Six layers of a prediction system
- Target: define the event, settlement period and output, such as home-win probability or home goals.
- Information cutoff: record what was genuinely known before each historical fixture.
- Features: transform provider data into consistent inputs without using future corrections.
- Model: fit a declared algorithm and preserve its parameters and code version.
- Evaluation: compare later-match probabilities with simple baselines using proper scoring and calibration.
- Publication: state attribution, coverage, refresh timing, failure behaviour and limitations. The evaluation and leakage controls in this sequence are grounded in football benchmark research and scikit-learn's common-pitfalls guidance.
Football model-evaluation research demonstrates why results depend on the benchmark data, target, feature set and protocol. scikit-learn's common-pitfalls guidance documents leakage and inconsistent preprocessing risks.
What model labels do not tell you
"AI", "neural network" and "ensemble" describe broad technical families. They do not reveal the target, data quality, prediction timestamp, calibration, test period or comparison baseline. A simple model can outperform a complex one on a particular sample, and the ranking can reverse after a provider or league change.
| Claim | Evidence needed |
|---|---|
| "The model is accurate" | Named metric, denominator, target, period and comparator |
| "Probabilities are reliable" | Reliability table or diagram and proper score on later fixtures |
| "The system updates live" | Event feed, latency record, state handling and replay test |
| "The model uses injuries" | Timestamped availability source and missing-data behaviour |
| "The service finds value" | Reproducible price capture, margin method, settlement and results |
Calibration guidance explains that forecast probabilities require frequency checks across comparable cases. TimeSeriesSplit provides one framework for preserving chronological order.
Product attribution
The current SportSignals prediction path uses provider-supplied probabilities and performs mapping, price comparison, publication and settled-result reporting. The Sportmonks prediction documentation supports provider-level prediction availability; the SportSignals methodology states the product boundary. This collection does not relabel a provider architecture as an in-house model.
Build an evidence record before choosing a route
The same forecast can support several user questions, but each question needs a different record. Use this minimum evidence map:
| Reader decision | Record that must exist | Page that should answer it |
|---|---|---|
| Understand a probability | Target, cutoff, inputs and calibration | Prediction pipeline |
| Reproduce a model | Data snapshot, code, parameters and split IDs | Build guide |
| Compare algorithms | Same fixtures, features, tuning budget and score | Model comparison |
| Assess live output | Ordered events, receipt latency and replay result | Live predictions |
| Assess a service | Attribution, immutable forecasts, coverage and results | Service evaluation |
| Interpret SportSignals | Provider boundary and maintained product logic | SportSignals methodology |
This division prevents one broad “AI prediction” page from becoming the supposed authority for every question. It also gives search and retrieval systems one canonical answer for each intent.
Minimum release packet
Before a forecast method is described as evaluated, preserve the target contract, data manifest, feature availability rules, chronological split, baseline predictions, candidate predictions, probability scores, calibration table, coverage failures and code version. Football benchmark research supports like-for-like model and feature comparison; scikit-learn's calibration guidance supports checking probability reliability separately from winner classification.
When an artifact is unavailable, name it as unknown. A concise evidence gap is more useful than a confident explanation assembled from unrelated sources.
Assumptions and limitations
This pillar explains methods and evaluation, not guaranteed outcomes. Coverage, provider schemas, prices and product features can change. Worked examples are illustrative unless a dated dataset is cited beside them.
GambleAware
GamCare
Gambling Therapy
GAMSTOP