Factor Evaluation

Factor evaluation treats the factor DataFrame as cross-sectional scores.

Higher scores are better.

IC and ICIR

For each date, bagelquant-bt computes the cross-sectional correlation between factor scores at date t and asset returns from t to t+1.

Factor evaluation outputs both Pearson correlation and Spearman rank correlation in result.ic:

result.ic.select("time", "pearson_ic", "spearman_ic")

result.ic_summary includes mean, standard deviation, and ICIR for each method. The compatibility fields ic_mean, ic_std, and icir use Spearman IC.

icir is:

mean(IC) / standard_deviation(IC)

Quantile Returns

Each day, assets are sorted by factor score from highest to lowest and split into quantiles: q1 contains the highest scores and qN the lowest.

Each quantile return is the equal-weight average forward return of assets in that bucket.

The spread is:

q1_return - qN_return

TOP N Backtest

The TOP N backtest converts factor scores into long-only equal weights:

top N assets each day -> 1 / N weight each

The resulting weight frame is passed through the same backtest engine as a normal portfolio-weight DataFrame, including transaction costs.

Spread and Lag Analysis

Factor evaluation also builds a spread portfolio: long q1, short qN, and passes it through the same cost-aware backtest engine.

lag_analysis evaluates TOP N and spread portfolios with factor signals lagged by 0, 1, 2, 3, 4, 5, 10, 20, 30, and 60 observations.

lag_returns contains gross and net cumulative return time series for the same portfolio and lag combinations.

ic_decay reports mean Pearson and Spearman IC at the same lags and is plotted as an IC decay line chart.