How the model works

Horse Truth reads a horse from its own race movement. You should know how much to trust each part, so here it is plainly.

Validated. Tested against real results.

• Class rating. A head-to-head rating across every track. Tested on races it had never seen, it picked the stronger horse about 70% of the time, against 50% for a coin toss.
• Performance rating. Pace, ground and class adjusted. Winners rate above beaten horses in the right order, and stronger grades rate higher.
• Two-lens horse type. Class against figures, so a real class horse is told apart from a false-figure horse whose numbers flatter it.
• Trouble watch. Only signals that held up in the data: a wide trip, or hard early work then beaten close. Both lift the next-start placing.
• Draw read. Built on the real barrier bias here, where inside draws win about twice as often as wide ones.
• Biomechanical read. Where a horse fades, how it spends energy, and whether it grinds late, from the raw splits.
• Closing 600m versus par, and pace dependence. A track-and-distance closing par, and whether a horse needs a truly run race or a soft one.

Experimental. Shown, but not yet proven.

• Soundness watchexperimental A stride-drop flag. A watch signal, not a diagnosis, and not yet validated against outcomes.
• Readinessexperimental First-up and backup spacing patterns. Useful but still rule-based and unproven.

What we will not claim.

• We do not sell tips, selections or predictions. This explains the horse.
• We do not use jockey, weight or odds, because they are not in this data. If a read would need them, we do not make it.
• Cardio, power and toughness are inferred from movement, not measured, and are labelled as inferred.
• One horse does not run from enough barriers to prove a personal draw preference, so we use the population bias, not a made-up one.

Every rating shows its reliability by how many runs support it, so a three-run read never looks as solid as a twenty-run read. Back to Horse Truth