point-in-time-researchlisted

# Point-in-Time Quant Research A historical price series is not enough to make a backtest point-in-time. The moment a strategy reads a financial fact, that fact must already have been **observable, admissible, and executable** by a real investor. Most "great" backtests are not fraudulent — they are *contaminated* by information that did not exist yet. This skill is the discipline that prevents it. ## When this matters Trigger this skill for any research that turns data into a claim about future returns: factor IC, long-short alpha, Sharpe, a screener edge, an ML signal, a strategy backtest. The user usually will **not** say "look-ahead bias" — they will say "does this strategy work?", "backtest momentum on SPY", "is this factor predictive?". That is exactly when the leaks below do their quiet damage. ## The three silent leaks 1. **Period-end dating.** A fact for the quarter ended 31 Mar describes activity through 31 Mar, but the market only learns it weeks later (earnings release, 10-Q in May). Joining on `period_end` hands the simulated investor weeks of free foresight. Anchor on the **filing/acceptance** time, never the period end. 2. **Latest-value overwrite.** Databases answer "what is the value for this period *now*?" — overwriting the originally-reported number with later restatements. A backtest then trades on a revision that did not exist at the time. Use the **originally-filed** version, not the latest. 3. **Same-session execution.** A release