What Is a Confidence Score?

A confidence score is the model's estimated probability of a prediction being correct. When ExPrysm shows a 72% confidence on a Home Win prediction, it means the model estimates a 72% chance that the home team will win that match.

It's not a guarantee — it's a probability. A 72% confidence prediction will be wrong roughly 28% of the time, and that's perfectly normal. The key question is whether the model's stated confidence matches reality over many predictions.

Confidence scores are not the same as odds. A 70% confidence doesn't mean you should bet at any odds. You still need to compare the confidence against the bookmaker's implied probability to find value.

How ExPrysm Calculates Confidence

ExPrysm's confidence scores come directly from the CatBoost gradient boosting model's probability output. Here's how the process works:

1
Feature Input
150+ features are fed into the model: Pi-ratings (Constantinou & Fenton, 2013), ELO ratings, form indices, head-to-head records, home/away splits, and league standings.
2
Probability Output
CatBoost outputs raw probabilities for each outcome (Home/Draw/Away for match result, Yes/No for BTTS, Over/Under for goals). These are the base confidence scores.
3
Calibration
Raw probabilities are calibrated against historical outcomes to ensure that when the model says 70%, it actually wins ~70% of the time. This is what makes the scores reliable.

Confidence vs Accuracy

The real test of a confidence score is whether it matches actual outcomes. Here's how ExPrysm's confidence bands map to real-world accuracy:

Confidence RangeExpected AccuracyObserved AccuracyStatus
50–60%~55%~52%Slightly conservative
60–70%~65%~63%Well calibrated
70–80%~75%~74%Well calibrated
80%+~85%~82%Strong reliability

The close alignment between expected and observed accuracy shows that ExPrysm's model is well-calibrated. When it says 70%, it means 70% — not 60% or 80%.

Higher confidence doesn't always mean better bets. A 60% confidence pick at odds of 2.50 can have more value than an 85% confidence pick at odds of 1.10. Always combine confidence with value analysis.

How to Use Confidence Scores

Filtering Picks

The most straightforward use: set a minimum confidence threshold. For example, only consider predictions with 65%+ confidence. This reduces volume but increases the average accuracy of your selections.

Combining with Value Detection

The most powerful approach is combining confidence with value. A high-confidence prediction that also shows positive edge against bookmaker odds is the strongest signal ExPrysm can give. Look for picks where both confidence and edge are above your thresholds.

Bankroll Allocation

Some bettors scale their stake size based on confidence. Higher confidence picks get larger stakes, lower confidence picks get smaller stakes. This is a simplified version of the Kelly Criterion approach.

Staking Example
50–65% Confidence
0.5 units
Conservative
65–80% Confidence
1.0 units
Standard
80%+ Confidence
1.5 units
Aggressive

What Makes a Well-Calibrated Model

Calibration is the alignment between predicted probabilities and actual outcomes. A perfectly calibrated model would show that among all predictions made with 70% confidence, exactly 70% turn out correct.

Calibration Curves

A calibration curve (or reliability diagram) plots predicted probability on the x-axis against observed frequency on the y-axis. A perfectly calibrated model produces a diagonal line from (0,0) to (1,1). Points above the diagonal mean the model is underconfident; points below mean it's overconfident.

Brier Score

The Brier score measures the mean squared difference between predicted probabilities and actual outcomes (0 or 1). It ranges from 0 (perfect) to 1 (worst). A lower Brier score indicates better calibration and discrimination combined.

Brier ScoreInterpretation
0.00 – 0.15Excellent calibration
0.15 – 0.25Good calibration
0.25 – 0.35Fair calibration
0.35+Poor calibration

ExPrysm's Calibration Performance

ExPrysm tracks calibration across all markets in real-time. The platform's performance data from 7,800+ analyzed matches shows strong calibration across key markets:

54.4%
Match Result (1X2)
80.8%
Double Chance (DC)
55.6%
BTTS
59.7%
Goals O/U 2.5

View detailed calibration curves, Brier scores, and daily accuracy trends on the Performance page.

Tips for Using Confidence Effectively

Don't Chase High Confidence Blindly

An 85% confidence pick on a heavy favorite often comes with very low odds (1.10–1.20). The payout doesn't justify the risk. One loss wipes out many wins. Always check the odds alongside confidence.

Consider the Market Type

Confidence scores behave differently across markets. Double Chance predictions naturally have higher confidence (80%+) because they cover two of three outcomes. Match Result (1X2) predictions rarely exceed 70% because three-way markets are inherently harder. Compare confidence within the same market type, not across markets.

Sample Size Matters

Don't judge the model on 10 predictions. Confidence scores are probabilistic — they describe long-run frequencies. You need at least 100+ predictions in a confidence band to meaningfully evaluate whether the calibration holds.

Combine Multiple Signals

The strongest approach uses confidence as one input among several: confidence score + value edge + league tier + market type. No single metric tells the whole story.

Conclusion

Confidence scores are one of the most useful features on ExPrysm, but only when you understand what they represent. They're calibrated probabilities — not certainties. A well-calibrated 65% is more valuable than an uncalibrated 90% from a random tipster.

Use confidence scores to filter, prioritize, and size your selections. Combine them with value detection for the strongest signals. And always remember: the goal isn't to be right every time — it's to be right often enough, at the right odds, to profit over time.