The Short Answer
A betting model is only as good as the history it learns from, and what matters most is depth, consistency and honest timestamps. You need several seasons of the same competitions recorded the same way, with every event carrying a real match time, so the model trains on the past and is tested on the future. Breadth, more leagues and more exotic markets, is worth little if the underlying events change meaning between seasons or leak information from after the whistle.
What a betting model actually eats
A model does not bet on a team name. It bets on a number: the probability it assigns to an outcome, compared with the probability implied by the odds. Everything upstream of that number is data work, and the data that feeds it best is event data, not match statistics.
Match statistics are totals counted by someone else: shots, passes, possession. Event data is the rows those totals were counted from: every pass, shot, tackle and touch, each with a timestamp, a player and team, an XY pitch location, and qualifiers describing how it happened. In Dribble's own record of Crystal Palace against Arsenal on 24 May 2026 there are 1,574 events, and 920 of them are passes. Shots, the events most summaries are built around, account for just 25.
That ratio is why events matter to a modeller. From totals you can build a model on shots and goals, and that is roughly where most beginner models stop. From events you can build features a totals model cannot see: where shots are taken from, how a team's passing network changes after a substitution, which players receive between the lines, and how often a side creates chances from open play rather than set pieces. Our explainer on football event data walks through what one event row holds, end to end, using the same match.
The practical rule: if a dataset cannot answer a question your model needs, no amount of clever modelling fixes it. Choose the data by the features it can produce, not by the number of leagues on the tin.
Depth: several seasons of the same competitions
The first thing that matters is depth, meaning how far back the history runs for the competitions you actually model. A model trained on one season of a league has seen each team play each other roughly twice. That is enough to fit a model, and it is not enough to trust one. Form, injuries, managerial changes and promoted sides all move faster than a single season can measure.
Dribble's standard history runs six completed seasons back to 2020/21, plus the current season, across 32 competitions: the Premier League, Serie A, La Liga, the Bundesliga, Ligue 1, the Eredivisie, the Primeira Liga, the Champions League, the Europa League and the Conference League, the main domestic cups, and more. Anything before 2020, or a competition outside the standard set, is sold separately through Salad Subscriptions. The history guide at /blog/historical-football-data-xg sets out what those seasons cover and how to use them.
Depth only helps if it is deep in the right place. Ten seasons of one league will usually beat one season of ten leagues for a model that bets on that league, because the model learns how that league behaves, how its teams defend, and how its referees call fouls, rather than averaging across leagues that play differently. Buy the seasons for the competitions you bet on first, and treat extra competitions as a second purchase, not the first.
Consistency: the same definition, every season
The second thing that matters is consistency: the same event, recorded the same way, in every season you train on. This is where models quietly break. If a provider redefines what counts as a tackle, changes its coordinate grid, or starts recording a qualifier it did not record before, the model's history stops being one history. It becomes two datasets stitched together, and the stitch is invisible in the totals.
Three consistency checks to run before you buy:
- Definitions. Is a pass the same thing in 2021 as in 2026, and does the provider document its event vocabulary? A fixed vocabulary of types with an outcome flag is what you want. If definitions drift, your features drift with them.
- Coordinates. Every event should carry a location on one standardised grid, x 0 to 100 from goal line to goal line, y 0 to 100 from touchline to touchline, with the attacking direction normalised. In the Palace and Arsenal sample, Merino's shot in the 79th minute sits at x 94.8, y 48.8, deep in the attacking end. If coordinates change meaning between seasons, every location feature you build is wrong in the same silent way.
- Qualifier depth. Qualifiers carry the difference between a five yard sideways ball and a defence splitting through ball, and they carry derived values too: in Dribble's tables, xG travels as qualifiers on the shot event. Check whether the tier you buy carries the full qualifier set. In Dribble's match_events, the Starter shape carries 19 columns, Pro 25 and Max 31, and the six Max-only columns are period_id, goalmouth_y, goalmouth_z, blocked_x, blocked_y and qualifiers_json. A model that needs shot placement cannot be built on a tier that does not carry it.
Consistency is also about what is missing. A dataset that records lineups and formations in one season and not in another, or that adds a competition mid-history, will hand your model gaps it will happily fill with noise. Ask for the gaps in writing before you pay.
