eScored Prediction Engine 2.0: Our Biggest Football Model Update Yet

We have released a major upgrade to the eScored football prediction engine. The new system uses XGBoost machine learning to estimate expected goals, followed by a calibrated bivariate Poisson score matrix to calculate match probabilities across the markets available on eScored.

In a 2500-match live-style validation, the updated model reduced total-goal mean absolute error by 16.5%, cut total-goal bias by approximately 96%, and delivered lower Brier scores for 1X2, Over 2.5 and Both Teams to Score predictions compared with our previous Complex Poisson model.

This is not a cosmetic change. We rebuilt the most important part of the prediction pipeline: how eScored estimates the scoring potential of each team before a match.

What changed in the eScored prediction model?

Our previous model calculated expected goals through a complex set of manually weighted football factors. These included recent form, team attack and defence strength, xG, head-to-head data, ELO ratings, player availability, momentum and other adjustments.

That system was transparent and useful, but it had one important weakness: small errors from multiple manual adjustments could accumulate. In the latest benchmark, the legacy model overestimated the combined number of goals by an average of 0.619 goals per match.

The new model takes a different approach:

1. XGBoost evaluates historical team and match features.
2. It predicts separate expected-goal values for the home and away teams.
3. Home and away goal estimates receive independent post-model calibration.
4. A low-correlation bivariate Poisson matrix converts those estimates into score probabilities.
5. The score matrix produces coherent probabilities for 1X2, totals, BTTS, team totals, correct scores and Asian handicaps.
6. Final probability calibration is applied where validation shows a measurable improvement.

This hybrid structure combines the pattern-recognition ability of machine learning with the consistency of a full football score matrix.

XGBoost and Poisson: a hybrid model, not a simple replacement

Although this release introduces an XGBoost football prediction model, Poisson mathematics still plays an important role in the eScored engine.

XGBoost now estimates the two most important inputs:

• expected home goals;
• expected away goals.

The bivariate Poisson layer then transforms those expected goals into a complete distribution of possible scores. This ensures that related markets remain mathematically connected. For example, Match Winner, Double Chance, Over/Under, BTTS and Asian Handicap probabilities all originate from the same underlying view of the match.

The major improvement is therefore not the removal of Poisson. It is the replacement of the old manually constructed expected-goals engine with a data-driven model that learns nonlinear relationships from historical matches.

Live-style testing with pre-match data only

Offline model accuracy is not enough. A football prediction system must also work when predictions are generated under real pre-match conditions.

To test this, we ran a live-style historical backfill. For every match, the system:

• performed fresh XGBoost inference rather than reading a saved prediction;
• used a feature cutoff of two days before kickoff;
• excluded the final result from model inputs;
• rebuilt expected goals and market probabilities through the active production pipeline;
• compared the prediction with the result only after inference was complete.

The live-style results remained close to our offline validation metrics. This gives us greater confidence that the model’s performance is based on genuine pre-match information rather than post-match data leakage.

New XGBoost model vs the previous Complex Poisson model

The following benchmark compares the new production configuration with the previous eScored Complex Poisson model over comparable 2500-match validation runs.

Expected-goals accuracy

Metric	Complex Poisson	New XGBoost model	Improvement
Home-goal MAE	1.076	0.903	16.1% lower
Away-goal MAE	0.953	0.824	13.5% lower
Total-goal MAE	1.420	1.186	16.5% lower
Home-goal bias	+0.376	+0.086	77% lower absolute bias
Away-goal bias	+0.243	−0.064	74% lower absolute bias
Total-goal bias	+0.619	+0.022	96% lower absolute bias

Mean absolute error, or MAE, measures how far predictions are from the final number of goals on average. Lower is better. Bias indicates whether the model systematically predicts too many or too few goals.

The new model’s total bias of +0.022 is particularly important. Across this sample, its average total-goal prediction was almost perfectly centred on the observed results.

Market prediction performance

Market	Metric	Complex Poisson	New XGBoost model	Change
1X2	Hit rate	49.4%	53.8%	+4.4 percentage points
1X2	Brier score	0.2517	0.2339	7.1% lower
Over 2.5	Hit rate	57.6%	65.2%	+7.6 percentage points
Over 2.5	Brier score	0.2565	0.2220	13.5% lower
BTTS	Hit rate	55.4%	66.2%	+10.8 percentage points
BTTS	Brier score	0.2489	0.2210	11.2% lower

The Brier score measures the accuracy of probability forecasts, not only whether the most likely selection won. A lower Brier score means the predicted probabilities were closer to the outcomes that actually occurred.

This distinction matters. A useful football prediction model should not merely select winners—it should estimate uncertainty realistically.

Better Over 2.5 and BTTS predictions

The largest improvements appeared in Over 2.5 goals and Both Teams to Score.

