Historical betting odds: audit the data before the backtest

Reproducible audit: a provider-published EPL snapshot

We downloaded The Odds API’s published historical EPL example on 30 October 2026 and ran the downloadable audit script. Its snapshot timestamp is 29 October 2021 at 23:55 UTC. This is an independent demonstration on a provider example, with no client commission, provider endorsement or live-feed benchmark.

The file contains 20 events, 13 distinct bookmaker keys and 240 bookmaker-event records. Across those records are 270 markets; this audit’s completeness, source-age and overround calculations cover the 240 head-to-head markets and their 720 outcome rows. The test requires exactly the event’s home team, away team and Draw, with unique outcome names and finite decimal odds above 1.

• Reproduction: download the source JSON from the provider and the audit script below, then run: node odds-sample-audit.mjs historical-epl.json. The script reads a local file, makes no network request and writes only aggregate findings to standard output.
• Match the input SHA-256 in the downloadable results before expecting the same figures. A provider can revise the example in place; a different hash is a different input.
• The 30-minute flag in the code is an illustrative review threshold. It is not a provider commitment or an agreed client acceptance criterion.
• One snapshot supplies no season-wide expected event list, quote path, original delivery logs, betting-account eligibility or execution evidence. The clean checks do not establish those properties.

node odds-sample-audit.mjs historical-epl.json > my-audit.json

Keep three clocks with every quote

Retain the provider snapshot timestamp, the source or bookmaker update time when available, and the time your own system received the record. Store the requested time too: the timestamp you asked for can differ from the snapshot returned. The Odds API's historical documentation explicitly returns the closest snapshot at or before the requested time.

For a replay of your own system, a record must have arrived by the decision cutoff. For a historical reconstruction from a vendor archive, today's download time is not a claim that your model possessed the record then. State the archive-availability assumption and distinguish it from measured delivery. Never replace a missing timestamp with an event start time.

• Use UTC with an explicit timezone and retain original precision.
• Preserve missing source times as missing; report their frequency.
• Record event start-time changes, suspended prices, corrections and ingestion failures.
• Keep the raw response hash and the transformation version so a derived row can be traced.

Worked example: a quote that arrived five seconds too late

All event names, prices and timestamps in this example are invented. The decision cutoff is 18:00:00 UTC on 1 October 2026. Both records refer to a fictional full-game, two-way moneyline market with overtime included. A separate away outcome would still be required to calculate the whole market's overround.

Record B has the fresher source timestamp, but it reaches the replayed system after the cutoff. A nearest-timestamp join that chooses B gives the model information it did not yet possess. Under the stated replay rule, A is eligible and B is excluded. A's eligibility does not establish that a bet was available at that price.

Define an as-of join that cannot look ahead

First restrict the candidate rows to the exact event, bookmaker, market, line, period and settlement convention. Apply the availability cutoff before choosing a quote. For this replay example, require received_at ≤ decision_at and source_updated_at ≤ decision_at, then select the newest eligible source observation. Break ties deterministically and retain the chosen row's identifier.

Set a maximum source age before evaluating a strategy. If the latest eligible quote is too old, mark the decision missing instead of reaching forward to the next snapshot. Audit the missing decisions as well as the filled ones. A vendor archive may need a different availability rule; document it and test the effect of plausible delay assumptions.

Audit coverage and settlement before performance

Build an expected event universe independently of which rows happen to survive the download. Report the events and bookmaker-market combinations present, missing, duplicated and corrected. A dataset restricted to successfully matched final results can conceal failed collections, postponed games and settlement disputes.

Keep full-game and regulation-only prices separate. Totals and spreads need their numerical line; player props need player identity and participation rules. Record pushes, voids and dead heats explicitly. Compare outcome definitions before comparing model results across providers.

• Count expected versus observed events for each league, season and provider.
• Inspect duplicate identifiers and cross-provider event matches manually on a sample.
• Verify whether each outcome and line existed at the same observation time.
• Reconcile final outcomes, refunds and voids without deleting unresolved records.
• Calculate costs, stake limits and execution assumptions separately from quoted-price returns.

Keep the audit receipt beside the result

Save the extraction request, download date, provider terms, raw-file hashes, parsing version, expected universe and exclusion counts. Include a small sample of accepted and rejected joins. That gives a reviewer a way to distinguish an edge in a model from an edge created by the dataset.

The free CSV below contains the two synthetic records and their expected verdicts. Use it as a simple acceptance fixture for your own join logic, then expand it with genuinely observed samples you are entitled to process. The overround calculator can check a complete market once matching and timing are established.