A Practical Checklist for Factor IC Evaluation
How I evaluate an alpha factor before it ever enters a backtest — IC, IR, decay, turnover, monotonicity.
Most quant backtests fail not because the idea was wrong, but because the signal was never properly evaluated before being thrown into a portfolio. After mining hundreds of factors across futures and A-share equities, I have settled on a small, almost mechanical checklist that filters out 80% of weak signals before they cost me a backtest cycle.
1. Start with Rank IC, not Pearson IC
Cross-sectional rank IC is far more robust to outliers and non-normal returns. For A-share names with limit-up / limit-down clamps, Pearson IC is almost noise. I look for a mean rank IC above 0.02 and an IC IR (mean / std) above 0.1 as a floor.
2. Layer backtest for monotonicity
A factor with a strong top-bottom spread but non-monotonic middle groups is usually a disguised size or liquidity beta. I require the long-short return curve to be roughly monotone across the 5 deciles — anything with a U-shape is almost always a poor risk control disguised as alpha.
- Top-bottom spread annualized > 8% (gross, no costs)
- Decile returns roughly monotone
- Top decile Sharpe > 1.0 in the in-sample window
3. IC decay is more informative than IC level
Plot Rank IC against horizon (1d, 3d, 5d, 10d, 20d). A signal whose IC halves between 5d and 10d is a fast signal that demands low turnover cost assumptions. A signal with flat IC out to 20d is a slow, capacity-rich factor you can hold comfortably.
If you cannot articulate the horizon at which your factor decays, you do not have a factor — you have a coincidence.
4. Turnover is a tax, not a feature
Even a 0.05 IC factor is worthless at 200% monthly turnover in A-shares once stamp duty and slippage are deducted. I budget a maximum one-way turnover per rebalance and reject any signal that cannot deliver its long-short spread inside that budget.
5. Stability across regimes
Split the sample pre-2020 and post-2021. If sign of IC flips, or if IC IR drops by more than 50%, the factor is regime-dependent and I will not allocate risk budget to it without an explicit regime switch.
This checklist is not a substitute for full portfolio optimization, but it filters out the obvious failures early — and it makes the remaining backtest results far more credible.