Choose a race from the panel to see predictions & analysis.
🏆
Today's Picks
Loading picks...
Loading performance data…
Model Improvements
v3.3 — Model DAVID May 2026
New ensemble model — LightGBM + XGBoost + Logistic Regression retrained on 22 months of data (141k runners)
18 new features: Racing Post Rating, TopSpeed, trainer return-to-form, Betfair exchange vs bookie signal, race freshness — 67 features total (was 49)
Full Card Top Pick — ROI −3.9% → +3.6% (+7.6pp) · WR 31.9% → 36.5% · Max losing run 35 → 18 · 2,776 bets across the test window. Different metric to the daily Best Pick on the Performance tab — that's the model's single top horse per day (~100 picks since launch), not its top horse in every race.
⚡ High Conviction pick added — fires only when model gap between #1 and #2 is ≥10pp (cleaner signal, fewer false picks)
Value Bet strategy under review — not promoted on V3.3 (regressed on recent data, thresholds being re-evaluated)
Betfair Exchange live prices now fetched before every pipeline run (9am and 12pm) — model sees real money odds, not bookie prices
No change to Best Pick model structure — same selection logic, improved by better live pricing signal at run time
18-fold out-of-sample roll-forward confirmed independently by two AI systems: Best Pick +9.6% ROI · High Conviction +17.0% ROI · Best 3 +8.1% ROI
v3.2 21 Apr 2026
Value Bet tightened — min model confidence raised from 8% to 18%
Odds cap introduced: Value Bet limited to 6/4–7/1 (market must agree)
Market rank guard added — only top-4 market horses considered
Minimum EV threshold raised from 3% to 8% — eliminates noise picks
Re-trained David on 400 days of data (was 280 days)
v3.1 April 2026
Added Betfair Exchange live odds as primary odds source
Improved going code mapping (7 levels instead of 5)
49 features across form, market, trainer/jockey, course & race conditions
Walk-forward backtest across 6 months of UK racing data
Expected Value (EV) calculation vs Betfair market
Confirmed Out-of-Sample Results (V3.3)
18 weekly folds, Jan–May 2026. Each fold trains only on past data and tests the following week — no data leakage. Results independently verified by two AI systems (bit-for-bit match).
Best Pick (XGBoost)
+9.6%
ROI · 66.1% win rate · 121 bets
High Conviction (LightGBM)
+17.0%
ROI · 56.4% win rate · 110 bets
Best 3 (Ensemble 20/40/40)
+8.1%
ROI · 60.7% win rate · 366 bets
ROI Value Bet (Exchange odds)
+17.9%
ROI · 26.7% win rate · 105 bets · avg 5.3/1
Full Card Top Pick (LightGBM): +3.6% ROI · 36.5% WR · 2,592 bets
What Is Backtesting?
Backtesting tests how well the model would have performed on historical races it never saw during training.
We use walk-forward validation — train on early data, test on later data, then slide the window forward.
Each row in the table below is one fold: a separate test period with independent results.
A high AUC-ROC (above 0.65) means the model reliably ranks winners above losers.
Backtest Results
—
Folds
—
Avg AUC-ROC
—
Avg ROI
Strategy Performance (avg across folds)
🏆 Best Pick
💰 Value Bet
🎯 Best 3
🥈 Contender
Fold-by-Fold Results
Loading backtest data...
💡
How It Works
A machine learning model that analyses 27 signals and 67 features to find value picks in UK horse racing — every day, fully automated.
The 27 signals the model analyses
📋 Form (weighted recency)⭐ Official Rating📈 Market odds🏇 Jockey win rate (90d)🎯 Trainer win rate (90d)🤝 Trainer-Jockey combo🏞️ Course win rate🔢 Course runs📏 Distance preference🌱 Going preference⏱️ Days since last run🎩 Headgear change⚖️ Weight carried🔵 Draw position👥 Field size🏆 Race class💰 Prize money📊 Handicap flag🐴 Age📡 Expected Value vs market🧮 Log-odds calibration⚡ Racing Post Rating (RPR)🏃 TopSpeed (TS)🔄 Trainer return-to-form rate📊 Betfair Exchange vs bookie ratio🏅 Race prestige (G1–Listed)🌿 Field freshness (days since last run)
How picks are chosen
🏆
Best Pick
The single horse with the highest ML win probability across all of today's races. Scored using XGBoost — confirmed +9.6% ROI / 66.1% win rate in 18-fold out-of-sample testing. One bet per day.
