EachWay.ai is a self-improving AI pricing engine for UK and Irish racing. Six specialist AI analysts price every runner independently of the market, a Bayesian graph hunts for the fingerprints of informed money, and every single prediction is scored against the Betfair market — in public, with proper statistics, forever.
No "guaranteed winners". No screenshots of cherry-picked slips. Just calibrated probabilities, audited daily. 18+ · GambleAware.org
A tipster shouts about the 12/1 winner and quietly forgets the eleven losers. A "win rate" tells you nothing without the odds taken. And no human pundit can honestly tell you the difference between their skill and their luck — because they never do the maths that would expose it.
Most tipping services have no immutable record. Ours is a database, not a Twitter feed: every prediction is timestamped before the off, scored after the line, and never edited.
Anyone can hit 60% of winners by tipping odds-on favourites — and still lose money. The only honest question is: were your probabilities better than the market's? We answer it with the Brier Skill Score, race after race.
"Have a few quid on" is not a staking plan. EachWay.ai sizes every recommendation with a commission-adjusted Kelly criterion, quarter-strength, hard-capped at 3% — the same discipline professional syndicates live by.
EachWay.ai never looks at the odds before forming its opinion. The agents produce tissue prices — independent probability estimates for every runner in the field — and value only exists where our calibrated numbers materially disagree with the market's. That separation is what stops an AI from simply learning to copy the crowd.
"The market is the smartest pundit in the world.
So we made beating it the only score that counts."— the EachWay.ai design principle
Each AI analyst is locked to a distinct analytical identity — it cannot drift into groupthink, and a divergence monitor (Jensen–Shannon, since you asked) forces them apart if they ever start agreeing too much. Diversity is what makes a consensus worth having.
Lives in the ground. Drainage profiles, rainfall ensembles, going changes — knows which horses float on soft and which drown in it.
Class transitions, weight-for-age, official ratings. Spots the horse dropping into a grade it has already outgrown.
Sectional timing, track bias, tactical shapes. Maps how the race will actually be run before it is.
Hunts what the crowd has overlooked — the form angle buried on page four that the market hasn't priced.
Trajectory and fitness. Believes a horse is what it has done lately, and weights the evidence accordingly.
Sire and dam statistics, trainer intent, ownership patterns. Reads the signals connections didn't mean to send.
Forget the equations for a moment (they're here when you want them). The informed-money graph is simple to picture — four ideas, one map:
Tipsters, trainers, jockeys, stables, owners — five kinds of dot, thousands of them, the whole cast of UK & Irish racing on one map.
This tipster keeps tipping that yard. This trainer keeps booking that jockey. This owner's horses keep attracting early money. Each observed pattern is drawn as a thread between two dots — 18 different kinds of thread, including the most telling kind: who moves before the news.
One number per thread: how strongly the evidence says this connection beats the market. Right and early, repeatedly — the thread thickens. Wrong — it thins. Silent — it slowly fades, because in racing, old information is no information. The map re-weighs itself after every single race.
When a runner sits where thick threads converge — a trusted tipster, a telling jockey booking, an owner whose money tends to be early — the system leans in. That convergence is what becomes your informed-money alert.
Thread thickness = strength of belief, on the evidence so far. Thick gold: strong, recent, market-beating. Faint: weak or fading. Dashed: suspected, still on probation. Numbers illustrative.
Honesty isn't a marketing tone here — it's enforced in the database schema. We call it the honesty wall, and it is why professionals and academics take this system seriously.
Predictions are written before the off with a cryptographic prompt hash and scored against actual results with proper scoring rules. There is no mechanism — none — for retroactively editing the record.
If we say 20%, it should happen about one time in five. Our Expected Calibration Error is computed continuously and shown on the dashboard. When we're overconfident, you'll see it before we can spin it.
The system must pass eight simultaneous statistical gates — positive skill on holdout data, significance under permutation testing, calibration quality, positive Closing Line Value — before live execution is even unlockable. Strategies that fail stay in the lab. That bar applies to us before it applies to you.
The system evolves its own analysts — one controlled mutation at a time, champion versus challenger, promoted only on statistically significant improvement. Self-improving, yes. Self-deluding, structurally impossible.
A screenshot of one winning slip. A "97% strike rate". A guarantee. Racing is a near-efficient market — edges are thin, statistical, and earned. Anyone who promises certainty in this sport is selling you their certainty, not the horse's.
What we show instead: full probability distributions over every field, the exact size of our disagreement with the market, our live calibration curves, and the discipline to say "no bet" — which is, far more often than the industry admits, the correct call.
