the obvious one after the paper comes out. There is also a much longer one before the paper’s sample even begins.

It is worth being precise about why that second period counts as out-of-sample, because there is an obvious objection. All of this is a backtest. The authors were themselves running a backtest when they wrote the paper. So what exactly separates their window from mine?

The answer is selection, not trading. The authors chose that signal, that sort, those breakpoints, and that sample because the combination worked in the window they examined. Every one of those choices is a degree of freedom, and every one was exercised against in-sample data. The pre-discovery window was never part of the search space. It is out-of-sample in exactly the way a holdout period is out-of-sample, and the fact that nobody was trading on it is beside the point. Some practitioners may well have been trading these signals before the papers appeared, and I have no way of knowing. What I can say is that the specification search that produced the published result never saw this data.

In plain terms. If you try enough variations of a trading rule against the same stretch of history, one of them will look excellent whether or not it means anything. The only way to separate a real edge from a lucky variation is to run it on data you never touched while choosing. The decades sitting before each paper's sample are exactly that data, and nobody has to have been trading on them for that to be true.

Linnainmaa and Roberts made this argument in the Review of Financial Studies in 2018. They hand-collected accounting data from Moody’s manuals back to 1918 so they could measure 36 anomalies before their discovery, and they found that the pre-discovery and post-discovery periods resemble each other while the in-sample period resembles neither. Their conclusion was that most accounting-based anomalies are largely a data-snooping artifact.

I wanted to run that design on a much wider set of anomalies, with portfolio construction that follows each original paper rather than a standardized factor build, and with the nine additional years of returns that have accumulated since they wrote.

Data and Methodology

Everything comes from the Open Source Asset Pricing project, the October 2025 release. Chen and Zimmermann replicate the cross-sectional asset pricing literature and publish both the signals and the resulting portfolios, and the project exists to be used and cited rather than licensed.

The dataset gives 212 predictors with monthly long-short returns from January 1926 to December 2024, built the way each original paper built them. I use that construction rather than a house-standard decile sort deliberately. The question is what happened to the strategy each paper actually published, not to a re-specified version of it.

The accompanying documentation file is what makes the design possible. For every anomaly it records the publication year, the journal, the start and end years of the original study’s sample, and the return and t-statistic the authors reported. Publication years run from 1973 to 2016. Original samples end between 1968 and 2014.

From those two dates I cut each anomaly’s return history into four regimes on its own clock. Everything before the paper’s sample begins is pre-discovery. The paper’s own sample window is in-sample. Everything from the end of that sample up to publication is the post-sample window. Everything after publication is post-publication.

The gap between sample end and publication is what makes this design work. It is never zero in this dataset. The minimum is two years, the median is four, the maximum is eleven. That window is a period when the anomaly’s most recent behavior was unknown to its own authors and unknown to everyone else, which separates two explanations that otherwise get confounded. Decay that shows up there cannot be arbitrage, because there was nothing to arbitrage yet. It has to be overfitting.

In plain terms. Every anomaly gets a timeline with four stretches: the long run-up before its paper's data even starts, the paper's own window, a short gap after that window closes but before the paper appears, and everything since. The short gap is the useful one. Nobody outside the authors knew about the finding yet, so anything that falls away there is the original result being flattered by its own sample rather than traders competing it away.

Two choices are worth stating. I treat a paper published in year Y as becoming public on 31 December of Y, which assigns the whole publication year to the pre-publication side. That is the conservative direction, as it biases the design against finding a publication effect. And I require at least 24 months in each of the three regimes the main regression uses, which drops two of the anomalies.

Regressions are pooled across anomalies and months with standard errors clustered by calendar month. That correction matters here more than it usually does. These 210 strategies hold overlapping positions in the same stocks and move together, and treating their monthly returns as independent observations would make everything look far more significant than it is.

Here is what the design looks like on four anomalies you will recognize.

Figure 1: Cumulative long-short returns for four familiar anomalies, with each one's regimes shaded on its own clock. Light blue is the paper's own sample. Grey is the window after that sample ended but before the paper appeared. The dashed red line is publication.

Results

Taking in-sample as the reference, pre-discovery averages 0.402% a month, which is 0.269 percentage points lower with a t-statistic of −4.58. The post-sample window averages 0.424%, down 0.247 at t=−3.77. Post-publication averages 0.334%, down 0.338 at t=−6.56. In-sample itself is 0.671%.

Total decay from in-sample to post-publication is 50.3%. About three quarters of that, 36.9 percentage points, is already present in the post-sample window, before anyone outside the authors could have read the paper.

Those two numbers are directly comparable to McLean and Pontiff, and the comparison is informative. Their out-of-sample decline was 26% against my 36.9%, and their post-publication decline was 58% against my 50.3%. So I find more decay before publication than they did and less after it, which leaves a residual of 13.4 points where they found 32.

The comparison that carries this post, though, is between the two out-of-sample periods. Pre-discovery minus post-publication is +0.069 percentage points with a z-statistic of 1.04 and a p-value of 0.299. Whatever an anomaly earned in the decades before its discovery, it earned about the same after everybody had read about it, and the difference between those two is well inside noise.

The in-sample window is the outlier. Step outside it in either direction and you get roughly half the advertised number.

In plain terms. The average anomaly earned about 0.67% a month in the window its discoverers examined and somewhere between 0.33% and 0.42% everywhere else. Whether “everywhere else” means the forty years before the paper or the twenty years after it makes almost no difference. Roughly half the advertised return was a property of the window rather than of the strategy.

I have deliberately left the arbitrage question out of this post. That 0.338 post-publication coefficient looks like an arbitrage effect, and establishing whether it is one takes more space than I have here. It is the subject of Part 2.

Read more

Let’s get into it.

Return-Optimal Regimes: Labeling Markets by What You Should Have Held

Defining a market “regime” as the holding path a reluctant trader would have wanted in hindsight, rather than a hidden statistical state, roughly doubles the risk-adjusted return of a simple industry rotation strategy.

Most regime models ask what state the market was probably in. Li and Mulvey ask a blunter question: Knowing what happened, which asset should you have held each day, assuming you hate trading?

