Striking findings from academic finance, selected by a hedge fund veteran. Go deeper into the mechanism, the evidence and the details that can change the investment case.
Get the finding that caught our attention, the original paper, and the practical question it raises. Full reviews and house-test details are available with a paid subscription.
A backtest can earn your attention before it earns your capital. This issue starts with an options model reporting 43.6% annualized mean return before costs. Another study follows two familiar stocks into a trade that never reaches its exit. The useful part is what happens when each headline meets the assumptions behind it.
Which papers deserve a closer look—and which are useful tools rather than trading strategies? This edition examines seven studies, separates the authors’ findings from our assessment, and explains what we would test next.
A paper estimates a ninefold median capacity increase from slower rebalancing among factors with positive initial capacity. The signals stay the same; the trading rule changes. These are historical model estimates.
In Avramov and coauthors' historical model, applying a common 5% partial-adjustment rule produces a median ninefold increase in estimated capacity among factors with positive capacity at full monthly adjustment. The forecasts are unchanged; the portfolios respond more slowly. Nine times the capacity does not mean nine times the return.
We tested the turn-of-month effect on SPY. From 2016 through September 2026, its gross lead was just 1.1 basis points per session. Random dates often did better.
For SPY from January 2016 through September 2026, those four sessions earned 7.04 bps per day before costs, versus 5.95 elsewhere. The difference was just 1.09 bps, with a month-level t-statistic of 0.20. About 41% of the 500 sampled random four-session placements did better.
Ten independent trials can produce a 1.57 Sharpe from noise in a one-year model. Refusing to let a strategy fail is where the damage starts.
A 2014 paper by David H. Bailey, Jonathan M. Borwein, Marcos López de Prado and Qiji Jim Zhu shows how little searching can be enough. In their one-year, independent-trial model, ten strategy variants with zero true Sharpe produce an expected best backtested annualized Sharpe of about 1.57. That attractive number can come entirely from selecting noise.
In our synthetic 3x Nasdaq test, a 200-day filter cut maximum drawdown from 95% to 66%. A statistical pass can still be a brutal ride.
In the full simulation, the slow filter reduced maximum drawdown from 94.97% to 66.19%. Gross compound annual growth was 35.87% versus 34.78% for constant 3x exposure. After 5% annual financing on borrowed notional and 5 bps per unit of exposure changed, growth was 24.91% versus 21.98%.
Our sector-seasonality test earned money, but a neighbouring calendar lag earned slightly more in 2016-2026. The timing story still has something to prove.
In our historical adaptation, that rule earned 4.68 percentage points of annual gross excess over an equal-weight sector basket. Charging 10 basis points per dollar bought or sold reduced the advantage to 2.82 points. Those are annualised average-return differences, not differences in compound growth.
Moreira and Muir (Journal of Finance, 2017) size exposure by last month's variance. We calibrated one implementation on 1990 to 2016 data and evaluated it over the next 108 months.
We reproduced Bruce Knuteson's "Strikingly Suspicious Overnight and Intraday Returns" from the same public data, then split it by era. The overnight return is still there.
We ported an open-source Gaussian mixture regime model line for line and traded its signal from June 2004 to September 2026. It agrees with a one-line momentum rule on 80% of days, and a randomly shifted copy of itself did better 42% of the time.
Brogaard, Han and Kim's reversal appears in our high-VIX, two-hour rebuild. Sampled entry quotes imply a spread hurdle more than three times the gross edge under full spread crossing.
Drexler, Guettler and Sun (Chicago Fed, August 2026) report rising equity exposure. Our daily-data adaptation supports the rise but differs on the separate Nasdaq loading.
A valuation-based switch reduced monthly drawdown in our house test. Its return advantage was much less convincing.
10.36% annualized with a 40.88% maximum monthly drawdown, versus 10.29% and 50.31% for the market control. The tested idea switches between trend following and staying long the market as valuations change.
Our six-market house test found a profitable slow rule and an unprofitable fast one. Execution and data limits still matter.
The slow futures trend rule earned 6.96% annualized after modeled costs, with a 9.62% maximum drawdown at recorded session opens. The fastest rule lost 0.61% annualized under the same cost convention.
A simple monthly allocation rule beat its fixed-lookback control in our house test. The question is whether volatility timing deserves the credit.
13.05% annualized after modeled trading costs, using five ETFs and a monthly decision. Our VIX-adaptive rotation finished ahead of the fixed-lookback version, which returned 11.51%.
Devanathan, Tzikas and Boyd (arXiv, September 2026) report 11.6% with an 18.1% drawdown. Our reconstruction lands within a tenth of a point. The question is what you can trust.
Our reference reconstruction produced 11.65% annualized return and an 18.07% maximum drawdown, against the paper's 11.6% and 18.1%.
Concretum Research (June 2026) proposes a pre-close entry on volatility stress. Our daily-data rebuild reproduces the return curve and exposes what daily data cannot prove about the signal.
maximum drawdown is 39.1% measured daily and 20.7% measured at month-ends
Harvey, Mazzoleni and Melone (NBER, March 2025) document the flow. We rebuilt the trade on ES/ZN and on an ETF proxy, charged turnover, and got two different answers.