Leakage: the quiet way models lie to themselves
The third thing that matters is leakage, and it is the one that costs modellers the most money. Leakage is information from after the event reaching into a prediction that is supposed to be made before it. It makes a model look brilliant in testing and useless in production, because in production the future is not available yet.
The common leaks in football modelling:
- Training on the future. Splitting matches at random into train and test sets, instead of splitting by time, so the test set contains matches played before some of the training matches. A model that has seen how a season ended will predict its middle far too well. Split by date, always, and keep the most recent season as the honest test.
- Post-shot values as pre-match features. Post-shot xG, the value of a shot as actually struck, is a fine measure of finishing. It is not available before the shot is taken, so it must never sit in a feature that predicts a match. Pre-shot xG is the one that belongs there. The same rule applies to any field that is only known once the action has happened.
- Results baked into inputs. A feature built from a match's final score, or from a table position that already includes that match, is the answer wearing a disguise. Rebuild every feature from events with timestamps at or before the moment the prediction is made.
- Timestamps you cannot trust. This is why honest timestamps matter as much as the events themselves. Every event needs a period and a clock time, and the dataset needs to say when it was recorded. Dribble's data is post-match, refreshed daily once matches complete, and the complete events for a match are typically available within 24 hours of the final whistle. That is a clear, checkable statement. A feed that cannot say when its rows were written cannot prove its features were knowable at kickoff.
One honest limit of our own data, stated plainly: Dribble carries no odds. A betting model needs odds from elsewhere, to compare its probabilities against the market and to measure whether it beats the closing line. What Dribble gives you is the event history to build the probabilities from, not the market to test them against. Any provider that claims its data alone makes a betting model is selling you half the system.
A minimal honest build
If you are starting from scratch, the shape that works is deliberately boring:
- Pick one competition and one market. The 1X2 market in a single league, for example. Depth in one place beats breadth across many.
- Build features from events only. Shots by location, pass networks, set piece threat, and pre-shot xG, all cut by time, all knowable at kickoff.
- Split by date. Train on the older seasons, test on the most recent one, and do not touch the test season while you are still choosing features.
- Compare against the market. Your model earns its keep only if its probabilities differ from the odds in a direction that survives the test season. Without odds in the loop, you have a ranking, not a betting model.
- Keep the pipeline reproducible. The same events, the same definitions, the same cut, every time you re-run it. If a re-run gives a different answer, the data moved, not the model.
None of this needs a large model or a large budget. It needs a dataset whose history is deep enough, consistent enough, and timestamped honestly enough to survive steps 2 and 3. That is the whole test, and it is the test most datasets fail on consistency or leakage long before they fail on size.
Frequently asked questions
What football data do I need to build a betting model? Several seasons of event data for the competitions you bet on, recorded the same way throughout, with timestamps on every event. From events you can build location, sequence and xG features that totals cannot produce. You also need odds from a separate source, to compare your probabilities against the market.
Which football dataset is best for predictive modelling? The one whose history is deepest for your competitions, whose definitions and coordinates stay constant across seasons, and whose timestamps let you prove no feature uses information from after kickoff. Breadth across leagues matters less than depth and consistency in the leagues you model.
How much historical football data does a betting model need? More than one season. A single season gives each team only a couple of meetings with each rival, which is not enough to separate skill from noise. Six seasons, the depth Dribble carries back to 2020/21, lets you train on several seasons and still hold the most recent one out as an honest test.
Does Dribble data include betting odds? No. Dribble carries event, match and player data, post-match, and no odds. You bring odds from your own source and use Dribble's history to build the probabilities you compare them against.
Can I build a betting model from match statistics alone? You can build a simple one, on shots and goals, but it will miss everything events carry: shot locations, passing sequences, set piece threat and chance quality. Statistics are totals counted from events. If you hold the events, you can recount them any way your model needs.
The event counts and the walked row in this article come from Dribble's Max tier sample of Crystal Palace v Arsenal, 24 May 2026 (1,574 events, 920 passes, 25 shots), published in full on Kaggle for anyone who wants to inspect a real event file before buying anything.