The old Complex Poisson model’s positive goal bias pushed too much probability toward high-scoring outcomes. The new XGBoost expected-goals model, combined with separate home and away calibration, brought the overall goal expectation much closer to reality.

As a result:

• Over 2.5 Brier score improved from 0.2565 to 0.2220;
• BTTS Brier score improved from 0.2489 to 0.2210;
• Over 2.5 hit rate increased to 65.2%;
• BTTS hit rate increased to 66.2%.

These results do not mean every future prediction will be correct. They show that, across the validation sample, the new model produced substantially better probability estimates than the previous system.

More accurate probabilities for favourites and handicaps

During development, we found that the unadjusted model was too conservative with strong favourites and placed too much probability on low-scoring results such as 0–0, 1–0 and 0–1.

We corrected this through several changes:

• separate calibration of home and away expected goals;
• a lower shared-goal correlation in the bivariate score matrix;
• validation of 0–0 and one-goal score probabilities;
• dedicated favourite–longshot tests for 1X2 and Asian Handicap markets;
• post-model 1X2 probability calibration when it reduced out-of-sample Brier score.

In the Asian Handicap −1.5 validation, the original model materially underestimated the home side’s cover probability. After the expected-goals correction, the average predicted probability moved much closer to the observed cover rate and the Brier score improved by approximately 7%.

Performance across major football competitions

The 2500-match validation included matches from a broad selection of competitions, including:

• Premier League;
• La Liga;
• Serie A;
• Bundesliga and 2. Bundesliga;
• Ligue 1;
• Eredivisie;
• Primeira Liga;
• Belgian Pro League;
• Scottish Premiership;
• Süper Lig;
• Ekstraklasa;
• Saudi Pro League;
• Brazilian Serie A;
• Superliga.

The XGBoost model produced a lower total-goal MAE in 13 of the 15 listed competition comparisons. Some of the largest reductions appeared in La Liga, Ligue 1, 2. Bundesliga, Serie A and Eredivisie.

League-level samples are still too small for aggressive competition-specific tuning. We will continue collecting data before introducing any league-level calibration that could overfit short-term results.

What users will notice on eScored

The updated engine affects the full prediction experience:

• more realistic expected home and away goals;
• fewer inflated or overly conservative score projections;
• improved Over 2.5 and Under 2.5 probabilities;
• stronger BTTS probability estimates;
• better separation between favourites and underdogs;
• more consistent Asian Handicap and team-total markets;
• mathematically connected probabilities across related markets.

Because every derived market starts from the same expected-goals distribution, the predictions are designed to tell one coherent story about each match.

What comes next

Prediction Engine 2.0 is a major step, not the end of development. Our next priorities include:

1. expanding untouched live validation across more matches and competitions;
2. monitoring calibration drift over time;
3. increasing league-level sample sizes before considering local adjustments;
4. improving the modelling of exactly one-goal matches;
5. extending transparent accuracy reporting for team totals and Asian handicaps;
6. evaluating new features only when they provide measurable out-of-sample value;
7. retaining the previous model as a benchmark and fallback rather than blending it into the new λ estimates without evidence of an improvement.

We will continue publishing model changes and validation results as the eScored prediction engine evolves.

Frequently asked questions

What football prediction model does eScored use?

eScored now uses a hybrid prediction engine. XGBoost estimates expected home and away goals, while a calibrated bivariate Poisson score matrix converts those estimates into probabilities for match results, totals, BTTS, correct scores and handicaps.

Did eScored completely remove the Poisson model?

No. We replaced the previous manual method of calculating expected goals. Poisson mathematics remains part of the score-distribution layer because it provides a consistent way to connect multiple football markets.

How much more accurate is the new model?

In the 2500-match benchmark, the new model reduced total-goal MAE by 16.5%, reduced absolute total-goal bias by approximately 96%, and lowered Brier scores by 7.1% for 1X2, 13.5% for Over 2.5 and 11.2% for BTTS.

Does a lower Brier score guarantee profitable betting?

No. A lower Brier score indicates more accurate probability forecasts over the tested sample. It does not guarantee the outcome of an individual match or future betting profit. Odds, market margins, variance and responsible bankroll management still matter.

Can football predictions ever be completely accurate?

No. Football contains red cards, injuries, tactical changes, refereeing decisions and random events that cannot be predicted perfectly. eScored probabilities represent estimates, not certainties.

A more accurate and transparent eScored

The move to XGBoost gives eScored a stronger foundation for future football predictions. Compared with the previous Complex Poisson model, the updated engine produces more accurate expected goals, significantly lower bias and better calibrated market probabilities.

Most importantly, the new system has been tested through fresh pre-match inference rather than relying only on saved offline predictions. We believe transparent validation is essential: model updates should be supported by measurable results, not marketing claims.

The upgraded prediction engine is now the primary model powering eScored forecasts.