They solve that exactly with a small dynamic program, then train one gradient-boosted classifier across all 46 Fama-French industries to predict those hindsight answers a day ahead. The setup is stacked in three layers. First, each industry is rotated against semiconductors as the aggressive bet. Second, when enough industries vote to retreat, capital shifts into tobacco stocks or T-bills. Third, the best-behaved pairs get picked by trailing return-per-drawdown and scaled back when volatility spikes. Out of sample from 1995 to 2025, the full stack lifts the Sharpe ratio from 0.56 to about 1.0 while cutting the worst drawdown from 57% to 22%.

Figure 1: Cumulative wealth and drawdowns, 1995 to 2025, adding one layer of the strategy at a time. The defensive overlay does most of the work in 2000 to 2003 and 2008. Recreated from Li and Mulvey (2026), Figure 8.

The part that is most interesting is the head-to-head. Run the identical pipeline with classical hidden Markov labels and the defensive layer actually hurts. The labels, not the machinery, carry the result. This study argues that “regime” should mean “what you would have done,” and that a clean crash exit matters more than picking the right sector.

Li, Silu and Mulvey, John M., Return-Optimal Regime Labels and Progressive Risk Overlays: A Hierarchical Framework for Industry Allocation. Available at SSRN: https://ssrn.com/abstract=7434978 or http://dx.doi.org/10.2139/ssrn.7434978

The Geometry of a Turning Point

A rule that ignores price levels and asks only whether today's trading range fits inside yesterday's lands far from a random walk in nearly every market tested.

Most turning point detectors flatten a daily candle to a single closing price. Liu keeps the whole range, low to high, and asks a simpler question: How often does one day’s range nest entirely inside the previous day’s (or the reverse)? If prices followed a pure random walk, the answer is a fixed constant, about one day in five, and it doesn’t move with volatility or drift. That gives a clean benchmark to test against.

Figure 2: Same stock, same 200 days. Top: turning points after nested candles are merged and short swings filtered. Bottom: every three-bar high and low. Source: Liu (2026), Figure 4.

Across seven asset classes, US and Hong Kong stocks nest at roughly double that rate, currencies and crypto sit well above it, and only broad equity indices land on the random walk value almost exactly. The annoying part is that the usual suspects don’t explain the gap.

Volatility clustering, jumps, and stochastic volatility barely nudge the number, and the author concedes the paper is “leaving open which features of the data-generating process generate it.” For traders, the immediate payoff is the extrema detector itself: No lookback window to tune, no smoothing, and turning points that fall on actual highs and lows. The bigger message is that the shape of daily ranges carries structure that return-based models miss, and it shows up most where individual stocks trade.

Liu, Xinyong, Price Overlap Rate III: Calibration-Free Geometric Local Extrema in OHLC Interval Sequences (August 30, 2026). Available at SSRN: https://ssrn.com/abstract=7415218 or http://dx.doi.org/10.2139/ssrn.7415218

DeFi Interest Rates Follow the 10-Year Treasury

Interest rates on DeFi stablecoin loans track the U.S. 10-year Treasury yield, and the link runs through how heavily the lending pools get used.

Aave is the biggest lending protocol in crypto, and its loans look nothing like a bank’s. No maturity date, everything overcollateralized, and rates set by a formula that reacts to how much of each pool is borrowed out. You would expect those rates to live in their own world. Spoiler, they don’t.

Using daily data from January 2023 to March 2026 across 28 pools, Bhambhwani finds that borrowing and deposit rates on USDC, USDT and DAI move with Treasury yields, and the 10-year does most of the work once every maturity sits in the same regression. A one standard deviation rise in the 10-year lifts stablecoin borrowing rates by roughly 1.1 percentage points.

Figure 3: Aave's USDC borrowing rate alongside the U.S. 10-year and 3-month Treasury yields, January 2023 to March 2026. Each series is standardized and smoothed with a 7-day rolling average. Source: Bhambhwani (2026), Figure 1.

The mechanism is neat, with higher Treasury yields pulling up borrowing faster than deposits, so pool utilization climbs and Aave’s rate formula does the rest. Ethereum and Bitcoin pools show no such link, which makes sense, since most people don’t park ETH in a lending pool to compete with bond yields. The author’s read is that stablecoin rates are “not detached from traditional financial markets.” For anyone earning yield on stablecoins, the bond market is now part of the forecast.

Bhambhwani, Siddharth, DeFi Interest Rates and Treasury Yields (July 31, 2026). Finance Research Letters, volume 111, 2026[10.1016/j.frl.2026.110678], Available at SSRN: https://ssrn.com/abstract=7387780 or http://dx.doi.org/10.2139/ssrn.7387780

Herding in Green Hydrogen Stocks: Falling Markets, Liquidity, and the Hormuz Shock

Hydrogen stocks only herd when they are falling, and a real oil-supply shock makes them scatter instead.

Seventeen pure-play hydrogen stocks (electrolyser and fuel-cell makers, with the diversified giants stripped out) look calm on average. Across the full 2018 to 2026 sample, there is no statistical sign that investors copy each other. Condition on what the market is doing and the picture changes.

On down days, however, returns bunch together far more than the size of the move justifies, with the classic herding signature and the same crowding shows up during the acute phase of the Russia-Ukraine war and faintly during the first Covid months. The point made is loss aversion. When the transition story wobbles, holding a divergent position feels expensive, so people sell what everyone else is selling.

Figure 4: Herding in hydrogen stocks is a mood, not a habit. Each dot is the herding coefficient (γ) estimated over a trailing 250-day window. Recreated from Figure 2 of Benyahia, Ben Amar, Bouoiyour and Bouattour (2026), “Do Investors Herd in Green Hydrogen Stocks?”, SSRN preprint (not peer reviewed).

The 2026 Strait of Hormuz closure did the opposite. Dispersion widened, driven almost entirely by the American names, where gas-fed fuel-cell firms sit next to pure hydrogen plays and an oil shock hits them in opposite directions. The authors’ warning is that “the diversification sought within a hydrogen basket largely disappears precisely when it is most needed.” Thematic baskets protect you in the wrong states, and supply shocks reward stock picking over the sector bet.

Benyahia, Georges Ali and Ben Amar, Amine and Bouoiyour, Jamal and Bouattour, Mondher, Do Investors Herd in Green Hydrogen Stocks? Market States, Liquidity, and Energy-Supply Shocks. Available at SSRN: https://ssrn.com/abstract=7435238 or http://dx.doi.org/10.2139/ssrn.7435238

This week for paid subscribers

Paid subscribers are getting a look at what happens when a cost-aware dynamic allocation model meets real ETF prices. We reproduce Kolm and Ritter's closed-form result exactly, then run the same machinery walk-forward against a fixed 50/40/10 mix with real trading costs and estimated momentum forecasts. The passive mix wins on return, Sharpe, and drawdown, and a forecast-skill audit explains why: the signals were anti-informative, so the better optimizer just acted on bad information more precisely. Python backtest code included.

