Where AI genuinely helps with racing data, where it just invents things, and why a language model needs the table in front of it.
Working SQL and Python for the features most racing models start from, with the leakage traps that catch people out and the thin-history problem.
A finishing time tells you who won. Sectional times tell you how the race was run. Two real races, one UK and one Hong Kong, show what the splits look like.
What the string of figures next to a horse's name means, how racecards record recent runs, and how the same information appears in a database.
A horse race lasts between about a minute and nine minutes. Here are the distances and the winning times that go with them, measured on UK and Hong Kong results.
The Hong Kong racing season starts in early September, ends in mid-July, and has about 88 race days. Here is the pattern, measured from nine seasons of results.
A frank description of the UK and Irish racing archive: how many races, which types, what is recorded for each runner and where coverage thins out.
A reference to the letters that replace a finishing position in race results, in UK and Hong Kong racing, with how often each one occurs.
A reference table converting decimal odds to fractional odds from 1.01 to 1001, with the one-line formula and how prices are stored in racing data.
What to know before you point machine learning at horse racing data: how to frame the problem, the traps in the split, and how thin the history is for many horses.
Recent Comments