Definition
An alpha factor is a numeric signal — computed from price, fundamental, or alternative data — that is hypothesised to rank securities according to their expected future returns relative to one another. Going long the top of the ranking and short the bottom should, if the factor has true predictive power, produce a return uncorrelated with the broad market. That uncorrelated return is the alpha; the factor is the recipe for extracting it.
The vocabulary descends from Fama and French's three-factor model and the academic factor zoo that grew around it — size, value, momentum, profitability, low-volatility, and dozens more. In practitioner usage the term has broadened: any quant-tradable signal, from a textbook factor to a machine-learning model's output, is an alpha factor if it produces a cross-sectional ranking that the trader is willing to bet on.
Why it matters
How it works
A well-formed alpha factor is a function from a universe and a timestamp to a vector of scores — one score per security. To be tradable, the function must be computable using only information available at that timestamp (no look-ahead), must be stable enough to survive realistic data noise, and must have a coherent economic or behavioural story behind it. A factor that works in-sample with no story is almost certainly overfit; a factor with a strong story but no historical evidence is just a hunch.
Once defined, a factor is evaluated by ranking the universe daily (or at the strategy's natural frequency), forming long-short portfolios from the top and bottom buckets, and computing the spread return through time. Standard diagnostics include the information coefficient (rank correlation between scores and forward returns), turnover (how much trading the factor demands), and factor-factor correlations against known canonical factors (to check the signal is not just repackaged value or momentum). Production strategies typically combine several factors using either equal-weighting, inverse-variance weighting, or a learned model that estimates each factor's conditional expected return.