How do I avoid data leakage in a prop backtest?

For a simulated earlier bet, get the line from /api/v1/board?as_of or the latest history observation at or before the cutoff. A closing_line recorded later is not a valid earlier feature; inspect closing_observed_at, close_gap_seconds and close_confidence when evaluating close-time decisions.

Build rolling player features from prior completed games only, and pass end_date before the target game to team and opponent statistics. Fit preprocessing and model tuning on training data, then report performance on later held-out games. Random row splits can place near-identical props or the same game on both sides.

Historical box scores can be corrected after a game, so a present-day reconstruction is not automatically a point-in-time archive of every feature revision. Keep observation and revision metadata, state what your backtest can reconstruct, and never imply that an API alone proves a profitable strategy.