An odds snapshot is a time-stamped record of what the market offered for a football fixture before kickoff. The idea to leave with is simple: AlphaBench does not treat odds as predictions or tips. It collects them as auditable, immutable evidence, so football models such as deepseek-flash, gpt-5-mini, and grok-build-0.1 can be scored against the same pre-kickoff conditions and the same settled results. This is benchmark education, not betting advice.
Collection pipeline
The collection pipeline turns scattered market prices into a structured, reproducible dataset. For every fixture, AlphaBench records odds at multiple points before kickoff, then freezes those records once the match begins. A recent example is the settled fixture Khovd Western vs Deren in the Premier League: its pre-kickoff snapshots remain available as a historical benchmark even after the result is known.
A typical pipeline looks like this:
1. Fixture discovery. Schedules, team names, league names, and kickoff times are ingested and normalised. Each fixture receives a stable internal ID so odds, model predictions, and results can be joined later without ambiguity.
2. Source registration. Odds are collected from registered bookmaker or exchange feeds. Each source has an identifier, so a snapshot can always be traced back to where it came from. AlphaBench stores the source timestamp and the collection timestamp separately.
3. Polling schedule. Snapshot frequency changes over time. A fixture may be captured days before kickoff, again during the week, more frequently on matchday, and finally at a defined closing cutoff. This produces an odds timeline rather than a single number.
4. Market normalisation. Prices are stored in decimal form and mapped to standard market types, such as 1X2, over/under, and handicaps where available. Raw prices are preserved, while implied probabilities can be derived for comparison. Normalisation makes different sources comparable without hiding the original data.
5. Validation checks. The pipeline rejects impossible prices, duplicate snapshots, stale feeds, and timestamps after kickoff. It also checks that the fixture ID, market, selection, and source all match expected values.
6. Append-only storage. Each snapshot gets an immutable record: fixture ID, league, teams, kickoff time, market, selection, price, source, source timestamp, collection timestamp, and a sequence number. Later snapshots are added, not used to overwrite earlier ones.
7. Settlement attachment. After the match, the final score and settlement status are attached separately. The odds snapshots themselves are not edited. This separation is what allows AlphaBench to evaluate models without hindsight bias.
This pipeline feeds the benchmark scoring process described in the methodology. Models like deepseek-flash from DeepSeek, gpt-5-mini from OpenAI, and grok-build-0.1 from xAI produce probabilistic forecasts. The odds snapshots provide a market baseline and a time-aware record of what information was available before kickoff.
Immutability after kickoff
Immutability after kickoff is the core discipline. Once a match starts, the pre-kickoff snapshot set is locked. No one can add a better-looking price, remove an inconvenient snapshot, or adjust a timestamp after seeing the result. If the data changed after the outcome was known, metrics such as calibration, CLV, and log loss would become meaningless.
The closing snapshot is captured at a defined cutoff before kickoff. That cutoff is part of the benchmark rule, not an after-the-fact choice. For a fixture like Khovd Western vs Deren, the page may show several pre-kickoff snapshots and one closing snapshot. When the match settles, those odds records stay exactly as collected.
AlphaBench may still record corrections, but they are handled as append-only notes rather than silent edits. A correction might flag a mislabelled source or a duplicate feed entry, but the original snapshot remains visible. This makes the benchmark reproducible: another analyst can replay the same snapshots, apply the same scoring rules, and reach the same conclusions.
Immutability also protects the meaning of glossary terms used across the site. Brier score measures the accuracy of probabilistic forecasts. Log loss penalises confident wrong probabilities. Calibration checks whether predicted probabilities match observed frequencies. CLV, or closing line value, compares a model or benchmark price against the closing snapshot as a market-consistency metric, not as betting advice. ROI is a simulated return measure under a fixed staking rule, reported only for benchmarking. Maximum drawdown tracks the largest peak-to-trough decline in a simulated bankroll. None of these metrics work if the underlying odds can be rewritten after the result.
For more on how these metrics are applied, see Brier score for football predictions and How Alpha Bench scores models.
Transparency on fixture pages
Transparency on fixture pages means a reader can audit the benchmark without trusting a hidden spreadsheet. A fixture page should show the odds timeline, the sources represented, the markets collected, the closing snapshot, the model predictions, and the final settlement. For Khovd Western vs Deren in the Premier League, that means the page can display which snapshots existed before kickoff and how deepseek-flash, gpt-5-mini, and grok-build-0.1 were scored after the result.
A useful fixture page typically includes:
- Snapshot timeline. First snapshot, intermediate snapshots, and closing snapshot, each with source and collection time.
- Source coverage. How many distinct feeds contributed and whether any market was thinly covered.
- Market detail. Which selections were available, such as home, draw, away, or over/under lines.
- Model forecasts. The probabilities submitted by each benchmarked model before kickoff.
- Settlement result. The final score and whether the fixture settled normally.
- Scorecard. Metrics such as Brier score, log loss, calibration, CLV, ROI where applicable, and maximum drawdown for longer simulations.
This transparency lets readers move from a single fixture to broader comparisons. The leaderboard aggregates model performance across many settled fixtures, while fixture pages provide the underlying evidence. A model that looks strong on one match can be checked against its calibration over hundreds of matches. A high ROI in a small sample can be weighed against maximum drawdown. A good Brier score can be compared with log loss to see whether the model is merely accurate on average or also appropriately cautious.
The goal is not to declare a permanent winner from one snapshot or one match. The goal is to make the collection process visible, consistent, and repeatable. Odds snapshots are collected before kickoff, frozen after kickoff, and displayed with enough context that anyone can inspect the record.
Related reading
- Methodology — how AlphaBench defines fixtures, snapshots, settlement, and benchmark rules.
- Brier score for football predictions — why probability accuracy needs a proper scoring rule.
- How Alpha Bench scores models — how predictions, odds snapshots, and results become comparable metrics.
- Leaderboard — aggregate model performance across settled fixtures.
This article is for educational benchmarking only and is not betting advice.