Daily race cards with tissue prices for every runner, the panel's reasoning in plain English, value flags where our numbers beat the market's, informed-money alerts, and Kelly-sized staking guidance. Everything a professional sees, made beautiful.
Racing media, odds platforms, syndicates and operators: ship a calibrated AI racing product under your own brand in weeks. Full API, embeddable components, your colours, our six-layer engine — with the audit trail your compliance team will actually enjoy reading.
EachWay.ai grew out of an academically supervised research programme in probabilistic forecasting and graph-theoretic signal detection. The methods aren't buzzwords bolted on for the brochure — they're the load-bearing walls.
The externally-Bayesian way to fuse expert probability distributions (Genest & Zidek, 1986) — with James–Stein shrinkage so new analysts earn influence rather than inherit it.
Brier Skill Score versus the market baseline, log loss, and Expected Calibration Error. Scoring rules that mathematically reward honesty and punish bluffing.
Beta-Binomial inference with temporal decay over a five-entity network, Leiden community detection, and Benjamini–Hochberg false-discovery control. Informed money leaves footprints; we measure them properly.
Two self-improving loops — bandits for discrete choices, Optuna for continuous parameters — gated so the system cannot optimise itself on noise.
Expanding-window validation with strict temporal ordering and 20% holdout courses. A machine-checked invariant guarantees no future data ever touches a training decision.
Commission-adjusted quarter-Kelly, 3% hard cap, daily loss limits, drawdown tiers that halve stakes automatically, and a severe-drawdown switch that pulls the system back to paper.
Under the hood: 237 Python modules, 11,000+ automated tests, a 161-table production database holding 168,000+ scored probability records, and 13,077 weather forecasts verified against what actually fell at each course — engineering depth most consumer betting products simply don't have.
"Self-improving" is the most abused phrase in AI marketing, so let's be specific. EachWay.ai contains exactly three self-improving elements — three separate loops, each statistically gated so it can only learn from evidence, never from noise.
The orchestrator mutates an underperforming analyst's prompt one variable at a time, runs the challenger alongside the champion, and promotes only on statistically significant improvement. Each analyst's identity is immutable — they get better at their obsession, never converge into one voice.
Thirty-one parameters govern how the informed-money graph detects, decays and believes. Bandit algorithms experiment with the discrete choices, Bayesian optimisation searches the continuous dials, and a drift detector widens the search when the market changes regime. How the graph learns to learn →
Arriving: genetic algorithms evolve whole betting strategies — selection rules, genes and staking policy — under Pareto selection pressure and the same unforgiving rigor harness. The newest loop, and the one you'll get to drive yourself.
All three loops share one constitution: champion vs challenger, significance tests before promotion, and progressive data gates so nothing optimises itself on a small sample. Self-improvement here is an engineering claim, not an adjective.
The engine's next evolution is genetic — literally. Every betting strategy carries a genome: its selection rule, its numeric genes, its staking policy. Genetic algorithms mutate them, cross-breed them, and let a Pareto referee decide which offspring deserve to live. The same nursery is open to you: design your own strategy, then cross it with the engine's champions.
Three chromosomes: a selection rule (a typed, grammar-constrained logic tree — every offspring is valid and leak-free by construction), a vector of numeric genes, and a staking policy. Lineage recorded from birth.
Multi-objective Pareto evolution over return, Closing Line Value, drawdown and turnover — with statistical power as a hard admission constraint, so a tiny-sample fluke can never enter the gene pool, no matter how shiny its backtest.
Quality-diversity search keeps one champion per behaviour niche — front-runners on soft ground, longshot value in big fields, low-turnover grinders. Not one "best" strategy, but a stable of distinct bloodlines.
Two parallel evolution tracks — pure evolutionary computation, and an AI that proposes and recombines rule structures — raced head-to-head on the same fitness harness, the same budget, one combined statistical correction. May the honest one win.
A no-code strategy builder with gene sliders and locked safety rails. Author your own rule, backtest it against the full rigor harness in one click, then cross-breed it with the engine's elites — offspring "by The Engine, out of Your Idea".
An evolutionary search is a multiple-comparisons machine — it will manufacture fake edges unless every trial is counted. Ours counts them all: deflated performance metrics, overfit-probability checks, and a forward out-of-sample clock that is never, ever replayed.
The Strategy Studio is in active development and arrives for founding members first — who get a voice in what it becomes. One thing evolution will never touch: the staking safety rails are locked out of the genome permanently. You set your risk limits; evolution explores ideas within them — it can never loosen them for you.