How to Read Football Research and Confidence
Read match reports, NO PICK decisions, xG, confidence, calibration and performance samples.
Reading & evaluating evidence
Read the reasoning behind the headline pick.
This guide follows a fictional report from its identity and evidence through to settlement and review. Learn what the record actually supports, what remains unknown, and where a forecast stops being a valid comparison.
- 01Inspect the report
- 02Challenge its evidence
- 03Separate the metrics
- 04Review fairly
18+ learning material. Every case study, price and probability in these workshops is fictional unless explicitly labelled otherwise. These exercises do not place bets, reserve account funds, or establish that a strategy is profitable. Set affordable time and money limits; taking no bet is a valid decision.
Annotated fictional report
Select a line to unpack its meaning.
Evidence audit
Ask what changed, and what the data can really say.
| Evidence | Useful question | Overclaim to avoid |
|---|---|---|
| Recent results | Were opponents, venue, red cards and schedule comparable? | Five wins prove the next fixture is a win. |
| Expected goals (xG) | Which model, sample and shot context produced the numbers? | 1.8 xG means exactly two goals will be scored. |
| Squad news | Is the source confirmed, time-stamped and relevant to the market? | One missing attacker guarantees Under 2.5. |
| Tactical matchup | Which observable mechanism supports the conclusion? | A team’s reputation alone establishes an edge. |
| Price history | Did the exact market move before or after relevant news? | A shorter price proves the bookmaker knows the result. |
Expected goals assigns a modelled scoring probability to a shot from its characteristics. Shot values can be summed to describe chance quality across a match or sample. They do not replace the score, and different providers can value the same chance differently because their features and models differ.
Read the source: StatsBomb: model inputs and interpretation of expected goals ↗
Three different measurements
Confidence, probability and accuracy are not interchangeable.
| Measure | Example | What it does not establish |
|---|---|---|
| Confidence label | Medium: supporting evidence with unresolved questions | A fixed 60% or 70% event probability. |
| Probability estimate | Home win: 55%, with a stated model and timestamp | That this match must win or that the price offers value. |
| Historical accuracy | 7 correct from 10 eligible settled picks | That each future pick has a known 70% chance. |
| Ticket ROI | £12 net profit from £100 settled stakes = 12% | Whether the underlying forecasts were well calibrated. |
A probability is better evaluated across many predictions made under comparable conditions. If a group of forecasts is repeatedly assigned about 70%, roughly 70% outcomes in a sufficiently informative sample would be consistent with calibration. A short run can differ substantially. Outcomes with the same fixture or source assumptions are also not fully independent evidence.
Advanced example: why Brier score tells a different story
For a binary event, a Brier contribution is (predicted probability − outcome)², with outcome 1 for occurrence and 0 otherwise. A 70% forecast that wins contributes 0.09; if it loses it contributes 0.49. Average these over a defined sample and lower is better for that score. The score reflects several aspects of prediction quality, not calibration alone; inspect reliability by probability groups as well.
Read the source: scikit-learn: reliability diagrams and probability scoring ↗
Reading NovaTide performance
Follow the sample all the way to its denominator.
- Choose the market first. Match result, preferred goals, double chance, BTTS, primary exact score and scenario sets answer different questions.
- Check the period and date basis. The website’s Performance Centre filters on research dates in UTC; an end date includes that whole UTC day. A fixture kickoff date can fall on a different day.
- Read wins, settled selections and pending selections together. A percentage with two settled records deserves different weight from one with two hundred.
- Check which rows count. Earlier superseded forecasts, out-of-scope entries and source-review records can remain visible in an audit without contributing to the headline rate.
- Distinguish OS tracked research from documented pre-kickoff website publication. A post uploaded after a result does not retroactively prove prior public forecasting.
- Use outcome filters to inspect misses, not to rewrite accuracy. On this website, the outcome filter narrows the table; it does not remove losses from the headline calculation.