How Bookmakers Make Money on the Grand National: A Data-Driven Analysis

What Is a Value Bet in Horse Racing?

A value bet exists when your estimated probability of a horse winning is higher than the probability implied by the bookmaker's odds. If you believe a horse has a 25% chance but the odds imply only 20%, the price is too generous relative to your assessment. You have found value.

Value is independent of whether the horse actually wins. You can place a value bet and lose. You can place a poor-value bet and win. Over a large enough sample, value bets produce profit. Poor-value bets produce losses. The Grand National is one race, not a large sample, but the value question is still the right one to ask.

The challenge is that bookmakers do not price races at random. They employ teams of traders, observe how professional money moves on the Betfair Exchange, and adjust prices in real time as odds shift throughout the week. By race time, the market has absorbed a great deal of information. Finding a horse that is systematically mispriced against all of that is difficult.

How We Built the Prediction Model

The system works in two stages. We have written about the full process here, but the summary is straightforward.

Stage one: structured data model

We downloaded three years of UK National Hunt race data: results, form figures, going conditions, weights, jockey and trainer statistics, and official BHA ratings. We also pulled historical weather data from Open-Meteo for each racecourse on each race day as a secondary signal for ground conditions.

On top of that data, we trained an XGBoost classifier. XGBoost is a gradient-boosted decision tree model widely used in competitive machine learning for structured prediction tasks. For each runner in a race, the model outputs a win probability. Those probabilities are normalised within each race so they sum to 100%.

Stage two: retrieval layer

Form data does not capture everything. A horse returning from injury, a trainer known to target specific festivals, a jockey with a poor record in big fields. This qualitative information exists in race commentary and tipster analysis, not in the structured numbers.

We built a retrieval layer to access it. Publicly available race previews were embedded into a searchable index. At prediction time, the system retrieves the most relevant commentary for each horse and passes it to a language model, which returns a small probability adjustment with a one-line justification. The adjustment is capped at 10 percentage points either way to prevent the language model from overriding the statistical foundation.

The combination produces a probability per horse, per race. We then compared those probabilities against the bookmaker prices available before the 2026 Grand National.

Zero value bets across 34 runners

After running all 34 Grand National runners through the model at available pre-race prices, we found no horse where our estimated win probability was high enough to constitute a genuine value bet once the bookmaker's margin was factored in.

The table below shows the model's estimated expected loss per £1 staked on some of the main contenders, at starting prices:

Horse	SP	Expected loss per £1
I Am Maximus	9/2	~38p
Panic Attack	7/1	~34p
Jagwar	17/2	~32p
Grangeclare West	9/1	~30p
Johnnywho	12/1	~26p
Iroko	18/1	~20p
Marble Sands	66/1	~8p
Answer To Kayf	100/1	~5p

Model estimates at starting prices. Based on 143% overround with favourite-longshot bias adjustment.

Two things stand out. First, every horse carries a negative expected return. Second, the gradient: shorter prices carry heavier losses per pound staked, while outsiders lose less. The reason for both sits in how bookmaker markets are structured.

What Is Overround? How Bookmakers Price a 34-Runner Race

Take every horse's starting price, convert it to an implied probability (the chance of winning that the price represents), and add them all up. In a fair market, they would sum to exactly 100%. One horse wins. The probabilities should add to one.

They do not. On the 2026 Grand National, the starting prices summed to approximately 143%.

That extra 43% is the overround, sometimes called the vigorish or the book margin. It is the mechanism by which bookmakers guarantee their profit regardless of which horse wins. Every price on the card is shortened from its true value so that the sum of implied probabilities exceeds one.

The margin this represents on bookmaker turnover is calculated as:

Margin = (overround - 100) / overround

At 143% overround: 43/143 = 30%.

On an estimated £250 million wagered nationally on the Grand National, a 30% margin represents approximately £75 million in built-in bookmaker edge before a single race is run. That is a theoretical maximum. It assumes every bet is placed at SP, ignores best odds guaranteed refunds, free bet promotions, and the lower margins available on the Betfair Exchange. The real industry take is lower. But the scale of the structural advantage is clear.

Shopping around on odds comparison sites or using BOG offers reduces the overround you face as a punter. If you take the best available price across a dozen bookmakers, the effective overround might drop to 115-120%. Still above 100%. Still a negative expectation for every bet on the card.

The Breakeven Threshold: When Does a Bookmaker Lose Money?

The overround tells you the bookmaker's aggregate edge across the whole race. But a sharper question exists for individual horses: at what point does the bookmaker actually lose money if a specific horse wins?

The formula is simple.

At decimal odds D, the bookmaker pays out D times the stake on every winning bet. If too much of the total pool concentrates on one horse, the payout exceeds the sum of all losing stakes.

The breakeven threshold is: 1 / D

If more than 1/D of total money wagered goes on the winning horse, the bookmaker loses on that race.

Here is the breakeven threshold for the main runners at their starting prices:

Horse	SP	Decimal	Breakeven threshold
I Am Maximus	9/2	5.5	18.2%
Panic Attack	7/1	8.0	12.5%
Jagwar	17/2	9.5	10.5%
Grangeclare West	9/1	10.0	10.0%
Johnnywho	12/1	13.0	7.7%
Iroko	18/1	19.0	5.3%
Marble Sands	66/1	67.0	1.5%
Answer To Kayf	100/1	101.0	1.0%

The longer the price, the lower the breakeven threshold, and the less money needs to concentrate on that horse for the bookmaker to lose. That sounds like outsiders should be riskier for the bookmaker, but the opposite is true in practice: casual punters rarely put significant money on 100/1 shots. The favourite is where the volume goes.

