Probability infrastructure

Racing intelligence, quantified into tradeable probability.

Equimetrics ingests millions of historical races, scores every runner through fourteen validated signals, and emits a calibrated win probability. No tips. No look-ahead. Priced edge, gated by risk.

Request access View methodology

Calibrated · out-of-sample validated · fully audited

Race / 14:05 Kenilworth / R7 LIVE

No. 4 · Highveld Run

ROUTER R-23 · CALIBRATED · WINDOW 31d

18.4%

win probability

Calibration · predicted vs actual ideal model

Form score0.71

Class delta+0.42

Pace fit0.58

Going bias0.33

Risk gate · FIRE market 14.9% · edge +3.5%

2.5M+

Historical records

Leak-free feature store

Validated signals

Out-of-sample tested

Risk-gated routers

Fire / hold decisions

Look-ahead

Strictly point-in-time

The platform

Six parts. One disciplined pipeline.

Every component exists to do one thing: turn raw history into a probability you can price against, and to refuse the bets that don't clear risk.

Calibrated Probability Engine

Model output is isotonic-calibrated, so a stated 20% wins about one race in five. Probabilities you can stake against, not scores.

Isotonic · reliability

Signal Validation Lab

Every signal is tested out-of-sample before it reaches production. Nothing ships on a backtest alone, and nothing is fit to the test window.

Out-of-sample

Risk-Gated Routing

Eighty-three routers convert probability into fire or hold decisions under explicit exposure, bankroll, and drawdown limits. Optimism never overrides the gate.

26 active routers

Live Portfolio Health

Continuous monitoring of calibration drift, hit rate, and drawdown across every active router. Degradation surfaces before it costs anything.

Drift · hit · drawdown

Auto-Discovery Search

A three-phase grid search across all routers — grid, blind-verify, auto-apply — with a deliberate pause so nothing is tuned on data it will later be judged on.

Grid → verify → apply

Full Audit Trail

Every prediction, stake, and outcome is logged immutably. The entire chain — feature to fire to result — is reconstructable, end to end.

Immutable ledger

How it works

Ingest. Predict. Search. Gate. Account.

Five stages, point-in-time throughout. Each runner moves from raw history to a logged, risk-cleared decision — and never sees its own future.

Ingest

2.5M+ historical records normalised into a strictly point-in-time, leak-free feature store.

Predict

Ensemble models emit a raw probability, then isotonic calibration prices it against reality.

Auto-discovery tunes each router out-of-sample, pauses, then blind-verifies before applying.

Gate

Risk rules convert calibrated edge into fire or hold under bankroll and exposure caps.

Account

Stakes, results, and P&L post to an immutable ledger for full reconciliation.

Point-in-time data integrity Every input table queried live — row counts, coverage, and freshness, end to end.

Live table totals — row counts and date coverage for every pipeline table, each marked CURRENT.

Pipeline completeness — monthly runner counts coloured by whether ML race features are complete, partially backfilled, or source-ready.

Signal accuracy

Every signal earns its rank. Out-of-sample.

Fourteen signals compete on one scoreboard — top-1 accuracy measured on races they never trained on, and broken out by racing code.

Leaderboard

Ranked by what actually wins.

PRED_RANK_SCORE_PL leads at 25.1% top-1 accuracy over 4,890 races. The calibrated probability sits mid-table by design — calibration trades a little peak accuracy for honesty. Filter any signal by Flat, AW, NH, Chase, or Hurdle.

4,890 races scored per signal, every one out-of-sample.

Code-level breakouts so a signal that only works on the all-weather can't hide in the blend.

Per-signal top-1 accuracy leaderboard — PRED_RANK_SCORE_PL, GS_SCORE, P_RF, FFS and others, each with rank, race counts and an accuracy sparkline.

Win rate

Better than baseline, and measured.

Highest-FFS runners finish first 21.4% of the time against a ~10% random baseline — a measured edge of roughly +12%, reported month by month rather than asserted once.

3,078 races evaluated, 659 free-form wins.

Monthly breakdown — edge tracked over time, not averaged into one flattering number.

Bankroll discipline

Sized by the bankroll, not the hunch.

Every fire is staked as a fraction of a live bankroll under a sustainability curve. Exposure, drawdown, and growth are tracked continuously — the gate sizes the bet, then keeps watching it.

Fractional staking keyed to calibrated edge and the current bankroll.

Sustainability curve — growth is judged against survival, not a single lucky run.

Bankroll sustainability and growth curve at R100,000 with daily P&L average, bets, wins, top-1 rate and rank.

Domain-agnostic

One probability engine. Any market.

Nothing in the core is specific to racing. The same calibrated-probability, risk-gated pipeline repurposes cleanly to any market where a price implies a probability.

Price-implied edge. Wherever a quoted price implies a probability — equities, futures, options, prediction markets — Equimetrics re-prices it and surfaces the gap.

The same risk gates. Exposure caps, Kelly-fraction staking, and drawdown limits carry over unchanged across every asset class.

The same guarantees. Every signal validated out-of-sample, every decision logged. Discipline does not depend on the domain.

Cross-market edge AllFireHold

MarketModel pImplied pGate

Race · R7 No.4RACING

18.4% 14.9% FIRE

NQ Dec futureFUTURES

61.2% 57.0% FIRE

AAPL 220COPTIONS

33.0% 34.8% HOLD

Election · State APREDICTION

52.5% 47.0% FIRE