Backtest Event Ordering: Test Gaps Between Bars

Backtest event ordering: Specify when signals become available and use a gapped fixture to expose ambiguous timing.

The data-accuracy discussion described uncertainty in an LLM-built futures backtester, particularly around minute-bar gaps and timestamps. Accessible comments proposed integrity checks and independent calculations. No reviewed code established the engine as correct, so a small counterexample is more useful than assurance.

The source discussion is a starting question, not verified strategy evidence. Accessible comments were reviewed incompletely; some replies remained unavailable. Author results are unverified. The exercise below is original educational reasoning with hypothetical inputs, not a reproduction of those results.

Backtest event ordering: bar-gap engine test: make a tiny tape

Create a hand-written sequence with five bars, an overnight boundary and one missing interval. Store timestamps with a stated timezone. Choose prices that visibly differ between the previous close and the next open. This makes an accidental substitution observable without requiring a large historical dataset or an expensive provider subscription.

Write expected behavior before running the engine. Specify when the signal becomes available, the first eligible order time, how a missing bar is handled and which price marks an open position. Do not silently fill a missing interval with a tradable price. A forward-filled value may be useful for some calculations while remaining unavailable as an execution opportunity.

Worked example: the next open is not the last close

Imagine a signal calculated after a close of 100. The next session opens at 108 and later closes at 106. A rule that buys only after the signal is available cannot claim an entry at 100 merely because that was the latest stored value. For one share, buying at 108 and marking at 106 produces a loss of 2 before costs. These are synthetic values, not market observations.

Change the next open to 92 while keeping the earlier bars identical. Decisions made before that open should remain identical unless the documented rule legitimately waits for it. This is a simple truncation-invariance test: future edits must not change already completed historical decisions.

Test information time separately from row time

A row labeled with a session date can contain a field published later. Add an availability timestamp to a synthetic feature and prevent the decision function from reading it early. Shift that availability time while leaving the value unchanged. Only decisions after the new availability boundary should be able to respond.

For a rolling indicator, calculate the expected value by hand on the permitted rows. Repeat with a duplicate timestamp and an out-of-order bar. The engine should reject or explicitly normalize those inputs according to a stated policy. Quietly sorting and deduplicating can conceal an upstream feed issue that deserves a report.

Keep instrument conventions visible

Futures roll treatment and equity corporate actions need different explanations. A raw-versus-adjusted slogan does not specify contract eligibility, settlement, split handling or executable prices. Record the convention used in the small tape, then preserve it in larger runs.

A passing test establishes one defined behavior. Expand the tape with a halt, session change or missing symbol only when that case matters to the intended model. Independent expected calculations are the evidence; the engine repeating its own answer is not.

Reference and first useful step

TimeSeriesSplit provides chronological train/test indices and an optional gap. TimeSeriesSplit documentation.

Open Advisory Research to explore a local sample report and inspect its assumptions before comparing your own saved evidence. Choose a synthetic sample before importing offline JSON. Keep outputs advisory; this exercise does not place trades. For background, read backtesting fundamentals and paper trading.

Continue the research

Earnings Backtests: Use an As-Of Event Calendar and Backtest P&L Reconciliation: Audit Matched Trades develop adjacent checks. Use a small synthetic fixture first; these educational diagnostics do not validate a trading edge or change account settings.