Disclaimer: The content provided in this newsletter, "Alpha in Academia," is for informational and educational purposes only. It should not be construed as financial advice, investment recommendations, or an offer or solicitation to buy or sell any securities or financial instruments. Past performance is not indicative of future results. The financial markets involve risks, and readers should conduct their own research and consult with qualified financial advisors before making any investment decisions.

The interpretations, opinions, and analyses presented herein are those of the author and do not necessarily reflect the views of the original researchers, their institutions, or the full implications of the cited academic papers. While every effort is made to accurately represent the research discussed, readers should be aware that the summaries and interpretations may not capture the full scope or nuances of the original studies. The information contained in this newsletter is believed to be accurate and reliable at the time of publication, but accuracy and completeness cannot be guaranteed. The author and publisher accept no liability for any loss or damage resulting from reliance on the information provided.

This newsletter may contain links to external websites or resources. The author is not responsible for the content, accuracy, or reliability of these external sources.

By subscribing to or reading this newsletter, you acknowledge that you have read and understood this disclaimer and agree to hold the author and publisher harmless from any liability that may arise from your use of the information contained herein.

There is a comfortable story in quantitative finance that says the right optimization, aware of trading costs and of how signals decay, should convert a static asset mix into a smarter, cost-conscious one. A recent paper makes that case with unusual rigor and a clean headline number: its dynamic rule recovers slightly more than half of the performance gap between a naive tactical portfolio and a hypothetical oracle that knows the future.

We reproduced that result exactly. Then we did the thing the paper does not do, which is to hand the same machinery real ETF prices, real estimated forecasts, and real trading frictions, and let it compete against a portfolio that simply rebalances to fixed weights and otherwise does nothing. Across thirteen years, a second asset universe, and a battery of pre-specified controls, the do-nothing portfolio won.

This piece explains why.

The Question

Tactical asset allocation promises to add value on top of a strategic policy by leaning into assets when their expected returns look high and away when they look low. The academic version of this promise is more careful than the marketing version. It acknowledges that every tilt costs money to put on, that forecasts decay at different speeds, and that a portfolio which chases each new signal will churn itself into the ground. The state of the art therefore frames allocation as a dynamic control problem: choose today’s trades while accounting for tomorrow’s costs and the persistence of today’s information.

Kolm and Ritter (2026) solve exactly that problem in closed form and show, in a controlled numerical setting, that their value-optimal rule captures 51.1 percent of the objective loss that a conventional tactical rule leaves on the table relative to a dynamic oracle. It is an elegant result. The natural question for a practitioner, and the one this study asks, is narrower and more stubborn: when you replace the paper’s known model parameters with quantities you have to estimate from data, and when you charge real trading costs, does any of that theoretical advantage survive contact with the market?

Our answer is that it does not. The optimizer works. The forecasts do not. And once the forecasts fail, every layer of sophistication built on top of them, dynamic planning, entropic risk adjustment, restricted policies, inherits the failure. The rest of this piece walks the evidence.

Reproducing the Model

Before testing a framework it is worth confirming the framework is sound. We rebuilt the paper’s Section 8 numerical illustrations from the stated equations, using no author code, no market data, and no paid datasets. We estimated the restricted policy parameters independently rather than copying the published values. Every one of the twelve entries in the paper’s central table reproduced to its displayed precision.

Reproduced values match the paper to five decimal places on all twelve entries. The value-optimal rule recovers 51.113574 percent of conventional TAA’s loss relative to the oracle, against the paper’s reported 51.1 percent.

Our independently optimized restricted policy landed at retention and tracking parameters of (1.16929, 1.00000, 0.88727), matching the paper’s (1.1693, 1.0000, 0.8873). Fixed-point residuals for the oracle and reported policies sat below 1e-12, and a separate 350-point parameter search failed to beat the optimized loss, a numerical cross-check rather than a proof of global optimality. In short, the mathematics is not in question. The 51.1 percent figure is a reduction in a model-implied objective, not an investment return, and reproducing it tells us the engine runs. It says nothing yet about whether the fuel, estimated forecasts, has any energy in it.

Read more

Let’s get into it.

The Hidden Cost of One-Period Thinking in Tactical Allocation

Institutional investors who re-optimize tactical bets one period at a time are quietly leaving roughly half of their achievable alpha on the table.

Kolm and Ritter tackle a governance habit that almost every large allocator follows: set the strategic policy, then let the tactical team re-solve a single-period problem each rebalance. That workflow looks disciplined, but it ignores a simple fact. Today’s trade becomes tomorrow’s inherited position, and the cost of moving it around is real. The authors show that the fully dynamic solution splits a question the one-period rule blurs together, namely where the portfolio should go versus how quickly it should get there. Signal persistence turns out to be the pivot. Below a specific threshold, short-lived alphas should be sized down relative to conventional tactical allocation (the trade lingers after the edge has faded), and above it, persistent views deserve larger positions than the standard rule prescribes.

Even more striking, in their three-asset example, a governance-compatible reformulation recovers about half of the value that repeated one-period optimization destroys, without dismantling the strategic/tactical split. For investors, the message is practical: the machinery you already use can be recalibrated to capture most of the benefit of a full dynamic model, and the usual instinct to match a “smarter” model coefficient-by-coefficient is the wrong yardstick, because economic value and coefficient proximity are not the same thing.

Kolm, Petter N. and Ritter, Gordon, From Strategic Policy to Tactical Trades: Dynamic Asset Allocation with Predictable Returns and Trading Costs (August 31, 2026). Available at SSRN: https://ssrn.com/abstract=7384838 or http://dx.doi.org/10.2139/ssrn.7384838

The Fed's 98% Talk: How Words Move Markets More Than Actions

The Fed’s most powerful tool isn’t rate changes, it’s the carefully worded hints about what might come next.

