IC-Weighted Signal Blending

A quant strategy often has several independent signals predicting the same thing, say, future stock returns, and simply averaging them equally ignores the fact that some signals are consistently better predictors than others. The information coefficient (IC) measures a signal's predictive skill: it's the correlation between the signal's ranking of stocks and those stocks' subsequent realized returns, typically re-measured each period. IC-weighted blending uses each signal's historical IC as its weight when combining signals into one composite score, so a signal with an IC of 0.08 contributes proportionally more to the blend than one with an IC of 0.02, rather than the two counting equally.

Concretely, if signal A has historically averaged an IC of 0.06 and signal B an IC of 0.02, a simple IC-weighted blend assigns weights proportional to those numbers, roughly 75% to A and 25% to B (0.06 and 0.02 sum to 0.08, and 0.06/0.08 = 0.75), rather than the 50/50 split an unweighted average would use. This lets a blended signal capture more of the genuinely predictive information in the mix while still benefiting from diversification across signals that aren't perfectly correlated with each other.