ai-asset-pricing: Skill for Claude Code

Factor Construction Rules

These rules prevent look-ahead bias (LAB) when forming cross-sectional factors. LAB means using information beyond the portfolio formation date — it silently inflates factor returns and invalidates results.

LAB Audit Checklist

Run this checklist whenever forming portfolios or constructing factors. Flag any violation.

1. Signal is point-in-time at formation date t

• Accounting signals use data available at t (raw Compustat: ≥4-month lag; JKP: already aligned)
• Market data (price, ME, volume) uses end-of-period t values
• Regression-based signals (beta, loading estimates) use windows ending at or before t
• No .shift(-1) or forward-looking operation on the signal
• Signal does not condition on future outcomes (e.g., future default, future delisting)

2. Universe defined at t

• Stock universe uses only t-available information (listing status, exchange, share type at t)
• Stocks that delist after t are included in the t-formation universe (you didn't know they'd delist)
• Stocks that IPO after t are excluded from the t-formation universe
• No survivorship conditioning (don't require stocks to exist for N future months)

3. Breakpoints computed at t

• If using NYSE breakpoints: NYSE membership determined at t
• Percentile cutoffs computed from the cross-section at t
• No future data in breakpoint calculation

4. Returns from t+1

• Portfolio holds from t+1 (month after formation)
• VW returns use ME at t as weights (not t+1 ME)
• Delisting returns compounded with final trading-day return in t+1

5. Output dating convention

• Portfolio return series re-dated to the earning period (see convention below)

Core Convention: Lead Returns, Not Signals

Preferred approach: Keep signals at their natural date. Lead only the portfolio return series.

Step 1: At date t, sort stocks using signal known at t
Step 2: Compute portfolio return from stock returns at t+1
Step 3: Assign this return to date t+1 (the period it was earned)