Statistical Models You Can Use

Why hand-tuned greyhound ratings reach a limit

A rating is a set of rules you have written down: perhaps trap 1 is worth four points, last-run position counts double, and a dog returning after three weeks off loses a point. It works because it distils experience into a formula. The trouble is that the weights are guesses, variables interact in ways that are hard to capture, and there is no objective way to know whether trap 1 is worth four points or six.

Statistical models do not magically know the answer either, but they search the data in a structured way, balancing each variable against the win rates you have actually recorded. The trap draw shows why this matters. At a typical compact UK track over the standard sprint trip, trap 1 wins roughly 20 to 22% of races against a random expectation of 16.7%. The market often knows this, but the size of the effect, and how it changes by track, is exactly what a good model should measure for you.

Logistic regression: the workhorse model for win probability

If you learn only one model, make it logistic regression. You feed in your features and it returns a number between zero and one that you can read directly as a win probability. For each variable it learns a coefficient: positive values push the probability up, negative values pull it down. The results are added together and passed through a squashing function. The logic is gentle enough to follow with a calculator.

You do not need to code the algorithm yourself, as a standard Python library or a stats package in R will do it. The practical steps are: define a clear yes or no outcome, choose a few trusted variables such as trap, recent finishing position, days since last run and track, turn each into numbers or flags, split the data into training and holdout sets, fit on the training set only, then test the probabilities on the holdout. Always keep the 16.7% random baseline alongside, and make sure the model beats it clearly out of sample.

Historical analogues and similar-race lookups

Sometimes the most honest estimate comes from finding past races that looked like tonight's and seeing how runners in them performed. Suppose tonight's race is a six-dog A1 sprint at a track of a given size with the favourite in trap 1. Search your database for races sharing most of those features. If trap 1 won 24% of two thousand similar races, you have a defensible estimate with no coefficients required.

Analogues capture interactions between variables without you having to specify them, and they are easy to explain because you can point at the actual races behind the number. Their weakness is sample size: for unusual grades, distances or smaller meetings, the matching races can shrink to almost nothing. A good habit is to slice on three to five variables such as track, distance, grade, trap and a rough form band, and to treat anything built on fewer than a few hundred races as provisional.

Penalised regression: stopping noise from looking like signal

As you add variables, some will be useful and some pure noise. A standard logistic regression will happily give a noisy variable a strong coefficient if it lines up with a few lucky winners in the training data, and that is how overfitting begins. Penalised or regularised regression is designed to prevent this. Ridge regression shrinks coefficients towards zero, while lasso regression can shrink some all the way to zero, dropping weak variables and leaving the survivors the model considers worth keeping.

For greyhound data, with many plausible variables and not always a lot of clean data per race, penalised models are a sensible default. They also stay interpretable: you can still read each coefficient and ask why it is positive or negative, which is not true of a black-box network. Whichever approach you choose, remember the model is a pricing tool. A 22% probability is a price, and the price tells you whether a bet is worth taking.

Track geometry and why one greyhound model is not enough

Trap draw looks simple but behaves differently from track to track. At compact tracks with a short run to the first bend, the inside traps enjoy a strong advantage. At larger tracks with wider bends and longer runs, that effect is much softer. A single model fitted across every track and distance will average these effects and miss both extremes.

The fix is either to include track as a categorical variable, so the model learns a separate trap effect for each venue, or to fit separate models for groups of similar tracks. This is a quiet advantage over a hand-tuned rating, which gives trap 1 one value everywhere. A model can learn that trap 1 at a tight track is worth more than trap 1 at a bigger, straighter one.

Comparing model prices with the market and checking calibration

A probability is only useful when compared with the market, the consensus of every bettor including sharp ones. The overround costs a non-value bettor roughly 15 to 20 pence per pound staked over time; your model exists to find spots where the market is wrong often enough to overcome that. If your model rates a dog at 25%, fair odds are 3/1. At 4/1 you have a value bet; at 2/1 you do not, however much you like the dog.

If your model says 30% and the market says 12%, the burden of proof sits with the model. Plot your probabilities against actual win rates out of sample: 25% shots winning 25% of the time means good calibration, winning 30% means a conservative model, and winning 20% means an overconfident one that needs attention. Resist opaque tools with longer names; interpretable models let you spot silent errors, such as a track resurfacing, before they cost you a month's bankroll.