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.

Today we are looking at quarter-end funding pressure, and at whether it shows up in the rate everyone uses to measure it. Across the twenty-five ordinary quarter-ends in the SOFR record, the median spread between SOFR and the policy floor on the turn date is exactly zero basis points. Against ordinary month-ends the difference is 3.01 basis points with a p-value of 0.217. On the number that gets quoted, quarter-end is not a thing.

The same days, measured in the upper tail of the same distribution, look completely different. A quarter-end roughly doubles the dislocation there, and unlike the level result, it survives dropping September 2019. It scales with reserve scarcity: each percentage point lower on reserves as a share of bank assets raises it about 39%. The largest reading in the sample is not September 2019 either. It is June 2019, when the quoted rate ranked it sixth of twenty-five, unremarkable.

Let’s dive right in.

Introduction

The standard account of quarter-end in funding markets is that balance sheet constraints bind on reporting dates, dealers pull back from intermediating repo, and the overnight rate spikes. September 2019 is the canonical illustration. The mechanism is right. The question is where you can see it.

SOFR is a volume-weighted median. The New York Fed publishes it alongside the first, twenty-fifth, seventy-fifth, and ninety-ninth percentiles of the same day’s transactions, and the median is the only one of those numbers that gets quoted. On a normal day the choice does not matter much. On a turn date it matters enormously, because the marginal borrower who cannot find balance sheet is not transacting at the median. They are transacting in the tail, and the tail is a different series with different behavior.

I want to be careful about what follows. This is not a trading strategy, and there is no instrument at the end of it. It is a measurement argument: a widely discussed phenomenon has been evaluated with the wrong statistic, and the right statistic tells a cleaner story about reserve scarcity than the wrong one does.

Data and Methodology

Everything here comes from two sources. The New York Fed’s markets API publishes SOFR, the tri-party and broad general collateral rates, and the effective fed funds rate, each with the full percentile distribution and daily volume, from 3 April 2018. FRED supplies interest on reserves, overnight reverse repo balances, reserve balances, the Treasury General Account, and total commercial bank assets. No API key is required for either.

Two construction choices drive most of what follows, and both are worth stating plainly.

Turn dates come from the observed rate calendar, not a calendar offset. Roughly a third of quarter-ends fall on a weekend and settle on the prior business day. Using MonthEnd or QuarterEnd offsets misaligns those, and since the effect is concentrated in a two or three day window, misalignment destroys it. I label every trading day by its position in the published SOFR series and define the turn window as one business day either side of the last observed trading day of the period.

The premium is measured relative to the ambient level, not to the policy floor. This one changed the whole analysis. When reserves are scarce, SOFR trades above the floor every day of the month, not only at turns. A model that regresses the raw turn-date spread on reserve scarcity will score well by predicting the ambient level while explaining nothing at all about the turn. So for each turn I compute the median spread over a reference window spanning twenty to five business days before and five to twenty business days after, and subtract it. What remains is the part specific to the turn date.

The correlation between the raw turn-date spread and the ambient level is 0.778. Most of what looks like quarter-end pressure in the unadjusted series is simply the funding regime you happened to be in that quarter.

The sample runs from April 2018 to August 2026 and contains twenty-five ordinary quarter-ends, sixty-seven ordinary month-ends, and eight year-ends. Year-ends are held out of every model and reported separately, because G-SIB scoring is a point-in-time measurement on 31 December and pooling them contaminates both groups. Twenty-five observations is a small sample, and I will come back to what that rules out.

Predictors are read five business days before each turn, using only vintages published by then. Reserve balances and bank assets are weekly for the week ending Wednesday and appear in the H.4.1 the following Thursday, so the code enforces that Wednesday W is not available until W+1 rather than merging on nearest date. The reserve ratio is reserve balances as a percentage of total commercial bank assets, which ranges from 7.99% at the September 2019 blowup to 19.26% at the peak of the abundant-reserve period, and sits at 11.48% today (the most recent reading).

Results

Here is the level measure, and it is a null result.

The turn-specific excess averages 7.58 basis points at quarter-ends and 4.57 at ordinary month-ends. The difference is 3.01 basis points with a Welch p-value of 0.217 and a Mann-Whitney p-value of 0.177. Neither test comes close to conventional significance. On the raw unadjusted spread the picture is worse: the quarter-end mean is 2.16 basis points, the median is exactly zero, and twelve of the twenty-five quarter-ends printed a negative spread, with SOFR below the floor.

Pooling quarter-ends and month-ends into a single regression with a quarter-end dummy, controlling for the reserve ratio and overnight reverse repo balances, gives a dummy of +3.33 basis points at p=0.034. That looks like a result until you remove September 2019, at which point it falls to +1.86 and p=0.163.

Fitting the excess on reserve scarcity gives an in-sample R-squared of 0.346 and a leave-one-out R-squared of 0.091. Removing September 2019 takes the reserve coefficient from −2.02 to −0.91 and its p-value from 0.014 to 0.136. Restricting to 2020 onward, which is the regime anyone would actually care about, the leave-one-out R-squared goes negative, at −0.195. Worse than predicting the sample average.

September 2019 has a Cook’s distance of 1.011 and a studentized residual of 4.76. Those are not the numbers of an influential observation. They are the numbers of a different data-generating process that happens to be sitting in the sample.

The level measure says quarter-ends are not special, the relationship with reserves is one repo crisis, and nothing here supports a forecast. But the percentile columns say otherwise.

Read more

Let’s get into it.

Fifteen Minutes Before the Close

The 4:00 PM close, the timestamp the entire derivatives industry marks its books against, has quietly become the worst fifteen minutes of the day to measure volatility.

Two identical measurements, fifteen minutes apart. The researchers built the same at-the-money, shortest-maturity implied volatility measure twice a day, once at 3:45 and once at the bell, using identical filters. The 3:45 version produces a usable number on 99.8% of trading days. The closing version works on 36.5%.

