Earnings Backtests: Use an As-Of Event Calendar

Earnings backtests: Freeze calendar availability at the decision time before testing whether earnings features add information.

The earnings-model discussion compared a baseline, an upcoming-earnings indicator and an exclusion rule. Reported return and drawdown tradeoffs motivated the exercise here. The models, performance and historical calendar coverage were not independently reproduced, and accessible replies did not resolve every data-timing question.

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.

Earnings backtests: earnings feature ablation: freeze the comparison

Write three variants before scoring: no earnings field, an earnings field and a hard exclusion. Hold the remaining features, training schedule, position sizing, costs and evaluation dates constant. Record any variant-specific missing data. This isolates the intended change as far as the supplied dataset allows; it does not automatically create an unbiased selection process.

Keep the calendar's availability timestamp alongside the scheduled event date. A date appearing in today's historical file may not be the date a researcher knew at the original decision time. Date revisions and missing announcements should have explicit handling. Treat unknown as a distinct state rather than quietly converting it into no event.

Worked example: one announcement, two timelines

Imagine a company is initially expected to report on Thursday. On Tuesday afternoon, the announced date moves to Friday. A Monday decision can use only the information available on Monday. If a reconstructed Monday feature uses Friday because that is the final calendar entry, the exercise has changed the information set. The dates are hypothetical and illustrate availability, not a provider's behavior.

Build two calendar rows with publication timestamps and query the latest permitted row at each decision. Confirm that Monday uses the earlier schedule and Wednesday can use the revision. Then remove the Monday row. The resulting feature should report missing history, not infer what must have been known.

Measure the tradeoff instead of worshiping importance

Feature importance describes a model's internal use of a field under its fitted setup. The useful experiment asks whether adding the field changes net portfolio outcomes on held-out dates. Report return, drawdown, exposure, turnover and missing-calendar coverage together. A lower drawdown with much less invested time may need a different interpretation from a similar return at similar exposure.

When the exclusion removes an order, retain that order in an audit table. Show why it was excluded and what alternative capital treatment the model used. Do not compare the surviving trades alone while forgetting the cash periods created by the filter.

Preserve the research boundary

Choose the candidate and decision criteria using development windows. Freeze them before a separate final interval. If the final result motivates a new filter, log the revision and seek new prospective observations. Selecting the nicest return/drawdown combination after inspecting all variants is still a research decision.

A well-constructed ablation can explain what changed and when. It cannot establish that an earnings filter is universally helpful, that the underlying calendar is complete, or that a model's reported historical results will persist.

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

Backtest P&L Reconciliation: Audit Matched Trades and Intraday Trade Clustering: Resample Portfolio Days develop adjacent checks. Use a small synthetic fixture first; these educational diagnostics do not validate a trading edge or change account settings.