Cieslak, Hansen, and Pang dig through 43 years of FOMC transcripts (365 meetings from 1976 to 2019) to show that the Fed’s policy “tilt,” the forward-looking signal about where rates might head, functions as a distinct policy tool separate from the actual rate decision. Over half (51%) of policy statements in meetings are future-oriented rather than about the current move, and these tilts consistently predict rate changes up to a year ahead, even after controlling for the Fed’s own economic forecasts. The mechanism driving tilts is risk management: policymakers lean hawkish or dovish based on which mistake would be costlier to reverse, not just what they expect to happen.

The market impact is real and measurable. Hawkish tilts compressed the ten-year term premium by roughly 90 basis points from 1996 to 1998 while the funds rate barely moved, and the 2020 framework’s retreat from preemptive language coincided with the largest term-premium spike since 1994. For investors, this means parsing FOMC language (not just decisions) is essential for anticipating where long rates and risk premiums are heading.

Cieslak, Anna and Hansen, Stephen and Pang, Hao, Risk Management in Monetary Policy: A Review with Asset Pricing Implications (August 25, 2026). Available at SSRN: https://ssrn.com/abstract=7350618 or http://dx.doi.org/10.2139/ssrn.7350618

Your Factor Model Is Reading Yesterday's Newspaper

Fama-French factors are built on financial statements that are, on average, almost a year old, and fixing that staleness quietly rewrites thousands of alphas.

Bowles, Reed, Ringgenberg, and Thornock point out something hiding in plain sight: standard Fama-French factors reassign firms to portfolios only once a year, every June, using financial statements that are already months old. Rebuild the portfolios monthly using each firm’s most recent annual filing, and the average age of the underlying information drops from roughly 347 days to 203 days. HML (value) and CMA (investment) change the most because their inputs move fastest.

The interesting twist is that swapping fresh factors for stale ones does not clearly improve asset pricing. What does work is treating the “update return” (the difference between fresh and stale) as its own separate factor, because assets load on the slow-moving fundamental component and the fast-moving information component in different ways. The practical stakes are real: benchmark choice flips the sign of nearly 1,800 earnings-announcement CARs and reshuffles roughly 15% of top-decile mutual funds, meaning which managers look skilled depends partly on when you refresh the ruler.

Bowles, Boone and Reed, Adam V. and Ringgenberg, Matthew C. and Thornock, Jacob, Factor Time (August 31, 2026). Available at SSRN: https://ssrn.com/abstract=7384719 or http://dx.doi.org/10.2139/ssrn.7384719

Momentum on ETFs Doesn't Work. The Rules Around It Do.

The alpha in ETF momentum isn’t in the signal, it’s in the boring operational rules wrapped around it.

Across 3,215 U.S. ETFs and 25 years of weekly data, textbook momentum applied to ETFs is a disaster. The four standard lookbacks (1, 3, 6, and 12 months) deliver Sharpe ratios ranging from slightly positive to negative, with drawdowns between 69% and 82%, worse than simply holding SPY. What makes this paper interesting is what happens when the authors bolt on three unglamorous layers that any disciplined retail investor could implement: a technical filter to confirm the trend is real, mandatory diversification across asset-class buckets so you don’t accidentally hold three S&P 500 funds at once, and a 7% trailing stop on every position.

The same universe, same costs, same momentum ranking, now produces a Sharpe above 1.3 and a maximum drawdown of just 7%. Notably, the extra return isn’t the story; the strategy barely beats equities on absolute return. What it does is slash tail risk by roughly forty percentage points of drawdown. The one place it still breaks is when VIX pushes above 25, because diversification stops working when correlations converge. For investors, the reframe is uncomfortable but useful: stop hunting for the perfect signal and start auditing your exit discipline.

Magner, Nicolás and Sanhueza, Aliro Joel, Momentum Strategies in ETFs under Simple Operational Rules: Economic Value and Conditional Predictability. Available at SSRN: https://ssrn.com/abstract=7379989 or http://dx.doi.org/10.2139/ssrn.7379989

A Dollar Is a Dollar, Until You Read the Fine Print

Once households learn the institutional differences between digital currencies, they demand a much larger interest premium to hold stablecoins than tokenized bank deposits, and both more than a CBDC.

Castagnetti, Cillo, Gurrado, and Masciandaro ran a randomized experiment with 800 participants in France, Germany, and Italy to test whether people actually distinguish between the three main flavors of digital money: central bank digital currencies (CBDCs), tokenized commercial bank deposits, and non-bank stablecoins. Baseline beliefs treated all three as roughly interchangeable, which matches the “money is money” intuition most consumers carry around. But a short informational nudge about how each issuer is (or isn’t) backstopped produced sharply asymmetric updates: perceived default risk fell for CBDCs, stayed flat for commercial banks, and rose for non-bank issuers.

When the researchers translated choices into a structural model, informed participants demanded meaningfully higher interest to hold commercial-bank digital money over a CBDC, and a premium roughly three times larger to hold stablecoins over commercial-bank money. For markets, this suggests stablecoin yields may need to rise materially as retail financial literacy improves, and that CBDCs enjoy a real (not just rhetorical) trust advantage once institutional differences become salient.

Castagnetti, Alessandro and Cillo, Alessandra and Gurrado, Giuseppe and Masciandaro, Donato, Digital Money, Default Risk, and Financial Information: An Experiment on Europe (September 01, 2026). BAFFI Centre Research Paper No. 284, Available at SSRN: https://ssrn.com/abstract=7395438 or http://dx.doi.org/10.2139/ssrn.7395438

This week for paid subscribers

Paid subscribers are getting a look at why the H100 rental curve quit pricing obsolescence at the end of 2025 while the A100 curve never moved, and what that structural divergence reveals ahead of CME’s new compute futures launch on October 5. This post isolates chip-specific demand from general compute duration repricing, documents how H100’s 36-to-12-month term slope collapsed from -23% backwardation to flat across late 2025, and details the benchmark governance risks and data artifacts behind Silicon Data’s index. Python notebook and data included.

Disclaimer: The content provided in this newsletter, “Alpha in Academia,” is for informational and educational purposes only. It should not be construed as financial advice, investment recommendations, or an offer or solicitation to buy or sell any securities or financial instruments. Past performance is not indicative of future results. The financial markets involve risks, and readers should conduct their own research and consult with qualified financial advisors before making any investment decisions.

