Conditioning a Signal on Liquidity

Illiquid names are often where a signal shows its strongest raw statistical edge, and also where that edge is hardest to actually collect. Fewer analysts cover small, thinly traded names, so genuine mispricing lingers longer — but the same thinness means wider spreads and thinner order books eat into whatever edge is there. Conditioning a signal's backtest on liquidity separates "does this work" from "can I trade it," which is a distinction a single pooled IC number erases.

Splitting the test, not just the universe

The naive fix — just restrict the whole backtest to liquid names — throws away information about how the edge varies. The more useful test buckets names by a liquidity measure (average daily dollar volume is a common, simple choice) into terciles or quintiles, and reports IC separately in each bucket, alongside an estimate of round-trip trading cost in that bucket. That produces two curves side by side: raw statistical edge by liquidity bucket, and cost by liquidity bucket. Where they cross tells you the honest boundary of where the signal is tradeable.

A worked example

A quality signal's gross IC by liquidity quintile: 0.045 (most illiquid), 0.030, 0.020, 0.013, 0.006 (most liquid). Estimated round-trip trading cost in the same quintiles: 60 bps, 30 bps, 15 bps, 8 bps, 4 bps. Converting IC to an expected return contribution and comparing to cost shows the two most illiquid quintiles are net negative after costs, despite having the highest gross IC — the edge is real but smaller than what it costs to capture there. The signal is only net profitable from the middle quintile outward, which is a much smaller and more liquid universe than "the whole cross-section" the raw IC number suggested.