T-Statistic Hurdles for New Factors

A t-statistic of 2.0 is the textbook cutoff for "statistically significant" in an introductory statistics class, corresponding to roughly a 5% chance of seeing that result by pure luck if there's actually no effect. Applied naively to factor research, that standard is far too weak: academics and practitioners have collectively tested many hundreds of candidate return factors, and if you test enough random noise variables, roughly 5% of them will clear a t-stat of 2.0 purely by chance — meaning a large fraction of "significant" factors in the literature are likely statistical accidents, not real effects.

The fix researchers now use is raising the bar substantially — a widely cited threshold is a t-statistic of at least 3.0 for a genuinely new factor, which pushes the chance-alone probability down enough to survive the fact that so many factors have already been tried. Some researchers go further and apply an explicit multiple-testing correction that scales the hurdle with the number of factors known to have been tested historically, since every additional factor tried and discarded silently raises the true bar for the next one to be believed.

This higher bar isn't about being conservative for its own sake — it's a direct, quantitative response to the file-drawer problem: unpublished failed factors don't show up in any dataset, so a published significant result is already the survivor of a much larger, mostly unseen search.