Building Your First Rating System

Choosing four to eight variables for a greyhound rating

The first decision matters most: which inputs is the model allowed to see? For a first rating, four to eight variables is the right range. Fewer probably ignores information that matters, while more builds something you will struggle to interpret. A helpful approach is to sort variables into three loose buckets: trap variables such as early race position, form variables such as recent finishing positions and consistency, and context variables such as track, distance, grade and going.

Mixing a couple of inputs from each bucket stops any single dimension dominating. Early pace nearly always earns its place, because in short sprint races the dog leading at the first bend has a large advantage; at tight UK sprint tracks, around 60% of races are won by the bend leader. A sensible beginner shortlist covers recent finishing positions with a recency bias, best recent time at the distance, past first-bend position, current grade with step-up or step-down adjustments, trainer strike rate at the venue if captured cleanly, and days since the last race.

Normalising raw numbers into comparable 0 to 100 scores

Raw inputs do not combine well. A best time in seconds and a trainer strike rate in percent live on different scales, so adding them together only adds noise. The next step is to convert every variable to a common scale, usually 0 to 100 with higher meaning better. A spreadsheet column is all you need.

There are three main options. A linear rescale maps each runner's value between the field's minimum and maximum, keeping the real gaps between dogs but reacting badly to one extreme value. Rank-based scoring gives points such as 0, 20, 40, 60, 80 and 100 by position in the field; it is robust and easy to read, though it treats every gap as equal. Z-scores measure distance from the mean in standard deviations, which is statistically tidy but hard to read on a scorecard.

For a first model, rank-based scoring is usually the friendliest, because you can read every line of output and see where each runner sits. Whichever method you choose, apply it within the race, not across your whole database. The question is not whether a dog is good in general, but whether it is good compared with the other dogs in this race today.

Setting weights that reflect what matters in greyhound racing

With every variable on the same scale, you combine them by giving each one a weight. Weights tell the model, for example, that early pace matters more than days since the last run. Give the strongest predictors the largest share: early pace might carry 1.5 or 2.0 at sprint venues, while days since last race might carry 0.5. The exact figures matter less than a defensible view and a sensible total.

A clean method is to start with equal weights and adjust them deliberately using what you already know about the sport. At this stage you are encoding prior knowledge, not optimising on past results. Each runner's rating is simply the sum of weight times normalised score, one formula per dog in a spreadsheet. A useful test is to write your weights on paper beside the screen: if you cannot justify one in a sentence, it is probably wrong for now.

Turning ratings into rankings, market edges and confidence

A rating on its own does not tell you what to back. The simplest output is a ranked list, with your top one or two runners as model picks. A more useful version compares the rating with the market: divide each rating by the race total to get a rough probability, then set it against the bookmaker's implied probability. If your model says 25% and the market says 18%, that gap is worth investigating.

A third output is a confidence band. Flag whether the top runner is clearly ahead or whether the ratings are tightly bunched. A bunched race is hard to call and is probably not worth a large bet. Reading all three outputs together gives a far clearer picture than the ranking alone.

Sanity-checking your rating against real race results

Check the model against reality, not against how it feels. First, measure how often your top pick wins. Favourites win around one race in three, varying between roughly 31% and 42% by venue. If your top pick wins at about the same rate, you have built an expensive way of backing the favourite. Second, check whether your rankings differ from the market. A model that always agrees with the odds cannot find an edge, because the market already carries the bookmaker's margin.

Third, study where the model is wrong. Take the last twenty races your top-rated dog lost and read the form by eye for a trap draw issue, a grade rise or a long lay-off that your variables missed. Avoid fixing the model after a single bad result: look for patterns across at least thirty races before changing any weight or variable.

Why a one-page model description keeps you honest

Write one page describing the system in plain language: the variables, how each is normalised, the weights and the reasons for them, and the intended output. Update it every time you change something. Months later, it turns the spreadsheet said so into a clear explanation you can audit. If you cannot describe the model to a friend in a couple of minutes, simplify before adding anything new.

A good first model uses a few clearly defined variables, gives stable rankings that do not simply copy the favourite, and makes sense in racing terms when you review its winners and losers. It will not be profitable yet, and that is fine. Profit depends on staking, market comparison and steady iteration; the rating is the transparent foundation the rest of the book builds on.