Krok Odds
Analysis

We Built a Machine Learning Model for Australian Racing — Here's What 233,000 Races Taught Us About Market Efficiency

The betting market knows more than most models give it credit for. We trained on 233K races to find out exactly where the edge lives — and where it doesn't.

James Whittaker
James Whittaker
Senior Market Analyst
12 min read·Published 10 Aug 2026·Updated 6 Nov 2026

Most racing models either mirror the market or don't work. We spent the last few weeks building one that does both — and the most interesting findings were about where the market is efficient, not where it isn't.

At Krok Odds, we train machine learning models on Australian racing data to power our racing comparison tool. This isn't a tipster operation or a paid predictions scheme. It's data engineering — building features, testing configurations, and finding the narrow bands where a model actually has an edge over the market.

Here's what we built, what we found, and why the results surprised us.

The dataset: 233,000 races, 1.78 million runners

The model trains on historical Australian racing results from January 2024 to August 2026 — thoroughbred, harness, and greyhound. That's 233,000+ races and 1.78 million individual runners, each with form data, connection statistics, barrier draws, track conditions, and market prices.

The data pipeline is the part nobody sees. We scrape and backfill speed figures and ratings across 80+ Australian race courses, pull exchange prices with near-total coverage, and build per-venue barrier bias statistics from rolling historical results. Sectional data, connection strike rates, and weather conditions round out the feature set.

All of this data is publicly available. No paid APIs. No proprietary feeds. The competitive advantage isn't the data — it's the feature engineering and model selection.

What we tested

We ran 66 different model configurations — LightGBM, CatBoost, XGBoost, logistic regression, and various ensemble approaches — and evaluated 4,462 combinations of features, thresholds, and weighting schemes. The goal wasn't to find one model that works. It was to find the configuration where the model consistently beats the market at a meaningful coverage level.

The final system uses a base scoring layer (applied to every race) plus an ML override that kicks in when confidence is high enough. Three sub-models must agree before the system escalates a pick — and the escalation thresholds are calibrated against backtested strike rates, not arbitrary confidence scores.

The results: tiered confidence with real separation

The model operates across three confidence tiers, each with a different trade-off between coverage (how many races get a pick) and accuracy (how often that pick wins). The figures below are the backtest (January 2024 to August 2026):

TierBacktest Strike RateBacktest CoverageWhat it means
Overall (no filter)42.0%100% of racesEvery pick, baseline
🟢 Value66.0%~20% of racesModel has a reasonable edge
🟡 Standard73.2%~10% of racesModel is fairly sure
🔴 Premium78.8%~5% of racesAll sub-models agree

Those are backtest numbers. Here is what the tiers have actually delivered on live picks since 1 October 2026, priced at the Betfair starting price (SP):

TierPriced PicksLive StrikeSP-Implied StrikeCoverageROI at SP
🟢 Value38365.8%63.9%2.17% of races+5.1%
🟡 Standard11677.6%78.0%0.69% of races−0.4%
🔴 Premium3076.7%86.7%0.15% of races−12.6%

The key insight: the model matches the market at full coverage (42% = always backing the favourite), and on live picks it has not yet beaten the market at any tier. At Premium the SP implied a clearly higher strike rate than the picks delivered — the market priced 86.7% and the picks won 76.7%. At Standard the two agree (78.0% vs 77.6%), and at Value the picks delivered slightly more than the price implied (65.8% vs 63.9%). Only Value is ahead of the price on return (+5.1% at SP over 383 priced picks), and that comes from a broad base — 252 of those 383 picks won, at an average winning price of 1.60 — not from a couple of long shots.

The edge exists. But it's narrow, and it requires discipline to skip most races.

How this compares to just backing the favourite

The favourite wins roughly 42% of Australian races across all codes. That's your baseline. Any model that doesn't beat this at meaningful coverage is adding noise, not signal.

Our model at full coverage produces a 42% strike rate — it matches the market. This is actually a good sign: it means the model isn't adding noise at scale. The separation happens when you filter for confidence:

  • Always back favourite: 42.0% strike rate
  • Model at full coverage: 42.0% (matches the market)
  • Live Value picks: 65.8% (SP implied 63.9%)
  • Live Premium picks: 76.7% (SP implied 86.7% — the market is the sharper end)

The market gets smarter as you filter for confidence — and so does our model. The gap narrows at the sharp end because the market already prices in most of the information our features capture.