The interpretations, opinions, and analyses presented herein are those of the author and do not necessarily reflect the views of the original researchers, their institutions, or the full implications of the cited academic papers. While every effort is made to accurately represent the research discussed, readers should be aware that the summaries and interpretations may not capture the full scope or nuances of the original studies. The information contained in this newsletter is believed to be accurate and reliable at the time of publication, but accuracy and completeness cannot be guaranteed. The author and publisher accept no liability for any loss or damage resulting from reliance on the information provided.

This newsletter may contain links to external websites or resources. The author is not responsible for the content, accuracy, or reliability of these external sources.

By subscribing to or reading this newsletter, you acknowledge that you have read and understood this disclaimer and agree to hold the author and publisher harmless from any liability that may arise from your use of the information contained herein.

On October 5, 2026, CME will list two cash-settled compute futures on NYMEX, one on Silicon Data’s H100 rental index and one on the B200. Each contract represents a month of rent. Before they list, I wanted to know what the existing term structure for GPU rentals looks like, and whether it has been stable.

Silicon Data’s H100 term curve was backwardated by roughly 23% from twelve months to thirty-six months in April 2025, meaning long-dated compute was much cheaper than near-dated. By March 2026 that gap had closed entirely. The A100 curve, built by the same vendor using the same methodology, did not move. It hovered near -10% throughout and ended the sample at -8.7%. This change belongs to Hopper and was not a general repricing of compute.

Let's dive right in.

What is a Compute Future?

The thing being priced is rent on a graphics card. If you want to train or serve a model and don’t own hardware, you rent H100s by the hour from a cloud provider, and that hourly rate is the price. Silicon Data collects those rates across providers and publishes a daily index, in the same way that Platts publishes an assessed price for a crude grade nobody trades on a screen.

A futures contract on that index works like any other cash-settled future. Nobody delivers a GPU. Two parties agree on a price today for a contract that will settle against the published index over some future month, and whoever was on the right side collects the difference. CME’s two contracts each represent a month’s worth of rent on one card, one referencing the H100 index and one the B200. The final settlement mechanics are in CME’s contract notice and the exact averaging convention will matter for anyone trading them.

The reason to want this is straightforward on both sides. If you’re a startup whose costs are dominated by inference spend, your exposure to GPU rental rates is real and currently unhedgeable except by signing long reserved contracts, which locks up capital and commits you to a specific chip. If you’re a neocloud that has borrowed against a fleet you’re renting out at spot, you have the opposite exposure and no way to lock in revenue. Futures let both sides move that risk without touching the physical machines. Whether enough of them show up to make a liquid market is the open question, and October 5 is only the start of the answer.

How Compute is Different

A commodity forward curve usually has an anchor. If you can buy oil today, store it, and sell it forward, then the forward price cannot exceed spot plus storage and financing without someone taking the free money. Cash-and-carry pins the curve to physical reality.

Compute has no such anchor. You cannot store an idle GPU-hour and sell it in March. This puts compute in the same bucket as electricity, but with a complication power doesn’t have. Electricity’s generating fleet depreciates slowly and predictably. A GPU’s economic life is governed by when its successor ships, and the successor ships on a roughly annual cadence that everyone can see coming.

So the compute forward curve is doing something specific: pricing the expected obsolescence of a machine against the expected growth in demand for what it does. When obsolescence dominates, the curve slopes down. When scarcity dominates, it doesn’t. That ratio is not a fixed property of the asset, and the interesting question is whether it moves.

Data and Methodology

Silicon Data publishes a daily forward curve for H100, A100, and B200 out to thirty-six months, in two forms. The term rate is today’s locked-in price for a rental ending at a given tenor, built from observed contracts at standard lengths. The forward rate is the implied price of a short rental beginning at that tenor, derived from the term curve by differencing.

I pulled the numbers from Silicon Data’s portal: set the date, read the table, move on. That constrains me to the tenors the portal displays, which are twelve, twenty-four, and thirty-six months.

I sampled in two passes. The first walked the second of each month from January 2025 to September 2026. That pass put a sign change somewhere between the December and January snapshots, so a second pass filled in weekly observations through November and December 2025. The uneven spacing is a direct consequence of that order.

Two construction artifacts turned up in validation and both are worth naming. On January 2, 2025, every tenor carries an identical value on both rate types. Through March 2025 the twenty-four and thirty-six month points are identical, which means the long end was extrapolated flat rather than built from observed contracts at those lengths. Neither is a market observation. The usable sample starts in April 2025 and runs to September 2026, twenty-five as-of dates for H100. Throughout, I summarize curve shape as the thirty-six month rate divided by the twelve month rate, minus one. Negative is backwardation.

Results

Start with the shape of the curve itself, six weeks apart.

Here is the H100 term slope at monthly resolution, with the December weeks filled in.

On November 17, the term curve fell from $1.83 at twelve months to $1.47 at thirty-six. Six weeks later, on December 30, it rose slightly, $1.82 to $1.86. The forward curve made the same move harder, from $1.65 down to $1.08 in November and from $1.63 up to $1.93 by the end of December.

A regression of the term slope on time across the whole usable sample gives 19.6 percentage points a year, R² of 0.66, p below 0.00001. Splitting at December 15, the pre-period mean is −16.4% with a standard deviation of 5.2, and the post-period mean is −0.6% with a standard deviation of 3.1. The two ranges do not overlap, though only just: the highest pre-period reading is −7.3% and the lowest post-period reading is −7.0%.

The first column of the table is substantially less reliable than the second. The forward rate is produced by differencing the term curve, which amplifies whatever moves underneath. It shifts 13.3 percentage points between consecutive observations on average during 2025, against 3.6 for the term rate. It ranged from −3.9% to −35.6% within 2025 alone. The −3.9% reading on June 2, 2025 is barely more negative than the +0.6% on December 23 that sits after the supposed change. The forward curve moves in the same direction and further, and I would not quote its crossing date as if it meant anything precise. So the claim I’m willing to defend is about the term curve, where a −16% mean becomes roughly flat and stays there for nine months.

Read more

Let’s get into it.

The Pre-BOJ Drift

Japanese stocks do almost all of their work in the days before a Bank of Japan meeting, and none of it after.

Maeda runs the Lucca and Moench pre-FOMC test on Nikkei 225 and TOPIX data from 2009 to 2026, and Tokyo behaves the same way. Buying at the close three sessions before a scheduled policy meeting and selling on the morning of the decision earned about half a percent per meeting, and those windows, roughly one trading day in eight, produced nearly half the cumulative index return.

