Weighting Training Samples by Tradability

A cross-sectional model trained to predict next-month stock returns treats every observation in its training set equally by default, whether that observation is a heavily traded large-cap or a thinly traded microcap that a real portfolio could barely have bought or sold without moving the price. If small, illiquid names disproportionately drive the largest apparent return anomalies, which they often do, because a handful of shares changing hands can swing a stock price sharply, an unweighted model can end up fitting patterns that a real strategy could never actually capture.

Weighting training samples by tradability means assigning each observation a weight (used in the loss function during training) that reflects how easily the position could have been traded at that point in time, typically some function of average daily dollar volume, bid-ask spread, or market capitalization, so illiquid observations count for less and liquid ones count for more. A stock trading $50 million a day might get a weight ten times that of one trading $500,000 a day, roughly mirroring the difference in position size a real fund could actually put on.

This is a training-time fix, distinct from simply excluding illiquid names outright: down-weighting still lets the model learn from the full cross-section (illiquid names carry real information about the broader pattern) while nudging the fitted model toward the signal that survives in the liquid universe a live strategy would actually trade.