Do any betting systems actually beat the market?

Everyone who has ever studied a racecard has, at some point, believed they were one rule away from cracking it. Back the favourite — it wins more often than anything else. Back the in-form horse. Follow the longshots for the big-price thrill. Each-way the outsiders so you collect even when you don't win. Stack four jollies into an accumulator and turn a tenner into a payday. There is a betting "system" for every instinct, and each one feels, on a good afternoon, like it works.

We wanted to know whether any of them actually do — not over an afternoon, but over years. So we took 20 popular betting systems and ran every one of them through our racing database: 27,676 real GB races, from October 2023 to June 2026 — roughly 2.7 years of Flat, all-weather and jumps racing. Every bet is settled at the industry Starting Price actually returned, with the awkward outcomes left in: non-runners and withdrawals are treated as refunds (not wins or losses), races with no recorded winner are dropped, joint-favourites split the stake, and fallers, pulled-up and unseated all count as losers — because filtering those out would flatter every number on the page.

The result is a clean, one-sided league table, and that clarity is the value of it. Every single system loses money. The bet that holds up best — backing odds-on favourites — still returns about minus 4.8%; the most popular bet in racing loses about 8.8%; the famous "big-price" angles fare worst of all, and a four-fold of random horses returns barely a third of your stake. We even put our own AI model on the page, backing its top pick in every race, and it lost too — a touch more than simply backing the favourite.

This is the honest version nobody selling a system will show you. Below: why people swear by these systems, the catch that sinks all of them, exactly how we tested, and the full table — every number with its error bars. (Past performance is not a guide to future returns; these figures are SP-only and cover 2.7 years, not a lifetime.)

Betting systems endure because each one is built on something that is genuinely true — just not true enough to beat the price.

Back the favourite. The favourite is the horse the whole market has judged most likely to win, and it duly wins more often than any other runner — about a third of the time across our sample. Backing it feels like backing the form, the trainer, the smart money all at once. The win column ticks over often enough that a losing run feels like bad luck rather than the system.

Back the second favourite. The mirror reasoning: everyone piles onto the jolly, so its price gets too short to be worth it, and the second favourite is nearly as likely to win but pays a bit more. It sounds like getting the same kind of horse at a fairer price.

Back the in-form horse, or the top-rated one. A horse that is well-rated, or that the handicapper makes best in the race, must simply be the best horse — so the folk wisdom says follow it. It scratches the deepest instinct in punting: back the one with the strongest figures.

Each-way the big prices. Back a 16/1 shot each-way and you collect a place even when it doesn't win — insurance built into the bet, the feeling of two ways to be right.

Chase the longshots, or stack the jollies. The purest dreams of all: the 50/1 winner that pays for everything in between, or four short-priced favourites folded into a four-fold that turns a tenner into a small fortune.

Follow the AI. And the modern one: if a computer model ranks the runners, surely backing its number-one pick beats a mug punter squinting at the form?

Every one of these has a kernel of fact behind it. That's exactly why they're so hard to give up — and why testing them honestly, over tens of thousands of races, matters. The feeling of a system working and a system actually working are two very different things, and only one of them shows up in the long-run numbers.

Here is the catch that sinks all of them at once: the price already knows.

Every bookmaker's set of odds on a race adds up to more than 100% — that surplus is the overround, the margin built into the market. Across our 27,676 races it averages about 12% per race, and it grows with the field: small fields are close to fair, while it climbs towards 30% in big fields of 16 or more runners. Even if the prices perfectly reflected each horse's true chance, backing across the field would still lose you that overround. The margin isn't a fee you can dodge by being clever about which horse; it's priced into every single runner.

So a selection rule doesn't change whether you're paying the margin — it only changes which races you pay it on. "Back the favourite" concentrates your bets on short-priced horses; "chase longshots" concentrates them on big-priced ones. Both are still betting into the same loaded book. And because the rules use information everyone can see — who's favourite, who won last time, who's top-rated — the market has already folded that information into the price. You're not finding an edge the market missed; you're re-betting the market's own opinion, minus the margin.

That's why the systems land roughly where they do: they lose about the margin, plus or minus how far each rule tilts you toward overbet runners. It also explains the most surprising pattern in the data: the systems built on the strongest-sounding folk wisdom — chase the longshot, each-way the outsider, stack an accumulator — lose the most, because the crowd has overbet exactly those horses (the favourite-longshot bias), and because multiplying prices multiplies the margin with them.

