factor-backtest

Factor Backtest

A library of individually-decent factors is not a strategy. This skill combines them into one composite signal and backtests the portfolio that signal implies — the level at which transaction costs and capacity actually bite.

Workflow

1. Combine and backtest

factorminer combine output/run1/factor_library.json \
  --data path/to/market_data.csv \
  --method all --fit-period train --eval-period test

• --method — equal-weight, ic-weighted, orthogonal, or all to compare every method.
• --fit-period — split used to fit weights / run selection (use train).
• --eval-period — split used to score the composite (use test).
• --selection — optional pre-filter: lasso, stepwise, xgboost, or none.
• --top-k — keep only the top-K factors before combining.

The report gives composite IC Mean, ICIR, Long-Short return, Monotonicity, and Avg Turnover.

2. Generate tearsheets

For the visual portfolio view — quintile returns, IC time series, correlation heatmap:

factorminer -o output/run1 visualize output/run1/factor_library.json \
  --data market_data.csv --period test --tearsheet --quintile --correlation

What to look for

• Monotonicity — quintile returns should step up Q1→Q5. A non-monotone composite is fragile regardless of headline IC.
• Long-short return net of turnover — high Avg Turnover means the gross return is optimistic; FactorMiner's transaction-cost model is what makes the net number honest.
• Method spread — if orthogonal and equal-weight disagree sharply, the library has redundant or unstable factors; revisit factor-evaluation.

Guardrails

• Fit weights on train, score on test — never fit and score on the same split.
• The backtest estimates historical behavior; it is not a forward return promise. Present it as a research artifact for review.
• Report net-of-cost numbers as the headline; gross numbers only as context.