I Am Maximus: The One Horse That Cost Bookmakers Money

I Am Maximus won the 2024 Grand National, then finished second in 2025 behind Nick Rockett. Returning in 2026 to regain the title, he was available at 10/1 and longer in the ante-post markets.

Then the money arrived.

He was backed steadily from 8/1 through 7/1, 6/1, 11/2 and into a starting price of 9/2 as punters piled on the only horse since Red Rum to regain the Grand National title. At 9/2 (5.5 decimal), the breakeven threshold is 18.2% of total pool money. Market estimates suggest around 20% of Grand National bets went on I Am Maximus on the day, driven by his record and the weight of public sentiment.

He was the only horse on the 34-runner card where that threshold was likely breached. He won.

The Racing Post reported that bookmakers were hit for tens of millions of pounds following the result. Betfred founder Fred Done described the payout as a "gladiatorial bloodbath." William Hill confirmed payouts north of £10 million on the race alone. One punter walked away with approximately £800,000 from a £100,000 bet placed at 8/1 on course.

For any of the other 33 outcomes, the overround holds. The bookmaker pays out, pockets the margin, and moves on. The I Am Maximus result was the one scenario, on the one horse, where the concentration of money exceeded the breakeven point. It happens. It is priced into the long-term business model.

Why Outsiders Look Like Good Value (And Why They Are Not)

The loss-per-pound figures tell a story that bookmaker marketing never does: the margin is not distributed evenly across the field.

On a 100/1 shot, the estimated loss per £1 is around 5p. On the favourite, it is around 38p. Outsider prices are tighter. Favourite prices carry far more margin.

This is deliberate. Casual Grand National punters bet on names they recognise. The previous winner. The horse their colleague mentioned at lunch. A trainer who dominated Cheltenham. Bookmakers know exactly where the volume goes and price accordingly: tight margins on outsiders that rarely need paying out, heavy margins on the horses attracting the big money.

The outsider odds are generous because the bookmaker can afford to be generous on them. If Answer To Kayf wins at 100/1, the payout is large but the number of bets is small. Total liability is manageable. The attractive longshot prices serve a purpose: they draw in casual once-a-year bettors who pick a name off the racecard, place a tenner on something at 66/1, and that money almost never comes back.

The favourites are where the profit is made. The outsiders are the bait.

This is the architecture of a well-constructed betting market. It is not a flaw in the system. It is the system working as designed.

Where Value Betting Genuinely Exists in Horse Racing

If the Grand National is not the place to find value, where is it?

Genuine value in horse racing tends to appear where the bookmaker's information advantage is smaller:

• Betting exchanges: On the Betfair Exchange, you bet against other punters rather than a bookmaker. The overround is close to zero (typically around 102%), though Betfair takes a commission on net winnings (base rate 5%). The Exchange Starting Price is widely considered the most efficient price available.
• Smaller fields with less attention: A midweek novice hurdle at Sedgefield attracts less professional trading than the Grand National. The odds are set less precisely. Edges, where they exist, are more likely to survive.
• In-play markets: Prices move faster than traders can respond during a race. Moments of mispricing occur, though exploiting them requires real-time data and fast execution.
• Early ante-post markets: Ante-post betting months before a race happens in a thinner market. Bookmakers sometimes overprice or underprice before information firms up. The trade-off: you lose your stake if the horse does not run.

On marquee races like the Grand National, with hundreds of millions in turnover and professional quantitative traders active on both sides, the market is about as efficient as horse racing gets. Finding a systematic edge requires either information the market does not have, or a model calibrated well enough to identify mispricings consistently across hundreds of races. We are building toward the latter. On 34 Grand National runners, in a 143% book, it was not close.

Moral of the story? Stick to poker.

Key Takeaways

• The 2026 Grand National starting prices summed to 143% implied probability, representing a 30% margin on bookmaker turnover.
• On an estimated £250 million in bets, that margin represents roughly £75 million in structural bookmaker edge.
• Our XGBoost model with retrieval layer found zero value bets across all 34 runners.
• The breakeven threshold for the favourite was 18.2% of pool money. Estimated actual share was ~20%. I Am Maximus was the only horse where bookmakers likely lost money.
• Bookmakers were hit for tens of millions on the result. That outcome is the exception, not the rule.
• Margin is loaded onto favourites and shorter-priced horses. Outsider prices carry less margin deliberately. This is structural, not accidental.
• The Grand National is one of the most efficient betting markets in UK racing. It is not the place to start looking for value.

Methodology and Data Sources

Race data: Three seasons (2022/23, 2023/24, 2024/25) of UK National Hunt results. Features include last-three finishing positions, beaten lengths, weight carried, days since last run, jockey/trainer strike rates (14-day, course-specific), official BHA ratings, race class, and field size.

Weather: Hourly historical observations from Open-Meteo, matched to racecourse location and race date.

Odds: Starting prices from Sky Sports and Racing Post, cross-referenced against Betfair Exchange closing prices where available. Overround calculated from the full 34-runner SP field.

Retrieval layer: Race commentary and previews from public sources, embedded using a sentence-transformer model, retrieved at prediction time and passed to a language model for probability adjustment capped at plus or minus 10 percentage points.

Estimated pool shares: Derived from exchange market data and industry reports. No bookmaker publishes exact bet distribution per horse, so these figures are approximate and should be treated as indicative.

The model remains in active development. We are expanding through more seasons and running walk-forward backtests across the full National Hunt calendar.