Figure 1: Correlation between each implied volatility signal and the next day's realized variance, by year. Same options, same selection rules, fifteen minutes apart.

Blame 0DTE options, which now account for much of the near-the-money volume and carry essentially zero remaining time value by market close, leaving the conversion from option price to implied volatility unstable or outright impossible. A wide gap in usefulness follows. Next-day realized volatility barely responds to the raw closing series, while the 3:45 series moves with it. On days when the close does return a real figure, that figure predicts perfectly well, which points to a broken thermometer rather than a market with nothing to say. Running the whole surface through a convolutional network adds accuracy, though a plain EGARCH stays annoyingly hard to beat. In the authors' words, “a later timestamp is not necessarily a better volatility signal.” Risk systems and valuation models fed by closing marks absorb that noise daily.

Clark, Brian J. and Palepu, Sai and Potì, Valerio and Siddique, Akhtar R., Last Fifteen Minutes: Equity Options Volatility at the Close (August 13, 2026). Available at SSRN: https://ssrn.com/abstract=7279624 or http://dx.doi.org/10.2139/ssrn.7279624

Borrowed Elasticity

Hedge funds hardly react to what it costs them to borrow a bond, because the size of the position was settled somewhere else entirely.

Insurers and pension funds routinely want more government bonds than actually exists. The gap gets filled by hedge funds, who sell the bond short and borrow it in the repo market so they can deliver it. Working from regulatory data on every repo backed by German government debt, the authors show that these borrowers barely respond to the price of borrowing. Push the borrowing cost up 10% and their borrowing falls roughly 1%.

Figure 2: Same investors, opposite behavior. Each dot is one type of institution, plotted by how much its bond buying responds to price against how much its bond borrowing responds to price. Hedge funds, the bulk of the “foreign” dot, sit in the bottom right corner.

Strange, given that the same funds are among the twitchiest buyers in the cash bond market. What explains it is that the repo leg was never a decision in the first place. It is machinery supporting a short whose size somebody else's appetite for the physical bond had already determined. The adjusting happens on the lending side instead, largely at the German debt office and the ECB. Investors can take two things from this. Repo specialness (what you pay to borrow one particular bond) doubles as a real time pressure gauge on the cash market, and whoever sets the marginal price in Europe's safe asset funding market is sitting offshore, well past the reach of any European supervisor.

Poinelli, Andrea and Pelizzon, Loriana and Tomio, Davide and Nguyen, Benoît and Linzert, Tobias, Elastic in Cash, Inelastic in Repo: Hedge Funds in the Treasury and Repo Markets (August 07, 2026). SAFE Working Paper No. 492, Available at SSRN: https://ssrn.com/abstract=7258361 or http://dx.doi.org/10.2139/ssrn.7258361

Funding-Technology Feedback in the AI Buildout

The AI financing loop is close to the point where a shock stops fading and starts feeding itself, and the weak link is the leveraged miners, not NVIDIA.

Cao and Huang model the 2026 compute buildout (NVIDIA funding OpenAI, OpenAI committing to Oracle capacity, bitcoin miners pivoting into GPU colocation) as a circuit where a funding freeze blocks the next hardware refresh, obsolescence craters the collateral behind the debt, and the freeze deepens.

It reduces to one number: how much distress returns to a borrower after one lap around the loop. Below one it dies out, above one it compounds. Their reference scenario lands at 0.98, and two defensible corrections (refusing to count intra-loop revenue as a real buffer, adding idle capacity from weak demand) push it past 1.4.

The useful part is where the fragility sits. Cutting NVIDIA out of the graph barely moves the index, removing the capital-constrained miner cohort moves it a lot, and netting every bilateral exposure does almost nothing, because the binding loop lives inside one balance sheet rather than between two.

Figure 3: Removing NVIDIA from the network barely changes the system's fragility score. Removing the leveraged miner cohort does.

The authors call these “scenario-conditioned structural diagnostics,” not measurements. Still, the watch list they imply is utilization and the weakest GPU-backed borrowers, not the vendor everyone already monitors.

Cao, Zeyu and Huang, Shaosai, Stability of Compute-Capital Networks: Funding-Technology Feedback and Scenario Diagnostics (July 25, 2026). Available at SSRN: https://ssrn.com/abstract=7295260 or http://dx.doi.org/10.2139/ssrn.7295260

ESG Ratings vs. Portfolio Decarbonization

ESG ratings tell you almost nothing about which companies in a sector actually emit less per dollar of revenue.

Hwang and Patatoukas rank S&P 500 firms against their own sector peers on both ESG scores and carbon intensity (emissions per dollar of revenue), and find the two rankings barely relate. The environmental pillar, which you would expect to be the exception, tracks the composite score so closely that it is effectively the same measure. What drives that pillar explains why: two process indicators, one covering the quality of environmental disclosure and one covering how climate risk is framed in strategy, account for most of the variation. Firms are scored on how well they report and position, not on what they emit.

Figure 4: Average annual financed emissions per $1M invested, 2017 to 2024. Recreated from Table 10 (Panel A) of Hwang and Patatoukas (2026). Both tilted indices hold the same sectors in the same proportions as the S&P 500; only the within-sector weights change.

The part that should worry investors is what this does to a portfolio. An index tilted toward carbon-efficient firms cut financed emissions by 43% and matched the market's return. An index tilted toward high ESG scores raised emissions by 10% instead, because the highest scorers tend to be the biggest companies in each sector, and bigger companies emit more in absolute terms regardless of efficiency. As the authors put it, “sustainability ratings and sustainability performance are not the same thing.” If you hold an ESG fund for climate reasons, the label and the outcome are separate purchases.

Patatoukas, Panos N. and Hwang, Jinsung, ESG Ratings Undermine Portfolio Decarbonization. Available at SSRN: https://ssrn.com/abstract=7301860 or http://dx.doi.org/10.2139/ssrn.7301860

This week for paid subscribers

This week for paid subscribers: Paid subscribers are replicating the 2024 Polymarket lead-lag, rebuilding the 34-asset Trump-trade portfolio and testing whether prediction market moves predicted next-day returns in banks, rates, and FX. This post covers turning a noisy directional signal into a percentile-ranked position, a placebo panel that rules out broad equity beta, and where the replication diverges from the paper. Python backtest code included.