What we actually found interesting

The numbers are satisfying, but the findings about the market itself were more useful for future development. Three things stood out:

1. Jockey/trainer combinations are underpriced

We built a sub-model using only connection data — jockey and trainer win rates, shrunk toward base rate to avoid small-sample noise. At very selective thresholds, this model found +23% ROI. Not at high coverage — at the narrow end where you're backing a specific jockey/trainer combination at a specific track.

The market prices the horse. It prices the form. It prices the barrier draw. But it doesn't fully price the combination of who trains and who rides. Stable form — a jockey and trainer hitting a rich vein of form together — shows up in the data before the market adjusts.

2. Class fit is the strongest non-market feature

One feature outperformed everything else: how a horse's recent form compares to the field's median class. Not speed figures. Not sectionals. Not barrier draws. A simple comparison of the level a horse has been racing at versus the level of the current race.

This makes intuitive sense — a horse stepping up from a maiden to a benchmark 64 is a different proposition than one dropping from a Listed race to a benchmark 78. The market prices the horse, but class fit captures the context of the race in a way the market often misses.

3. The market is brutally efficient at 70% weight

When we weighted the model heavily toward market probabilities (70%+), it just mirrored the favourite. The exchange already knows most of what our 63 features know. Genuinely efficient markets don't leave much on the table.

But at 45% market weight — where the model balances market input against independent features — edges emerged. The market has a blind spot, and it's narrow. Finding it requires being willing to disagree with the market when the data says you should, and being willing to agree when it doesn't.

4. Predictability varies by code

Greyhounds are the most predictable code (favourite wins ~44% of the time), while thoroughbreds are the least (~35%). The model performs differently across thoroughbred, harness, and greyhound racing — and the optimal coverage thresholds shift accordingly. A one-size-fits-all approach doesn't work across codes.

What we're doing with it

The model is integrated into our racing comparison tool at krokodds.com.au/racing. It replaces our previous tipping system — which was essentially a market mirror with a ~20% hit rate — with a tiered confidence system that only generates picks when it has genuine backing from multiple sub-models.

No more tipping every race. The model now picks selectively across three tiers, and it's transparent about confidence levels. Premium picks carry the strongest signal but cover the fewest races. Value picks cover more ground but at lower strike rates. Punters decide which tier suits their approach.

What we're not doing

We're not selling tips. We're not taking your money for predictions. The model is a data point in a comparison tool — one input alongside odds comparison, form data, and market analysis. The goal is to give punters better information, not to run a tipping operation.

The margin structure of Australian bookmakers makes it hard enough to profit without adding a subscription fee on top.

The limitations

This is a backtest on historical data. Past performance doesn't guarantee future results — and in racing, it definitely doesn't. The model was trained and evaluated on the same dataset, which means out-of-sample performance could differ. We're collecting live data now to validate.

The coverage numbers were aspirational. The model analyses roughly 3% of Australian races, so Premium fires on 0.15% of the card, Standard on 0.69% and Value on 2.17% — not the 5/10/20% the backtest implied. And a 76.7% strike rate at 0.15% coverage means the model fires on roughly 1 in 670 races, where the market has already priced it at 86.7%. The edge is small and unproven, and bankroll management matters more than model accuracy.

Bottom line

The betting market is more efficient than most people give it credit for. At70% market weighting, our 63-feature model adds nothing over just backing the favourite. The edge lives in the narrow bands where independent features — class fit, connection data, barrier bias — add information the market hasn't fully priced.

If you're building a racing model, the lesson is: don't fight the market. Find the specific places where it's slow, and bet selectively. The edge is real — but it requires the discipline to skip most races and wait for the ones where the data says the market is wrong.

For punters who want to explore the data themselves, the model picks are available on our racing comparison page — free, transparent, and with no strings attached.

Disclaimer: Past performance doesn't guarantee future results. This article describes a backtest on historical data. Always gamble responsibly. If gambling is a problem for you, call 1800 858 858.

James Whittaker
About the author
James Whittaker
Senior Market Analyst

James covers the AU bookmaker market — pricing mechanics, line movement, promotional structures, and how the corporate books actually operate.