Rolling IC computes the Information Coefficient of a signal on a sliding window of time — for example, the IC over each overlapping 63-day (one quarter) period. The resulting time series of IC values reveals whether a signal's predictive power is stable, decaying, or regime-dependent.
Key uses
- Live signal monitoring — rolling IC trending toward zero in real-time is an early warning that alpha is eroding, possibly due to competitor crowding, regulatory change, or a structural market shift.
- Detecting regime-conditional behavior — some signals have high IC in certain regimes (e.g., low-volatility trending markets) and near-zero or negative IC in others (risk-off, high-vol). Rolling IC reveals these regime dependencies empirically.
- Signal weighting — IC-weighted ensemble methods use recent rolling IC to allocate more weight to whichever signals have been most predictive in the recent window.
The rolling IC time series also supports diagnostic statistics: ICIR (IC mean / IC std) is a signal quality metric analogous to a Sharpe Ratio; IC autocorrelation measures persistence; IC drawdowns identify the worst sustained periods of no signal value.