Alpha in Academia

The Odds Lead the Tape

Read more

a month ago · 7 likes · Alpha in Academia

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 taking inspiration and guidance from a new paper by Goldstein, Li, and Wang. This explores the 2024 U.S. presidential election, where changes in Polymarket’s Trump probability predicted next-day returns on Trump-sensitive assets by roughly 13 basis points per percentage point of movement, with a simple long-short strategy earning a Sharpe of nearly 2. We tested it. The lead-lag is real, the placebo is clean, and the story it tells about prediction markets is more interesting than the trading signal it produces.

Let’s dive right in.

The Setup

Prediction markets are supposed to be information aggregators. People bet real money on future outcomes, and the resulting prices, in theory, distill dispersed information into a single number. That number is useful only to the extent that other people look at it and act on it. This paper asks a specific version of the question about whether people look at it. During the 2024 U.S. presidential election, did traders in traditional financial markets watch Polymarket, and did they trade on what they saw?

The authors examine daily changes in the Polymarket Trump-YES probability and test whether those changes predict next-day returns on a portfolio of assets that market commentary flagged as sensitive to Trump’s electoral prospects. Their answer is yes. A one percentage point increase in Trump’s implied probability was associated with about 13 basis points of next-day return on the Trump-trade basket, statistically significant and economically meaningful.

The identification challenge is the usual one. Both markets might be reacting to the same underlying news, with Polymarket happening to move first. To distinguish real cross-market learning from sequential news arrival, the authors exploit on-chain wallet-level data to classify individual Polymarket traders as informed or uninformed based on their post-trade profitability. They then show that price impact from uninformed trades also propagates to traditional markets before partially reversing. Since noise cannot reflect fundamental information, its transmission establishes that traders in equities and currencies are genuinely extracting signals from Polymarket prices, not just responding to the same news feed at a lag.

For our test we focus on the price-level analysis in the paper’s Section 3. Wallet-level classification requires pulling and processing the full universe of on-chain Polymarket transactions, which is a separate exercise. What follows uses public price data from the Polymarket CLOB API and daily returns from yfinance for the 34 signed Trump-trade assets.

Figure 1. Polymarket-implied probability of a Trump win, January through November 2024. Key events annotated.

The Trump Trade Portfolio

The paper’s Trump-trade portfolio is not a factor model. It is a narrative-based classification. The authors reviewed contemporaneous financial media, primarily Bloomberg, the Wall Street Journal, and the Financial Times throughout 2024, and cataloged the assets that analysts and commentators repeatedly identified as exposed to Trump’s electoral prospects. The result is a portfolio of 34 assets across five categories: broad equity ETFs, bond ETFs, six large bank stocks, currencies and commodities, and a small “Connected” set consisting of Trump Media, Phunware, and Tesla.

Each asset gets a directional sign based on whether it would benefit or suffer from a Trump victory. Bank stocks and equity ETFs go long on the expectation of tax cuts and financial deregulation. Treasury bonds go short on the expectation of fiscal expansion and inflation. The dollar goes long against foreign currencies, reflecting tariff policy. Bitcoin, Ethereum, and gold go long. The three Connected names go long as direct campaign-linked equities.

Once we sign each asset’s return in the expected direction, the paper’s construction implies that if the market genuinely tracks Trump’s odds, all five categories should trend up together as those odds rise, and down together as they fall. This is exactly what we see in the data.

Figure 2. Cumulative signed returns by asset category. Each category is equal-weighted across its constituents.

The Connected category is extraordinary, mostly driven by DJT’s post-merger volatility, but the more instructive pattern is the broad co-movement of the other four categories through the second half of the year. Bank stocks, equities, and the dollar basket climb together as Trump’s implied probability rises from roughly 45 percent in April to over 60 percent by early July, then move sideways after the Biden withdrawal, then accelerate together into the election. Bonds are the mirror image, drifting slightly negative on the signed basis. This is a real portfolio-level exposure to a single political factor, and it makes the lead-lag test meaningful.

Read more

Let’s get into it.

Private Equity's "Free Lunch" Was Just a Pricing Illusion

When you value buyout funds at real market prices instead of sponsor-reported estimates, their famous risk-adjusted outperformance disappears entirely.

Private equity has long sold itself as the rare asset that beats stocks while smoothing out the ride, low volatility, low correlation, better returns. This paper tests that claim using a clever workaround: a set of buyout funds that trade on European stock exchanges, giving researchers both the official NAV (the fund’s own periodic self-appraisal) and the actual price investors are willing to pay for the same assets, every day. The gap is striking. Priced at NAV, these funds look tame, with volatility close to public stocks. Priced at market, volatility jumps to 29%, correlation with stocks climbs to 0.94, and beta lands around 1.5, meaning these funds are about 50% more volatile than the market, not less. Once that real risk is accounted for, the outperformance vanishes, with alpha statistically indistinguishable from zero.

The authors frame it plainly: NAV-based accounting lets buyouts “appear to generate significant alpha” that isn’t really there. For investors leaning on private equity as a smoother, higher-returning complement to stocks, this is a reason to check whether that cushion is real or just an artifact of how infrequently the assets get marked to market.

Ennis, Richard and Rasmussen, Daniel, Buyout Performance with Assets Valued at Market (July 01, 2026). Available at SSRN: https://ssrn.com/abstract=7157298 or http://dx.doi.org/10.2139/ssrn.7157298

Is Volatility Really "Rough"?

A new model suggests the popular “rough volatility” framework may be forcing two separate questions, how choppy volatility looks up close and how long its memory lasts, into a single number, and separating them changes the picture.