⚡
High Conviction
Top races where the model's top horse has the biggest probability gap over the runner-up — the model is most certain. Scored with LightGBM. Confirmed +17.0% ROI / 56.4% win rate in 18-fold OOS testing.
💰
ROI Value Bet
Where the model's win probability is furthest above what the Betfair Exchange implies. EV calculated using exchange odds (no overround) — confirmed +17.9% ROI in 18-fold OOS testing. A +39.5pp improvement over bookie-odds EV.
🎯
Best 3 Picks
Top 3 horses by ML win probability across today's races. Scored with the full 20/40/40 ensemble. Confirmed +8.1% ROI / 60.7% win rate in 18-fold OOS testing. Ideal for treble accumulators.
The model explained simply
🤖
Strategy-Specific Models (V3.3)
Different strategies use the best-performing algorithm for that task — confirmed across 18 OOS folds. XGBoost for Best Pick, LightGBM for High Conviction, full ensemble (20% LogReg / 40% XGB / 40% LGB) for Best 3 and Value Bet.
📅
Daily Schedule
8am: Racing API pro racecard — 30+ bookmaker odds, Betfair Exchange, RPR, TS, trainer stats in one call. 9am ML pipeline. 12pm re-run with latest market prices. Picks lock at noon — no changes after that.
📡
Betfair Exchange Pricing
ROI Value Bet EV is calculated against Betfair Exchange odds (net of 5% commission) — not bookie prices. Exchange has no overround, so overlays are genuine edges. This single change improved Value Bet ROI by +39.5 percentage points in OOS testing.
🔬
67 Features, 27 Signals
V3.3 added 18 new features from Racing API data: Racing Post Rating, TopSpeed, trainer return-to-form rate, Betfair Exchange vs bookie ratio, race prestige (G1–Listed), and field freshness. Feature importance confirmed exchange odds as the #2 most important predictor.
Frequently asked questions
How are picks selected?▼
Each day the model scores every runner in every UK race using 67 features — form, going, course record, jockey/trainer stats, Racing Post Rating, TopSpeed, Betfair Exchange odds and more. A win probability is assigned to each horse using three ML algorithms. The best-performing algorithm per strategy is used: XGBoost for Best Pick, LightGBM for High Conviction, the full ensemble for Best 3 and Value Bet.
What is a value bet?▼
A value bet is where the model's estimated win probability is higher than the probability implied by the Betfair market odds. If the model says 30% chance and Betfair prices the horse at 4/1 (implying 20%), that's +10% edge — expected value over many bets should be positive even if individual bets lose.
How accurate is it?▼
No model wins every bet — horse racing is inherently uncertain. The goal is positive ROI over a large sample of bets. Actual win rate and P&L figures are shown live on the Performance tab, updated daily with real results. Nothing is cherry-picked or hidden.
When are picks available?▼
Picks go live each morning from 9:00 AM BST. The model re-runs at 12:00 PM with updated Betfair prices, then picks lock — no further changes. The page refreshes automatically every 5 minutes. Results are scraped and updated throughout the afternoon.
Why are odds shown as fractions?▼
UK racing tradition uses fractional odds (e.g. 7/1, 5/2, EVS). The model fetches decimal prices from Betfair Exchange and converts them automatically. EVS means Evens — you win £1 for every £1 staked.
About this project
Built as a concept project during an Oxford University course — for fun and learning only. This is not a commercial tipping service and makes no guarantee of profit.
⚠️ Gambling disclaimer. AJ Racing Tips does not provide financial or betting advice. All picks are generated by an automated model and should not be treated as guaranteed predictions. Gambling involves risk — never bet more than you can afford to lose.