Conditional IC by Volatility and Liquidity Buckets

A signal's overall information coefficient — a single average number computed across the entire history and all stocks — can be a misleading summary if the signal's real predictive power is concentrated in specific market conditions. A value signal, for instance, might have a healthy average IC that's actually the blend of a strong IC in calm, liquid markets and a near-zero (or even negative) IC during high-volatility or illiquid periods. Splitting the sample into buckets by volatility or liquidity and computing IC separately within each bucket surfaces this directly, and it changes how you'd actually want to size or gate the signal in live trading.

An analogy: a batter's overall average versus their average against fastballs

A baseball player's season batting average is one number, but it can conceal a batter who hits fastballs very well and struggles badly against curveballs — the overall average is a weighted blend of two very different underlying skills. A manager who only looks at the season average might not realize the batter should be pinch-hit for specifically when a curveball pitcher is on the mound. Conditional IC analysis does the same thing for a trading signal: instead of one blended number, you split performance by the "pitch type" — here, the volatility or liquidity regime the market was in — and see whether the signal is a fastball hitter, a curveball hitter, or genuinely good at both.

The method, one symbol at a time

Choose a conditioning variable — market or stock-level realized volatility, or a liquidity measure like average daily dollar volume — and split all observations into buckets (commonly terciles or quintiles) based on that variable's value at the time of the signal observation. Within each bucket $k$, compute the information coefficient $\text{IC}_k = \text{corr}(s_i, r_i \mid \text{bucket } k)$ using only the observations that fall in that bucket. Comparing $\text{IC}_1, \text{IC}_2, \ldots, \text{IC}_K$ across buckets reveals whether the signal's edge is uniform across regimes or concentrated in specific ones — and the overall unconditional IC is simply a (roughly) weighted average of these bucket-level values, which can look "fine" even when it's masking a wide spread underneath.

Worked example: volatility-conditioned IC for a mean-reversion signal

Suppose a short-term mean-reversion signal has an overall IC of 0.03 across the full sample — modest but usable. Splitting the sample into low-, medium-, and high-volatility terciles (based on trailing 20-day realized volatility at the time of each signal observation) gives $\text{IC}_{\text{low}} = 0.01$, $\text{IC}_{\text{mid}} = 0.03$, and $\text{IC}_{\text{high}} = 0.07$ — the signal's overall average of 0.03 is actually the blend of a fairly weak effect in calm markets and a much stronger effect in volatile ones, which makes intuitive sense for a mean-reversion strategy that profits from overreaction, since overreactions are larger and more common when volatility is elevated. A manager relying on the blended 0.03 figure to size the strategy uniformly across all market conditions would be under-allocating during high-volatility periods (where the real edge is much bigger) and over-allocating during calm ones (where there's barely any edge at all).

Worked example: liquidity-conditioned IC revealing implementation risk

Splitting the same signal instead by liquidity (average daily dollar volume, in terciles) might show $\text{IC}_{\text{illiquid}} = 0.09$, $\text{IC}_{\text{mid}} = 0.04$, and $\text{IC}_{\text{liquid}} = 0.01$ — a common and important pattern, since mean-reversion and many other anomalies tend to be strongest precisely in the stocks that are hardest and most expensive to trade. The overall IC of 0.03 here is dangerously optimistic about implementable edge: most of the raw statistical signal lives in the illiquid tercile, where realistic transaction costs and market impact are likely to eat most or all of the apparent edge, leaving the liquid, cheaply-tradable tercile — where you could actually execute at scale — with almost no real predictive power at all.

What this means in practice

Conditional IC analysis is standard practice before scaling any signal's position sizing or turning on a live strategy, since it reveals whether the signal should be gated or up-weighted by regime — trading it harder in high-volatility periods, or down-weighting it in the illiquid names where the statistical edge doesn't survive real trading costs. It's also a diagnostic for overfitting: a signal whose IC is entirely concentrated in one narrow, hard-to-trade bucket (like the most illiquid decile) is a common signature of a backtest that looks good on paper but has little implementable edge once realistic constraints are applied.