For the past decade, quants have modeled market volatility as “rough,” meaning it looks jagged and unpredictable at short timescales, using a single parameter (the Hurst index) borrowed from fractal math. This paper argues that parameter is secretly doing two jobs at once, setting both how volatility scales over time and how much it remembers its own past, when those are logically different properties. Borrowing a tool from physics (originally used to model particles bouncing around in fluids), the author builds a more flexible framework that lets memory and scaling move independently. Testing it on real order book and stock data, two of the model’s predictions hold up clearly: volatility’s memory decays slowly rather than instantly, and there’s a measurable asymmetry where past price moves predict future volatility more than the reverse.

The short-term “roughness” question, though, turns out to be essentially unmeasurable with current data, neither confirmed nor ruled out. For traders and risk modelers, this matters because it suggests some volatility models may be more constrained than the data actually requires, and that memory in markets is a real, testable phenomenon rather than just a curve-fitting trick.

Itkin, Andrey, Beyond Rough Volatility: Decoupling Memory and Scaling via a Generalized Langevin Equation (July 29, 2026). Available at SSRN: https://ssrn.com/abstract=7202798 or http://dx.doi.org/10.2139/ssrn.7202798

When the Spread Doesn't Come Back: The Math of Knowing When to Stop

Capping your position size in a pairs trade doesn’t just limit your losses, it actually changes the optimal trade itself, because a smart investor starts hedging against future limits before they ever get hit.

Pairs trading lives and dies on one assumption: that two related stocks, after drifting apart, eventually snap back together. This paper asks the uncomfortable question that assumption usually skips over, what happens when they don’t. The author builds a formal model where a trader sets hard position limits on both legs of the trade, then solves for the mathematically optimal strategy under those limits. The twist is that this constrained strategy isn’t just the unconstrained strategy clipped at the edges. Anticipating that limits might bind later, the optimal trader adjusts positions earlier than you’d expect, even while still comfortably within bounds.

Tested on Ford and GM stock from 2022 to 2024, a period where their prices diverged and stayed diverged, the constrained strategy lost about 70 dollars per unit of capital versus roughly 280 for the unconstrained version. The takeaway for anyone running a spread trade: sizing discipline isn’t just risk management bolted on afterward, it should shape the trade from day one.

Chen, Ziyi, Optimal Pairs Trading with Position Constraints (July 30, 2026). Available at SSRN: https://ssrn.com/abstract=7205360 or http://dx.doi.org/10.2139/ssrn.7205360

Can the Shape of a Probability Curve Predict a Crash Before Volatility Does?

A new early-warning model that reads the geometry of market volatility, not just its size, flagged the COVID crash and the 2022 rate-hike selloff an average of 18 days before a standard volatility filter did.

Most volatility models, including the classic Markov regime-switching approach used across the industry, work by waiting for enough big price swings to pile up before declaring that markets have shifted into a stressed state. That’s inherently reactive. This paper tries something different, borrowing a tool from topology (the math of shapes and connectivity) to track how the pattern of a volatility signal reorganizes itself in the days before a real shift, not just how big it gets. Applied to JPMorgan stock and the S&P 500 from 2020 through 2024, this topological layer detected the two biggest volatility events of that period nearly three weeks earlier than the standard model, and the signal was statistically unrelated to VIX or realized volatility, meaning it’s genuinely picking up something different rather than just repackaging existing data.

The catch is that this early-warning system throws a lot of false alarms, roughly seven flagged windows out of ten turn out to be nothing, so it’s built to work as a tripwire that prompts a closer look, not a system that trades on its own. For risk managers and active investors, that tradeoff, faster warning bought with more noise, is worth understanding before leaning on any signal that claims to see trouble coming early.

Faris, Mahrus, Early-warning Volatility Regime Detection in Equity Markets: A Combined Markov Switching and Persistent Homology Approach (July 20, 2026). Available at SSRN: https://ssrn.com/abstract=7206123 or http://dx.doi.org/10.2139/ssrn.7206123

This week for paid subscribers

Paid subscribers are watching the G10 currency carry trade erase two decades of calm-regime gains across high-volatility selloffs, then testing whether an implied equity volatility filter can predict those crash regimes in advance. It isolates the funding-driven unwinds that destroy the trade and is blind to the basket’s own realized volatility, with a simple 80th-percentile threshold turning a zero-Sharpe basket into a 0.32 net Sharpe. 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 are looking at the G10 currency carry trade, and at the ETF that existed to sell it to retail investors between 2006 and 2023. Rebuilt from free data, the basket returned 0.74% annualized over nineteen and a half years on 9.8% volatility, which is a Sharpe ratio of 0.07. Taking realistic costs into consideration, it returned nothing at all.

That number is an average of two regimes that happen to cancel. In the calmest fifth of the sample, the basket earned 4.8% annualized, whereas in the most stressed fifth, it lost 15.6%. The stressed regime can be identified in advance, but only with implied equity volatility. The basket’s own realized volatility tells you nothing useful, and realized equity volatility gets you about half way. There is also a construction quirk in the original index that turns out to have been hedging the crash risk by accident.

Let’s dive right in.

Introduction

Carry is the oldest trade in currency markets. Borrow where rates are low, lend where they are high, keep the spread. Uncovered interest parity says the high-yielding currency should depreciate by exactly the interest differential and leave you flat, and it does not, which is why the trade has a forty-year academic literature behind it.

The Deutsche Bank G10 Currency Future Harvest Index formalized it about as plainly as possible. Rank the G10 currencies by yield, go long the top three at a third of NAV each, short the bottom three the same way. Gross notional of 200%, rebalanced quarterly. The index was calculated back to March 1993 at a base of 100, and by July 25, 2007, it stood at 315.27.

In September 2006, an ETF launched to track it: DBV, the Invesco DB G10 Currency Harvest Fund. It ran for a little over sixteen years and was liquidated on March 10, 2023.

So the index tripled, then the product arrived, then nothing happened for sixteen years. That sequence is what this post is about. The question is not whether the carry premium exists in the data, because it does. The question is what it did during the only window in which an ordinary investor could have bought it.

