Machine learning in betting means building a model that learns its own rules from past results and data. It outputs a probability for each outcome, such as each runner's chance of winning a race, and fair price = 1 / probability turns that into a price you can compare with the bookmakers'. What counts is how well it prices games and races it has never seen.
How a machine learning betting model works
A model takes inputs, called features, and fits them to past results. It is one of the betting tools in Australia that estimate prices rather than place bets:
| Step | Sports example | Horse racing example |
|---|---|---|
| Inputs | Team ratings, home ground, travel, team changes | Past race times, class, barrier, weight, days since last run |
| Output | Each team's chance of winning | Each runner's chance of winning |
Machine learning vs a betting algorithm vs AI
A betting algorithm follows rules a person wrote, and a machine learning model fits its own. AI gets attached to either, and to chatbots, as AI betting bots and AI sports betting show.
| Label | Who sets the rules | What can go wrong |
|---|---|---|
| Betting algorithm | A person, and they stay fixed | The rules contain no edge |
| Machine learning model | Learned from past data | It learns noise as well as pattern |
| AI | Depends on the tool: the wider field machine learning belongs to | Used loosely by sellers, the label says nothing about the method |
Three traps: overfitting, leakage and calibration
Overfitting
An overfitted model has learned last season's noise as well as its pattern. With illustrative numbers, a model picks 68% of head to head winners in the season it was fitted to and 57% in the next, which it never saw. The 11-point drop is noise it memorised.
Leakage
Leakage is information you could not have had when the bet went on: the closing price used to model a bet placed hours earlier, or anything you only learn later, such as a late scratching or the final margin. Fed the closing price, an illustrative model makes 7% in testing and loses 2% live.
Calibration
A calibrated model's 70% outcomes win about 70% of the time. A calibration table checks it (illustrative results):
| Model's rating | Selections | Won | Win rate |
|---|---|---|---|
| 50% | 300 | 147 | 49.0% |
| 70% | 200 | 124 | 62.0% |
| 85% | 100 | 71 | 71.0% |
This model is close at 50% but overconfident above it, so it would back short-priced selections at prices below their worth.
For a single number, use the Brier score: square the gap between each forecast and the result (a win scores 1, a loss 0), then average the squares, so 0 is perfect. Four illustrative forecasts of 0.70 (won), 0.70 (lost), 0.30 (lost) and 0.85 (won) score (0.09 + 0.49 + 0.09 + 0.0225) / 4 = 0.173. Run the same sum on the market's margin-free chances: a model worth using scores lower.
Testing a model against the market
Train on earlier seasons, then test on a later one the model never saw, run as it would have run at the time. Compare its prices with the market's, margin removed.
With illustrative prices, Team A is $1.72 and Team B $2.20: implied chances of 1 / 1.72 = 58.14% and 1 / 2.20 = 45.45%, a book of 103.59%. Dividing both by 1.0359, the simplest method in the fair odds calculator, leaves 56.1% and 43.9%.
Your model gives Team A 57%, more than the market does, yet its own fair price is 1 / 0.57 = $1.75. At $1.72 the bet is worth 0.57 x 1.72 - 1 = -2.0% per dollar, so there is no bet.
High accuracy is not profit either. With illustrative figures, picks that all go off at $1.38 and win 70% of the time return 0.70 x 1.38 = 0.966 per dollar, a 3.4% loss, since the price implies 1 / 1.38 = 72.5%. Betting models and backtesting betting strategies set out the full method.
Where B337 fits
B337 does not pick bets with a model: its strategies follow rules you set, and its racing strategy compares prices, rates no runners and has no machine-learning model. For your own model, API Odds delivers racing and sports odds from Australian bookmakers and Betfair over the B337 API, and places no bets; it is set up with the team and needs coding. Bookmaker names are trade marks of their owners. B337 is not affiliated with them.
Risk: Betting involves risk. A model's probabilities are estimates, a model that tested well can stop working as markets change, and there is no guarantee of profit. See responsible gambling for limits and support.
For free and confidential support call 1800 858 858 or visit gamblinghelponline.org.au.