Figure 1: Cumulative return on the Nikkei 225 and TOPIX in the days around a BOJ decision, anchored three sessions before the meeting. The line climbs into the announcement (dashed) and goes flat afterward. Shaded bands are one standard error.

Volatility inside them was also not higher than on ordinary days, and the gain stops the instant the decision lands. What separates a large drift from none at all is fear going in. When the Nikkei volatility index is elevated three days out, the run up is big. And when markets are calm, it vanishes. The same pattern shows up in the yen but not in JGBs, so this looks like payment for holding risky assets into an uncertain event rather than anything about interest rates. The practical point is that a large share of Japanese equity return arrives on a schedule anyone can read off the BOJ calendar, and Maeda argues the effect is “a general feature of equity markets rather than a U.S.-specific anomaly.”

Maeda, Jun, The Pre-BOJ Announcement Drift: Evidence from Japanese Equity Indices (August 26, 2026). Available at SSRN: https://ssrn.com/abstract=7356758 or http://dx.doi.org/10.2139/ssrn.7356758

Only the Tail Survives

The oldest crash trade in the book mostly does not work, and most of its apparent edge is an artifact of counting the same bad day many times.

Large declines do not arrive independently. They land on the same handful of calendar dates, because whatever knocks one index down knocks everything else down that same afternoon. An event study that treats each stock day as a separate observation is counting one shock over and over, and the evidence looks decisive when it should not. Count each date once and it falls apart. In Dashyan’s panel of single names, a t statistic of 6.76 drops to 1.68.

Figure 2: Same events, two ways of counting. Treating each stock's bad day as its own observation puts the evidence far above the significance line. Counting each calendar date once drops it below. The single name panel (right) is where the illusion is largest, since 28 stocks can crash on the same afternoon.

Correct three other counting problems and exactly one result survives out of everything tested across US indices, individual stocks, and crypto perpetual futures. Index drops of seven percent or worse are followed by a strong next session, and most of that move happens after the open, so a real participant could capture it.

The catch is frequency. The signal has fired 25 times in 76 years and beats Treasury bills by less than half a point a year. For investors the transportable lesson is about evidence rather than crashes, since it is “a good idea that mostly does not work,” and plenty of published backtests look strong for exactly this reason.

Dashyan, Alexandr, The Tail Is the Only Signal: Flush Reversion in Equity Indices and Crypto Perpetual Futures (July 01, 2026). Available at SSRN: https://ssrn.com/abstract=7363482 or http://dx.doi.org/10.2139/ssrn.7363482

The Investment Grade Fire Sale

When bond ETFs were forced to sell during the COVID crash, the price damage landed on safe investment grade bonds, not on junk.

The intuition says forced selling hurts the least liquid assets first, so high yield should have cracked. Spoiler, it didn't. The authors tracked daily holdings for 135 corporate bond ETFs and find that bonds caught in ETF fire sales lost about 3 basis points of benchmark adjusted return in normal times, and roughly 15 more during the four weeks ending March 20, 2020. Almost all of that extra damage sits in investment grade.

Figure 3: Cumulative abnormal returns around ETF fire sales during the COVID crisis window. Solid lines are bonds hit by fire sales, dashed lines are bonds that weren't. Source: Gao, Huang, Qin and Wang (2026), SSRN working paper, not yet peer reviewed.

The explanation is behavioral rather than mechanical. Money left IG ETFs after IG ETFs fell, chasing the decline downward, which is the feedback loop that turns an ordinary selloff into a spiral. High yield investors did the reverse and bought into the dislocation. Once the Fed announced its corporate credit facility on March 23, the penalty disappeared entirely. The takeaway is that stress shows up where selling is easy, not where credit risk is highest. Quality bonds were the ones people could actually liquidate, which is precisely why they broke.

Gao, Xin and Huang, Jing-Zhi Jay and Qin, Nan and Wang, Ying, Fire Sales by Corporate Bond ETFs During the COVID-19 Crisis. Available at SSRN: https://ssrn.com/abstract=7370046 or http://dx.doi.org/10.2139/ssrn.7370046

Governing the Founder

Dual-class shares now account for roughly a third of U.S. IPOs, and the standard story that they exist to lock founders into their seats does not survive contact with the data.

The authors built a new database of U.S. IPOs since 2000 and find the share listing with unequal voting rights rose from roughly 10 percent in 2000 to about 35 percent in 2025, with nearly all the growth coming from venture-backed technology firms led by their founders. Founder-CEOs and founder-directors are much more common at these companies, yet the rest of the governance picture looks ordinary (board size, independence, and committee structure barely differ from single-class peers).

Figure 4: Founders with supervoting shares do not outlast anyone. The two lines track the odds that the CEO at listing is still in the chair years later, and the fact that they overlap is the evidence against the entrenchment story.

The entrenchment story runs into trouble on turnover. Dual-class CEOs do not stay in office longer, and their odds of being replaced still respond to poor performance. As the authors put it, “the question is perhaps not how long a person holds a role” but how much they can accomplish while in it.

Recent theory pushes further, showing that separating votes from cash flow rights can make control transfers easier rather than harder. Voting structure alone is a weak signal of governance quality, and treating every dual-class listing as an automatic discount ignores how unsettled the evidence on value still is.

Adams, Renée B. and Ferreira, Daniel, The Governance of Dual-Class Firms (August 27, 2026). Forthcoming in B. E. Eckbo (ed.), Handbook of the Economics of Corporate Finance, Vol. 2: Corporate Takeovers and the Market for Corporate Control (North-Holland/Elsevier, Amsterdam, Netherlands), European Corporate Governance Institute – Finance Working Paper Forthcoming, Available at SSRN: https://ssrn.com/abstract=7360478 or http://dx.doi.org/10.2139/ssrn.7360478

This week for paid subscribers

Paid subscribers are getting a look at why the option market's volatility forecast is sharp at one week and actively harmful at one quarter, and how a simple out-of-sample rescaling that strips out the variance risk premium turns a negative forecast score positive. This post separates bias repair from genuine predictive skill, tests the fix across stress regimes, and maps it onto volatility drag in leveraged and inverse ETFs. Python backtest code included.

Disclaimer: The content provided in this newsletter, "Alpha in Academia," is for informational and educational purposes only. It should not be construed as financial advice, investment recommendations, or an offer or solicitation to buy or sell any securities or financial instruments. Past performance is not indicative of future results. The financial markets involve risks, and readers should conduct their own research and consult with qualified financial advisors before making any investment decisions.