Data and Methodology

Daily spot rates for the nine non-USD G10 currencies from Yahoo Finance, normalized to USD per unit of foreign currency so a rise always means the foreign currency strengthened. Three-month interbank rates from FRED’s OECD series. DBV, VIX and S&P 500 history are from Yahoo.

The sample runs from June 2006, where AUD spot history begins, to December 2025, where every currency still has published rate coverage.

Three construction choices, all taken from the fund’s final 10-K:

1. Ranking uses the previous month’s rate observation, lagged so that nothing in a given month depends on data published during it. The index actually ranked on a currency carry ratio (front-month futures over the three-month futures) rather than cash rates, and rebalanced quarterly rather than monthly. Interbank rates are observable live, so my lag is stricter than it needs to be.

2. The dollar is ranked but never traded. When USD lands in the top or bottom three, that leg is simply not established and gross exposure falls to about 1.67:1. This happened in 65.4% of months in the sample, which is more often than I expected, and it matters later.

3. Returns are excess returns: the spot move plus the interest differential against USD. DBV shareholders received excess return plus collateral income minus 0.78% in fees, so the comparison against the fund adds those back.

The Reconstruction Tracks the Fund

Before trusting any of this it needs to match the thing it claims to replicate.

Figure 1: The reconstruction net of fees against DBV’s actual returns over the fund’s listed life. Daily correlation is 0.29, which looks alarming until you notice it is a clock problem rather than a disagreement.

Over the fund’s life, the reconstruction returned −0.25% annualized against DBV’s −0.34%, on volatility of 10.5% against 12.1%, with daily skew of −0.64 against −0.86.

The daily correlation of 0.29 is a measurement artifact. Yahoo’s spot quotes are a 24-hour snapshot taken at an arbitrary time; DBV was a 4pm Arca close on futures that settled at 2pm. Different clocks. Aggregate the returns and the gap closes: 0.61 weekly, 0.88 monthly, 0.90 quarterly.

Twenty Years of Nothing

Here is the full sample.

Read the columns from left to right. The first is fourteen months, and it sits on the terminal blow-off of the largest carry run in modern history, so I am not going to use it for anything. The middle column is the fund's entire listed life. The one to its right is what happened after it closed: 1.4% annualized at a Sharpe of 0.26, and 2.3% at 0.40 once the accidental dollar position comes out.

Carry did not stop paying. It paid on either side of the window in which you could buy it. This shows that it is not underperformance against a benchmark. Over nineteen and a half years, the trade produced nothing, while taking a 37% drawdown and carrying negative skew the whole way. You held a left tail and were paid zero for it.

DBV tracked that faithfully. The fund was not the problem, and the fees were not the problem either. The strategy did this.

Figure 2: The reconstructed basket over the full sample. The shaded band is DBV’s entire life on the exchange, and the dotted line is the index’s all-time high in July 2007.

Zero is an Average

A twenty-year Sharpe of 0.07 invites the conclusion that carry stopped working, but I do not think that is what happened. Sort every day by where the VIX closed, cut it into quintiles, and the average falls apart immediately.

The premium is not missing. It is large, and then it is violently negative, and the two cancel. Four fifths of the sample pays, with the top fifth taking it all back. Volatility triples from Q1 to Q5, so the losses arrive levered as well as late.

The relationship is not monotonic, which is worth flagging rather than smoothing over. Q2 sits below Q1 in both the version I am showing and in the balanced version. The reliable statement is that the top two quintiles are where carry dies.

This is what “carry is short volatility” means in practice, and the claim deserves a proper test rather than an assertion. Regressing monthly returns on the monthly change in VIX gives a slope of −0.0017 with a t-statistic of −7.0. Adding a quadratic term, which tests whether the payoff bends the way a short option position does, gives a curvature coefficient with a t of −2.8 and adds 2.7 percentage points of R². Significant, and modest. The strong version of the short-volatility claim survives the test without running away with it.

Figure 3: Each point is one month. The slope says carry dislikes rising volatility, which nobody disputes. Only the curvature says the payoff is option-like, and that is the weaker of the two results.

So, can the destructive fifth of the sample be spotted in advance?

Read more

Let’s get into it.

Alpha from Mandelbrotian Prices

The most accurate option pricing model turned out not to be the most profitable one.

The authors ran five pricing models against S&P 500 option chains, then turned each into a simple trading rule: buy when the model says the market price is too low, sell when it says the market price is too high. On raw accuracy, classical Black-Scholes held up remarkably well, landing closest to actual prices across most medium and long dated contracts, which is a little annoying given how many of its assumptions are known to be false. But when those same models became strategies, Black-Scholes finished fifth out of nine. The winner was Mandelbrot's multifractal model, which treats prices as rough and self-similar rather than smoothly random, and which more than tripled starting capital in the backtest.

Figure 1: Every model-based strategy roughly tripled its starting capital. Buy-and-hold, momentum, and mean reversion finished essentially flat. Note that Black-Scholes, the most accurate pricer in most categories, sits mid-pack here.

The paper says it plainly, noting that “accuracy in option price calculation does not always translate” into trading results. The catch is that data limits confined the backtest to a single date in 2023, so read this as proof of concept, not verified edge. The broader lesson still holds, that the model that best explains yesterday's prices is not automatically the one that finds tomorrow's mispricings.

Bhattacharyya, Ritabrata and Goh, Zhi Hwee and Korovedzai, Rudairo Orpah and Chen, Jun-Han, Alpha Generation using Option Trading Strategies based on Modeling Options Prices considering Mandelbrotian Movement of Prices (July 21, 2026). Available at SSRN: https://ssrn.com/abstract=7154800 or http://dx.doi.org/10.2139/ssrn.7154800

ETF (Mis)pricing: Blame the Inventory

ETF prices drift from the value of their underlying holdings largely because the firms responsible for keeping the two aligned run into inventory limits, not because the underlying assets are broken.

