How NextGoalAI Evaluates Prediction Accuracy

Prediction History shows what happened to selections previously published by NextGoalAI. Its accuracy figure is a retrospective summary of evaluated prediction records, not a forecast of future success or a financial return. This guide explains the numerator, denominator and reporting period behind the displayed percentage.

Which outcomes count?

• PENDING — the prediction does not yet have a confirmed evaluation; it may appear in the History list but is excluded from the accuracy calculation.
• WON — the confirmed evaluated outcome matches the selected HOME, AWAY or NO GOAL prediction.
• LOST — the confirmed evaluated outcome does not match the selection.

The public evaluation rules require a resolved result. A provisional change on a live scoreboard is not, by itself, a confirmed WON or LOST result.

Accuracy is calculated per prediction

For a selected reporting period, the formula is WON ÷ (WON + LOST) × 100. The website rounds the result to one decimal place. The denominator includes confirmed evaluated predictions only, not PENDING rows.

Hypothetical calculation — not current NextGoalAI performance

• 10 published predictions have confirmed results.
• 7 are WON and 3 are LOST.
• Accuracy = 7 ÷ (7 + 3) × 100 = 70.0%.
• If two more records are still PENDING, those records are not included in the denominator.

This is a per-prediction calculation, not a per-match calculation. If a single fixture has two separately published, evaluated predictions, both can contribute to the counts. This differs from counting unique fixtures.

Why the list and summary counters can differ

History can display PENDING predictions while the summary label Evaluated Predictions counts only WON plus LOST. A paginated list may show fewer records than the summary total because it displays only a page of matching rows. Do not sum visible rows from one page and expect them to equal the complete-period totals.

The selected period also changes the population. Compare figures only when the same filters, period selection and record definitions are used.

How the date windows work

History uses Europe/London calendar boundaries, rather than each visitor’s device timezone. Today refers to the current London calendar day up to the time of the query. In the current product, 90 Days summarizes the preceding completed days and excludes Today. The interactive History summary identifies the cohort through period labels (such as Today or 90 Days) and Europe/London context; it does not currently display a separate start/end date pair beside those controls.

The public History view also applies a public-data start cutoff. Consequently, a 90-day selection may cover fewer than 90 complete days of records if the public dataset began more recently. The effective reporting period may therefore be shorter than the selected label alone suggests.

Build-time snapshots versus interactive results

The crawlable History overview may include a verified snapshot generated when the website was built. Its timestamp and reporting window identify the data it summarizes. The interactive History filters can show newer results or a different period. A build-time snapshot is not a live ticker and should never be read as the latest evaluation count without checking its timestamp.

What accuracy cannot demonstrate on its own

• It does not guarantee that an individual future prediction will be correct.
• It does not measure financial returns or return on investment.
• It does not mean the same result holds for all leagues, minutes or outcome categories.
• It describes the evaluated predictions in the selected sample, not every conceivable match or unpublished model assessment.

For current verified counts and the individual records behind them, visit Prediction History. The example above is illustrative and deliberately does not hard-code a marketing success rate.

Related Insights

• Understanding Next Goal Probabilities
• Probability vs Confidence
• Understanding NO GOAL Predictions

All Insights · How It Works · History · Live