RollingIC Node
Rolling information coefficient (prediction correlation)
Overview
Rolling Information Coefficient (IC) measures the correlation between predictions and actual outcomes in rolling windows. It quantifies predictive skill: IC = 0.0 (no skill), IC = 1.0 (perfect predictions), IC = -1.0 (perfectly opposite). IC > 0.05 is statistically significant predictive power.
IC is fundamental in systematic trading: it directly indicates whether your signal actually predicts prices. Declining rolling IC warns of signal degradation. This metric translates trading signals into measurable skill, enabling data-driven strategy refinement.
Formula & Calculation
Return: Actual return from t to t+1
Measure across rolling window
IC stability matters more than absolute AC
Parameters
| Parameter | Default | Description |
|---|---|---|
| window | 60 | Rolling correl window (quarters) |
| lookahead | 1 | Periods ahead to predict |
Common Use Cases
1. Signal Validation
IC > 0 = signal has edge. IC = 0 = noise (abandon signal). Positive rolling IC = skill.
2. Signal Combination
Weight signals by IC: High-IC signals get 2x weight, low-IC signals get 0.5x. Optimize signal portfolio.
3. Hyperparameter Tuning
For each indicator parameter set, calculate IC. Choose parameters with highest rolling IC stability.
4. Regime Adaptation
Declining rolling IC = signal working worse. Switch to backup signal or adjust parameters.
Advantages & Limitations
Advantages
- Directly measures predictive power
- Simple interpretation ([-1,1])
- Robust to ranking vs levels
- Works across timeframes
Limitations
- Requires sufficient samples
- Affected by outliers
- Assumes stationarity
- Can be overfit in studies