Authorized Participants (the large trading firms permitted to create and redeem ETF shares) are supposed to arbitrage away any gap between an ETF's market price and the value of its basket. Using FCA regulatory data covering 128 UK listed ETFs from 2018 to 2022, the authors observe each firm's daily inventory directly for the first time, and the pattern is consistent. When an AP holds more ETF shares than it wants, it quotes cheaper prices to offload them, and the fund slips to a discount. The adjustment shows up almost entirely in the ETF price, while NAV barely responds.

Figure 2: Corporate bond ETFs, February to May 2020. Median price to NAV gap (black), spread across funds (bands), Fed and BoE interventions (vertical lines). The gap closes as policy eases dealer funding pressure.

The effect also splits by asset class, with equity ETFs more sensitive to inventory and bond ETFs more sensitive to unexpected order flow. APs sometimes take directional positions instead of correcting gaps immediately, which undercuts the assumption that arbitrage is instantaneous. For investors, this reframes premiums and discounts (bond ETFs moved over 5% from NAV in March 2020) as a read on dealer capacity rather than fund quality, since inventory practices “can exacerbate mispricing, particularly during periods of stress.”

Kraus, Wladimir and Kirilenko, Andrei A. and Linton, Oliver B. and Xiao, Mingmei, ETF (Mis)Pricing (May 25, 2025). Available at SSRN: https://ssrn.com/abstract=7143458 or http://dx.doi.org/10.2139/ssrn.7143458

Climate Attention and the Bond Market

Public curiosity about global warming, measured by Google searches, predicts higher returns on U.S. Treasury bonds over the following year.

The authors built a monthly index of worldwide searches for “global warming” and test whether it forecasts the excess return Treasuries earn over cash. It does, and not marginally. Higher search interest reliably precedes higher bond returns across maturities from two years all the way out to twenty four, and the relationship survives controls for the yield curve, the standard bond forecasting factors, and five separate uncertainty measures.

More impressively, the signal holds up out of sample, where most predictors quietly die, cutting forecast errors by roughly 15% against the historical average benchmark. The mechanism is less exotic than it sounds. Rising climate attention travels with expectations of a softer economy (weaker production, higher unemployment) and a tilt toward safer assets, so investors mark down the expected path of policy rates. The effect sits almost entirely in expected short rates rather than term premia, and it barely existed before the Paris Agreement. For investors, that suggests attention data can proxy for shifts in macro expectations before conventional indicators register them.

Yu, Deshui and Tang, Jiachen and Li, Luyang and Zhou, Mingtao, Climate Attention and Treasury Bond Risk Premia. Available at SSRN: https://ssrn.com/abstract=7203448 or http://dx.doi.org/10.2139/ssrn.7203448

Trading on Language Models Under Market Frictions

Large language models beat word-counting sentiment on financial news, but their edge lives almost entirely in the sentences where meaning depends on structure rather than vocabulary.

The authors push nearly a million firm specific news stories through six sentiment engines, from the classic Loughran McDonald word list to an 8 billion parameter LLaMA-3, then force every signal through the frictions an actual desk faces (trades happen only after the news was genuinely observable, positions pay spreads and costs, and no trade can exceed a tenth of a stock's daily volume). The word list barely beats a coin flip and loses money once turnover is paid for. The decoder models survive, with LLaMA-3 posting the strongest net long short performance in a test window chosen to sit after its own training cutoff.

On plainly worded news, a 1980s word list trails an 8 billion parameter model by 17 points. On contrastive clauses (revenue beat, guidance cut) the gap more than doubles to 37, and the word list lands below a coin flip. Sample of 1,200 test articles, coded by two annotators. Chart: Alpha in Academia. Data: Table 13, MFAST working paper (SSRN 7213792).

The more useful result is where that edge comes from. LLaMA-3's advantage over the dictionary is smallest on plainly worded headlines and roughly doubles on negation, contrastive clauses, and forward looking guidance, exactly the sentences where a revenue beat sits next to a guidance cut. It also widens in less liquid names. The lesson for investors is that reading comprehension, not model size, is the thing being sold, and that a signal only counts once costs and capacity are in the room.

Kirtac, Kemal, Financial Language Models as Applied Artificial Intelligence Systems for News-Based Trading under Market Frictions. Available at SSRN: https://ssrn.com/abstract=7213792 or http://dx.doi.org/10.2139/ssrn.7213792

This week for paid subscribers

Paid subscribers are watching the equity correlation matrix collapse onto a single factor across the 2008 and 2020 crashes, then testing whether that same fragility measure warns of a crash in advance. It forecasts the crisis that built up endogenously and is blind to the exogenous one, with the first sustained signal crossing fifteen months before Lehman. Python backtest code included.

Alpha in Academia

When correlations concentrate

Read more

2 months ago · 10 likes · Alpha in Academia

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 will take a look at how when markets crash, the correlation matrix of the equity cross section collapses. The market factor absorbs a rising share of total variance and the effective number of independent factors falls sharply. This spectral concentration replicates cleanly across both the 2007-08 financial crisis and the 2020 Covid crash. But the same signal, run as a causal early-warning indicator, only forecasts the crisis that built up endogenously, and the exogenous shock is invisible to it in advance.

Let’s dive right in.

Introduction

Igor Halperin’s July 2026 paper introduces an approach called Observable Matrix Dynamics that tracks the equity cross section through the trajectory of a fixed-size distance matrix and its spectrum. We run two fixed-size matrix observables across the crisis decades using the current S&P 500 universe with daily adjusted closes, and the central empirical finding is that the correlation spectrum collapses onto the market factor at the 2008 and 2020 onsets.

A separate section tests whether these fragility signals actually forecast the crash they concentrate on, and finds that only the endogenously building 2008 crisis is predictable in advance.

“An endogenous fragility measure can forecast a crisis that grows out of the correlation structure itself, and cannot forecast an exogenous shock such as a pandemic, or a dispersed decorrelated unwind.”