The interpretations, opinions, and analyses presented herein are those of the author and do not necessarily reflect the views of the original researchers, their institutions, or the full implications of the cited academic papers. While every effort is made to accurately represent the research discussed, readers should be aware that the summaries and interpretations may not capture the full scope or nuances of the original studies. The information contained in this newsletter is believed to be accurate and reliable at the time of publication, but accuracy and completeness cannot be guaranteed. The author and publisher accept no liability for any loss or damage resulting from reliance on the information provided.

This newsletter may contain links to external websites or resources. The author is not responsible for the content, accuracy, or reliability of these external sources.

By subscribing to or reading this newsletter, you acknowledge that you have read and understood this disclaimer and agree to hold the author and publisher harmless from any liability that may arise from your use of the information contained herein.

Today we’re looking at a number that claims to tell you how turbulent the next month will be. It is called the VIX, and it is one of the most-watched figures in finance. The question this piece asks is narrower and more useful than the usual VIX commentary: if you actually tried to forecast future market volatility with it, when would it help you, and when would it quietly lead you astray? The answer turns out to depend almost entirely on one thing you would not expect to matter so much. How far ahead you are looking.

Let’s dive right in.

A signal that points forward

Most financial data looks backward. Yesterday’s returns, last quarter’s earnings, the trailing average of almost anything. These are facts about what already happened, and the hope is that the recent past rhymes with the near future. Option prices are different. When traders buy and sell options on the S&P 500, they are placing bets about what the market will do between now and the option’s expiry. Bundle those bets together the right way and you get a forward-looking estimate of volatility, an expectation of turbulence extracted from where real money is being wagered right now. The VIX is exactly this: the market’s expected volatility over roughly the next month, distilled into a single number.

That sounds like it should dominate any backward-looking measure. Why average the last twenty days of market movement when you could read the crowd’s forecast of the next twenty directly? For short horizons, that intuition is right. The trouble begins when you ask the signal to reach further than it naturally sees.

To test this properly we need to be strict about one thing. Every forecast in this piece is made using only information that existed at the moment of the forecast. When we estimate how the option signal should be adjusted, we use only data from before that day. This is called an out-of-sample test, and it is the difference between a strategy that works and a strategy that merely looks good in hindsight. A model fitted on the whole history and then tested on that same history is grading its own homework. We never do that here.

The measuring stick

To judge any forecast we need a scoring rule and a baseline. The baseline is deliberately humble: a running historical average of realized volatility, the simplest honest forecast anyone could make. The scoring rule is called out-of-sample R-squared, and it answers a single question. Did this forecast produce smaller errors than the humble historical average?

A positive score means the forecast beat the running average. A score of zero means it merely tied. And a negative score, which will matter enormously in a moment, means the sophisticated forecast was worse than simply assuming the future looks like the recent past. Negative is not a weak positive. Negative means you would have been better off ignoring the fancy signal entirely.

We test three time horizons. One week ahead, roughly five trading days. One month ahead, twenty-two trading days, which is about where the VIX is designed to point. And one quarter ahead, sixty-six trading days, where we swap in a longer-dated cousin of the VIX called the VIX3M that targets three months instead of one. For each horizon we compare the raw option signal, a calibrated version of it we will build shortly, and a well-constructed forecast built purely from historical volatility.

Read the top row and the option signal looks brilliant. At one week, the raw signal scores 0.352, crushing the history-only model’s 0.114. At one month it still leads, 0.184 against 0.143. The forward-looking signal is doing exactly what forward-looking signals are supposed to do.

Now read the bottom row. At one quarter, the raw option signal scores minus 0.203. It did not just lose its edge. It became actively harmful, producing forecasts meaningfully worse than a running historical average that a person could compute on a napkin. The same signal that was the star performer at one week is a liability at one quarter. That reversal is the whole story, and it deserves an explanation rather than a shrug.

Read more

Let’s get into it.

Social Security's Debt Problem Isn't Aging, It's Architecture

Social Security’s growing contribution to the national debt has less to do with an aging population than with benefit design choices made decades ago.

Boccia and Nachkebia’s central claim is that demographics get too much blame. Yes, the worker-to-beneficiary ratio has collapsed from 16-to-1 in 1950 to roughly 3-to-1 today, but the deeper problem is that Congress built in structural cost growth on top of that: early cohorts got outsized windfalls relative to what they paid in, wage indexing (adopted in 1977) mechanically raises each new retiree’s initial benefit faster than inflation, and the inflation measure used for annual adjustments overstates true cost-of-living increases. The result is $1.5 trillion added to federal debt since 2010, with another $3.4 trillion projected by 2032.

The authors test whether stronger economic growth alone could dig the program out, and it can’t, even at real wage growth nearly double the baseline assumption, deficits persist through 2099. As they put it, “neither faster economic growth nor higher inflation can, by themselves, close Social Security’s funding gap.” For investors, that’s a signal that the fix will have to come through legislated benefit or tax changes, not a growth surprise, which has direct implications for long-duration Treasury demand and fiscal risk premia.

Boccia, Romina and Nachkebia, Ivane, Social Security's Role in the Federal Debt Explosion: Past, Present, and the Reform Imperative (August 19, 2026). Wharton Pension Research Council Working Paper No. 2, Available at SSRN: https://ssrn.com/abstract=7315578 or http://dx.doi.org/10.2139/ssrn.7315578

The Sector Rotation Everyone Trades Isn't Actually There

The classic growth-leads-defensive sector rotation, a staple tactical signal, turns out not to be a real cross-sector relationship at all, it’s the market factor reaching defensive stocks a day later than it reaches growth stocks.

Traders have long used the fact that cyclical sectors like tech and industrials seem to move a day ahead of defensive sectors like utilities and staples as a timing signal. This paper tests whether that lead-lag is a direct link between sectors or just two sectors reacting to the same market move at different speeds. Once they strip out the shared market factor, the rotation disappears almost entirely, and two placebo checks confirm this isn’t just an artifact of over-controlling. What’s left is a market factor that reliably leads defensive sectors, concentrated almost entirely in low-diversification, high-stress periods, and largely absent in calm markets.

