Research notes, from paper to result
One published quant paper per note, tested on real market data, with the code and the finding.
Each note picks one published quant paper and tests it on real market data from scratch. The plan is written first, then the data loads. You see the code, the charts, and what came back, even when the answer is no. Not advice; a replication study.
All notes
11 min read
Did the five-factor model survive the years Fama and French never saw?
A preregistered test on the 150 months after the paper's sample ends. The failure its abstract admits to has gone, four of its five factors are now redundant against each other, and no model beats any other on the paper's own test portfolios because none of them fail.
9 min read
A factor with a 0.90 Sharpe ratio that breaks even at 7 basis points
The trade replaces the whole portfolio every month, so it pays only while costs stay under a break-even. That break-even is now 7 basis points, against a gross return that fell from 0.88% a month before publication to 0.14% after.
10 min read
For eighty years the US market paid nothing except at the turn of the month. Then it stopped.
A preregistered replication of McConnell and Xu: the four-day effect holds at t = 8.03 from 1926 to 2005 and clears t = 4 in three consecutive quarter-centuries. In the twenty years since it is 7% of its old size.
9 min read
Is the accruals anomaly dead? Testing Sloan thirty years on
The low-minus-high accrual spread paid 0.38% a month before Sloan was published and 0.08% since Green, Hand and Soliman declared it finished. The decade panel does not fit the explanation either paper gives.
10 min read
Does selling in May work? A hundred years of US data says no
Winter really does beat summer, in every subperiod and every variant tried. It never reaches significance on US data alone, and the strategy loses to buying and holding at any cost assumption including zero.
12 min read
Is the security market line still flat? Testing Frazzini and Pedersen out of sample
Tripling market beta bought a quarter more return over sixty-three years, and alpha falls almost monotonically across the deciles. The levered spread meant to harvest that clears its threshold as preregistered and falls below it at a 36-month beta window, at a 120-month window, and on quintiles.
9 min read
Are momentum crashes forecastable? Testing Daniel and Moskowitz out of sample
A preregistered replication on Ken French data: the bear-market and variance state effect is visible, the specific interaction the test demanded is not, and the out-of-sample win rests on a single month.
Research, not investment advice. Every result here is a test on historical data, and nothing in it is a recommendation to trade.