Data & Methodology

The setup is a rolling Pearson correlation matrix on the S&P 500 cross section over a 504-day window, from which we extract the eigenvalues and read two summary statistics. The first is the market-factor share, the largest eigenvalue divided by the sum of all eigenvalues, which measures how much of total return variance is absorbed by the single largest common factor. The second is the participation ratio, defined as the square of the sum of eigenvalues divided by the sum of squared eigenvalues, which counts the effective number of independent factors. Both are standard diagnostics from the random matrix theory of correlation matrices, and both are known to move sharply at crisis onsets.

We compute these statistics across two crisis periods and average them over the calm year before onset and the first two months of the crash. The universe is current S&P 500 constituents with full history back to each period’s two-year pre-roll, filtered to 349 names for the 2007-08 window and 353 names for the Covid window. This universe is survivorship-biased by construction, which means 2008-era failures like Lehman, Bear Stearns, and Washington Mutual are absent from our sample. If anything this should understate the correlation stress at the 2008 onset, since the survivors are disproportionately the firms that made it through the crisis intact.

Read more

If you have been reading Alpha in Academia for a while, you will know that the Thursday posts tend to go a little further than the weekly paper summaries. Sometimes I implement a strategy. Sometimes I replicate a result, work through a model, or find that an interesting idea does not quite survive contact with the data.

This page is where all of that work lives. I have also included the research companions I have available. Depending on the investigation, that may mean code, a notebook, data, tests, or reference results. They are there if you want to rerun the work, change an assumption, or take the analysis in another direction.

If you find something I missed, get a different result, or have an idea you would like me to look at next, please let me know.

Read more

Let’s get into it.

The Stock Market Never Learns

Halperin borrows a tool from AI research to watch how the entire stock market’s correlation structure moves over time, and finds that unlike a neural network, the market never settles into a stable pattern, even decades after a crisis passes.

The core idea here is clever: treat the whole S&P 500 like a machine learning model in training, and watch whether its “internal representation” (the correlation structure between all 500 stocks) ever stabilizes the way a trained neural network’s does. It never does. Across three crises (the 2001 dot-com bust, 2008, and 2020), the market’s correlation geometry collapses in dimension during a crash, as everyone starts moving together, then partially unwinds afterward, but never settles into the kind of clean, learned structure you’d see in a model that’s finished training. Separately, Halperin ranks stocks by relative performance and volatility each day and tracks those rankings as their own systems. Performance rankings shuffle fast (a week), while volatility rankings are sticky (a month or more), and only the volatility ranking shows a real “arrow of time,” meaning it behaves differently forwards than backwards, flaring hardest at the 2002 and 2008 lows. For investors, the practical takeaway is: the correlation-based fragility signals only gave advance warning for 2008, the slow-building crisis. They were blind to Covid.

As Halperin puts it, the market is best read as an object that a crisis “drives further from equilibrium rather than toward a new one.”

Halperin, Igor, Observable Matrix Dynamics of Stocks (July 20, 2026). Available at SSRN: https://ssrn.com/abstract=7149898 or http://dx.doi.org/10.2139/ssrn.7149898

Why a Yen Funding Rate Half a World Away Moves the Price of U.S. Corporate Debt

Japanese banks fund roughly a quarter of America’s CLO market through yen-dollar swaps, and the cost of that swap, not U.S. credit conditions, is what actually drives pricing at the top of the capital structure.

Here’s the setup that makes this paper fun: Japanese banks, led by the agricultural cooperative bank Norinchukin (nicknamed “the CLO whale”), hold about a quarter of all AAA-rated U.S. collateralized loan obligations. But they don’t fund those purchases with dollars, they fund them by swapping yen into dollars through FX derivatives. That swap has a price, called the cross-currency basis, and it turns out to explain over 60% of the quarterly variation in CLO issuance, more than VIX, credit spreads, or any domestic macro variable the authors threw at it.

The real hook, though, is that this sensitivity isn’t stable. After Japan tightened bank capital rules in 2019, the pass-through from funding costs to AAA spreads roughly tripled, because the shock pushed the price-insensitive whale out and let more rate-sensitive buyers set the margin. That’s a good reminder that “foreign ownership share” is a lazy proxy for exposure. What matters is who’s marginal, and how easily they can walk away.

Huber, Amy and Kundu, Shohini, When Funding Markets Move Credit Markets: Foreign Investors and U.S. CLOs (May 15, 2026). Available at SSRN: https://ssrn.com/abstract=7086758 or http://dx.doi.org/10.2139/ssrn.7086758

Deep Hedging Finds Free Money

A neural network trained to hedge derivatives will, if left alone, quietly convert itself into a leveraged bet on the market’s historical drift rather than an actual hedge.

Deep hedging replaces the classic quant approach of assuming a price model (Black-Scholes, Heston, etc) and deriving a formula, with a network that just learns the best trading policy directly from simulated or historical paths, optimizing for a risk-adjusted objective rather than a textbook Greek. The elegant idea, borrowed from Buehler and coauthors, is that “the price of a derivative is the cost of its hedge,” so if you can learn the optimal hedge, you’ve also learned the fair price. The catch shows up fast: when the author trained this framework on S&P 500 data from 2015 to 2025, the network didn’t learn to hedge options responsibly, it learned to go long the index, sell puts, and mostly ignore the derivative it was supposed to be hedging, because that combination was simply the highest-Sharpe trade over that specific decade.

Figure 2: shows the distribution of trading gains “Under P” (raw historical drift) versus “Under Q” (drift-adjusted, risk-neutral measure), which visually demonstrates how removing the statistical arbitrage collapses the fat right tail into a distribution centered near zero.

The fix involves reweighting the training paths into a new probability measure that strips out the drift, forcing the model to actually hedge instead of quietly rediscovering “stocks go up.” It’s a clean reminder that any backtested strategy, human or machine, can mistake a decade’s tailwind for genuine skill.

