Topic · Statistics & Econometrics
← All topicsStatistics for Alpha Research
34 articles · 6 checkpoints · 23 deeper reads · 5 reference notes
A standalone topic: it is on no roadmap, so read it on its own terms.
Every article, in reading order
plant a flag as you finish eachRead these first
The regression every quant equity researcher runs to test a candidate alpha signal is easy to get statistically wrong in three specific, well-known ways, and the checklist to catch it.
When consecutive observations are correlated, a series of N data points carries less independent information than N. This page gives the formula for how many "effective" independent observations you actually have.
The fundamental law of active management multiplies skill by the square root of the number of independent bets. Count correlated bets as independent and you overstate your information ratio by exactly the amount the correlation should have cost you.
An information coefficient is a correlation measured from a finite, noisy sample, so it has its own error bar. This page shows how to compute that error bar correctly, from the period-by-period series rather than by pooling every stock-month together.
Sampling long-horizon returns every month manufactures dependence between consecutive observations that plain OLS standard errors do not see. The Hansen-Hodrick correction accounts for exactly the amount of overlap you built in.
Testing fifty signals one at a time and testing the single portfolio that combines them are different statistical questions with different thresholds for what counts as significant. Confusing them is how mediocre research gets published as a discovery.
Then the rest