As the authors put it, “it is not a direct causal signal.” For anyone running a rotation overlay, this means the strategy isn’t harvesting a diversifying edge, it’s a leveraged bet on market direction that only pays off during stress, and should be sized and risk-managed accordingly.

Sudjianto, Agus and Setiawan, Sandi and Narain, Arpit, Is Sector Rotation Causal? A Geometric Test of the Growth-to-Defensive Lead-Lag (August 18, 2026). Available at SSRN: https://ssrn.com/abstract=7313339 or http://dx.doi.org/10.2139/ssrn.7313339

When Everyone Trades the Same Signal, the Market Can Get Stuck With Two Right Answers

Two portfolios can crowd each other out just by watching the same data, even if they never touch the same stock, and past a measurable tipping point the market stops having one correct price and starts supporting several self-fulfilling ones instead.

This paper draws a distinction most crowding research skips: there’s crowding from holding the same positions (which market impact already prices) and crowding from conditioning on the same drivers, like everyone running a regression on the same handful of macro series. That second kind erodes returns even when portfolios never overlap in what they actually own. The more interesting result is what happens once enough capital piles onto correlated signals: the author derives a single dimensionless statistic, built from cross-impact, covariance, and deployed capital, that tells you which regime the market is in. Below a threshold of one half, prices are pinned down uniquely. Above it, the same fundamentals can support a whole family of self-confirming “conventions,” essentially coordinated mispricings that persist because everyone trading them keeps them true.

The author’s own S&P panel test finds a modest but real signature of this crowding, stronger when borrowing costs (a proxy for short-selling pressure) are elevated. For investors, it’s a formal argument for why popular factors can go quiet for years and then break down all at once: the market isn’t drifting, it’s approaching a threshold.

Rodriguez Dominguez, Alejandro, The Market's Conditioning Representation: Equilibrium, Crowding, and Convention Multiplicity (August 09, 2026). Available at SSRN: https://ssrn.com/abstract=7309320 or http://dx.doi.org/10.2139/ssrn.7309320

VIX Is a Great 22-Day Forecaster and a Mediocre 66-Day One (And Recalibration Only Half-Fixes It)

Goyle and Revtsov ask a deceptively simple question: does the options market’s forward-looking volatility estimate actually help predict how much S&P 500 leveraged ETFs will drift from their stated daily multiple over time. The answer depends entirely on how far out you’re looking. At 5 and 22 trading days, raw VIX is a genuinely strong predictor of realized variance, easily beating both a historical average and a HAR-style model built from past volatility alone. But stretch the horizon to 66 days (roughly a quarter) using VIX3M, and the raw signal collapses, actually performing worse than just guessing the historical mean. The fix is a standard statistical recalibration that rescales the options signal to match its historical relationship with realized outcomes, which restores meaningful forecasting skill.

Even recalibrated, though, the options-based forecast doesn’t significantly beat a purely backward-looking model at this longer horizon, so its edge is really about fixing a scaling problem, not revealing hidden information. For a leveraged-ETF holder or risk manager, that’s the practical takeaway: the market’s implied volatility is a genuinely useful monthly gauge, but treating it as gospel for quarterly risk monitoring without adjustment can be actively misleading.

Goyle, Kartikay and Revtsov, Yevgen, When Does Option-Implied Variance Add to Physical Forecasts? Evidence from S&P 500 Leveraged ETF Drag (August 21, 2026). Available at SSRN: https://ssrn.com/abstract=7325618 or http://dx.doi.org/10.2139/ssrn.7325618

Your DCF Model Is Lying to You About Frontier Markets (Here's a Fix Built on Bayesian Priors and Currency Jumps)

A frontier-market investor doesn’t actually want one number, they want a map of everything that could plausibly happen to their investment, and this paper builds the machinery to draw that map.

This is less an empirical finding and more a proposed toolkit, which changes how we should treat it. The author’s argument is that standard discounted cash flow analysis quietly assumes a stable, well-behaved world (smooth currency moves, independent shocks, roughly normal outcomes) and that assumption simply breaks in frontier markets, where currencies sit still for years and then devalue all at once.

The fix combines three ingredients: a Bayesian approach that lets you blend thin historical data with expert judgment when estimating growth, a jump-diffusion model that treats currency crises as sudden, discrete events rather than smooth drift, and a Student-t copula that captures how FX shocks, inflation, and risk premiums tend to spike together rather than independently. Feed all of that into a Monte Carlo simulation and instead of one DCF number you get a full distribution of possible values, with an explicit downside case.

For investors, the appeal isn’t a better point estimate, it’s an honest accounting of tail risk that a textbook DCF simply ignores, or as the author puts it, uncertainty here should be “weaponized... into a competitive advantage.”

Chipili, Rackson, A Bayesian and Monte Carlo Framework for Equity Valuation in Frontier Markets (August 20, 2026). Available at SSRN: https://ssrn.com/abstract=7318419 or http://dx.doi.org/10.2139/ssrn.7318419

This week for paid subscribers

Paid subscribers are dissecting quarter-end funding pressure in repo markets, measuring dislocations in the 99th percentile versus the quoted SOFR median, and testing how bank reserve scarcity drives turn-date spreads. This post covers stripping ambient funding levels from turn windows, why the standard level measure yields a statistical null while tail dispersion more than doubles, and how the true 2019 peak hit in June rather than September. Python Code included.

Disclaimer: The content provided in this newsletter, “Alpha in Academia,” is for informational and educational purposes only. It should not be construed as financial advice, investment recommendations, or an offer or solicitation to buy or sell any securities or financial instruments. Past performance is not indicative of future results. The financial markets involve risks, and readers should conduct their own research and consult with qualified financial advisors before making any investment decisions.

The interpretations, opinions, and analyses presented herein are those of the author and do not necessarily reflect the views of the original researchers, their institutions, or the full implications of the cited academic papers. While every effort is made to accurately represent the research discussed, readers should be aware that the summaries and interpretations may not capture the full scope or nuances of the original studies. The information contained in this newsletter is believed to be accurate and reliable at the time of publication, but accuracy and completeness cannot be guaranteed. The author and publisher accept no liability for any loss or damage resulting from reliance on the information provided.

This newsletter may contain links to external websites or resources. The author is not responsible for the content, accuracy, or reliability of these external sources.

By subscribing to or reading this newsletter, you acknowledge that you have read and understood this disclaimer and agree to hold the author and publisher harmless from any liability that may arise from your use of the information contained herein.