Buehler, Hans, (Deep) Learning to Trade II - Deep Hedging, Model Uncertainty, Deep Bellman Hedging (March 18, 2026). Available at SSRN: https://ssrn.com/abstract=7086438

A Machine That Draws Arbitrage-Free Options Markets, and Then Learns to Simulate Them

A weighted sum of Black-Scholes prices, fit with plain linear programming, can capture an entire options market’s shape without ever creating a mathematical arbitrage, and once encoded into just twenty numbers, that shape can be simulated forward in time under a risk-neutral measure.

Building a model of the options market has always meant a tradeoff. Flexible models fit reality poorly, while accurate ones (like local volatility) are painfully hard to compute and calibrate without human babysitting. This paper’s answer, called SANOS, sidesteps the mess entirely by expressing every option price as a weighted mix of simple Black-Scholes prices across a grid of strikes. Because that structure is mathematically guaranteed to avoid static arbitrage, fitting it becomes a fast, off-the-shelf optimization problem rather than an art project, and in testing it priced 91.4% of real SPX options within the bid-ask spread, in well under a second.

Note: Fitted 1000 options within 2 ATM implied volatility standard deviations which had a Vega/sqrtT of at least 0.1% on 2025-05-06 across all 48 expiries from 1D to 657 business days. Options were chosen by closeness to ATM. The model fitted 91.4% of all options within bid/ask. Of those options not fitted the median error is just 21% of half spread.

The authors then compress five years of these fitted surfaces into a 20-number daily state, run PCA to find just five real drivers of surface dynamics, and use that to simulate future markets. As they put it, once set up, “it is fast.” For traders, this means a genuinely tractable way to stress-test strategies against markets that behave like the real thing, without hand-tuning a volatility surface every morning.

Buehler, Hans and Horvath, Blanka and Kratsios, Anastasis, DYSANOS - Generative Dynamic Smooth Arbitrage-free Non-parametric Option Surfaces (Presentation) (July 01, 2026). Available at SSRN: https://ssrn.com/abstract=7047878 or http://dx.doi.org/10.2139/ssrn.7047878

This week for paid subscribers

This week, paid subscribers get full access to a special two-part deep dive on pricing spread options and forecasting correlation when traditional models fail. Part 1 covers the pricing machinery and where standard approximations break down, while Part 2 tackles the hidden risk of correlation instability and how to actually select your parameters. Both posts include complete Python backtest code and historical datasets so you can run the entire pipeline yourself.

Part 1:

Part 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.

Yesterday we built three ways to price a spread option and found that the pricing machinery is not the problem. Kirk’s approximation is accurate to a few basis points in the region where it actually gets used.

We ended on the thing that is the problem. A spread option’s price depends on the correlation between its two legs about as much as it depends on either leg’s volatility. But unlike volatility, correlation has no market, no implied surface, and no vocabulary for expressing how unsure you are. It’s a number someone types into a box.

Today: which number, and what the choice costs.

Let’s dive right in.

How Much Does It Actually Move?

The case for treating correlation as a fixed parameter is that it does not wander much. Volatility clusters and spikes, but two things that are economically linked (crude oil and the fuels made from crude oil) ought to stay linked.

Over our sample, the sixty-day rolling correlation between crude and the 2:1 product basket has a full-sample value of 0.677. That single number is the one most likely to end up in the box.

Here is what the rolling estimate actually did.

Figure 1: Sixty-day and 250-day rolling correlation between crude and the product basket, 1986–2026. The dashed line is the full-sample value a model would typically use.

The sixty-day estimate ranges from 0.22 to 0.98, nearly the entire theoretical span. And the departures are not brief excursions around an otherwise stable mean. The rolling correlation sits more than 0.15 away from the full-sample value on 33.7% of all trading days.

Put differently, for a third of the past forty years, the number in the box was wrong by an amount we are about to put a price on.

What That Range Is Worth

A parameter wandering between 0.22 and 0.98 only matters if the price cares. So take our three-month at-the-money crack spread option, hold both volatilities fixed at their forty-year values, and change nothing at all except the correlation.

A 5.6x range. Same strike, same expiry, same volatilities, same underlying. The entire difference is an assumption.

And that is not a point made with implausible extremes. Restricting to the middle of the distribution (5th to 95th percentile of what correlation has actually done), the same option still spans $3.30 to $7.78, a factor of 2.4.

So, which number do you type in? Before answering that, it is worth understanding one property of this correlation, because it turns out to explain more than it looks like it should.

Does Correlation Really Spike in a Crisis?

There is a well-documented result in equity markets that correlations rise in a downturn. Diversification thins out precisely when it is wanted, because in a sell-off individual names stop trading on their own news and start trading on the market’s.

The question is whether that transfers here. A crack spread is not a portfolio of equities. It is one commodity against the fuels refined out of it, and the link between them is a physical production process rather than a shared risk appetite. So it is genuinely unclear, before looking, whether the same pattern should hold.

We can check directly. Split every day in the sample into five buckets by the prevailing crude volatility, from calmest to most stressed, and see what correlation was doing in each.

Correlation rises from calm into the middle of the distribution, which is the equity pattern working as advertised. And then, in the most stressed quintile, where crude volatility averages 61% annualized, three times the calm regime, it falls back to 0.696, below the mid-regime level. The dispersion widens too: the standard deviation of ρ is at its highest, 0.148, exactly where you would want it lowest.

So the shape is a hump, not a ramp. The equity intuition holds through ordinary stress and then inverts in genuine crisis.

The mechanism is not mysterious once you look at the individual events. In a real dislocation crude stops trading on the same information as the products. April 2020 is the cleanest illustration in the sample: crude collapsed on a storage constraint, a mechanical problem about where to physically put barrels, while gasoline and diesel kept tracking actual fuel demand. On the day WTI printed −$36.98, gasoline was still trading comfortably above zero. The legs had decoupled entirely.

Hold onto this finding. It looks like a piece of trivia about crisis behavior, and it turns out to be the reason the standard way of estimating correlation is biased.

Read more