And the obvious objection — surely a clever model can beat that? — is one we tested directly. We built our own AI on form, conditions and weather to hunt for a fairer price. It still lost. Untangling why is the most important part of this whole study.

We didn't simulate anything. We replayed each system, bet by bet, over real races and the actual Starting Prices returned.

The universe is 27,676 GB races between October 2023 and June 2026 — Flat, all-weather and jumps, every code. For each system we coded its selection rule (e.g. "the favourite", "the longest price in the race", "the top-rated runner", "back the AI's number-one pick"), placed a level notional stake on every qualifying runner, and settled it at the industry SP that horse actually returned. The accumulators (doubles, four-folds, Lucky 15s) are priced analytically off the same single-leg results.

The honesty is in the edge cases, because that's where ROI gets quietly flattered:

Methodology

• Races: 27,676 GB races (Flat, all-weather and jumps), October 2023 – June 2026 (~2.7 years). This is 2.7 years of data — not five, not a lifetime. Data as of 12 July 2026.
• Prices: all returns to industry Starting Price, flat stakes, no commission. We don't hold Betfair BSP, so every figure is the bookmaker SP a punter would actually have got — and a touch worse in the real world once commission and a moving price are added.
• What counts as a bet: a horse that ran and had an SP. Non-runners and withdrawals are refunds, not losing bets (excluded). Races with no recorded winner are excluded. Fallers, pulled-up and unseated count as losing bets — filtering them would inflate every ROI here, and that's exactly the dishonesty we're testing against. Joint-favourites split the stake across the tied runners.
• ROI: return on the stake actually placed, settled to SP.
• Error bars: race-clustered bootstrap 95% confidence intervals — we resample whole races, not individual bets, because runners in the same race are correlated. Thin samples get wide bands, and we say so. The analytic accumulators are expected values, so they carry no bootstrap band.

Two more guardrails. The selection rules use only information available before the off, so there's no hindsight. And the method was fixed before we read the results — we decided what we were measuring, then ran it, so the table isn't the product of hunting for a flattering cut. The engine that produced it is re-runnable, so the numbers can be checked against the database rather than taken on trust. (These figures were corrected after an earlier settlement bug wrongly counted some winnerless races as losses; every number below is the honest, winner-required version.)

Here is the full league table, all-time over the 2.7-year window, ranked best to worst. ROI is to industry Starting Price; the empirical systems carry their race-clustered 95% confidence interval. Read the error bars, not just the point estimates.

(The 6-month and 1-year windows tell the same story — the rankings barely move. That stability is itself a finding: this is structure, not a run of bad luck.)

What the table tells you

Every system loses money. The bet that holds up best — backing odds-on favourites — still returns about minus 4.8% (95% CI [-7.2,-2.4]), roughly the margin and no more. Backing the favourite, the most popular bet in British racing, loses 8.79% (95% CI [-10.2,-7.3]) and does so with remarkable consistency across all three time windows. It wins a third of the time and still goes down, because a 35%-strike-rate horse returning an average of around 5/2 simply doesn't pay enough to cover the two losers in between, once the margin is taken out. Stake £100 on the favourite, race after race, and you are left with about £91.21.

The famous systems lose worst

The angles people are surest about are the ones that cost the most. Backing the outsider — the longest price in every race — wins about 1 race in 33 (3%) and loses 34.7% (95% CI [-40.9,-27.2]): the big-price dream is, statistically, a steady drain. Each-way an outsider loses 30.0%, and the bias compounds in multiples — a four-fold of favourites loses 30.8%, and a four-fold of random horses loses 62.1%, returning barely a third of every pound. The clearest pattern here is the most useful one to know: the more exciting the bet, the more it tends to cost.

Even the AI can't beat it

We put our own model on the same table, and it lost like everything else. Backing the AI's number-one pick in every race lost 11.4% over 6,601 bets — a little worse than simply backing the favourite, because the top pick is usually the favourite or close to it, and the bookmaker has already shaved the value out of those short prices. The model's most-confident picks (the ones it rates 35 percent or better) come in at -0.8%, which looks tantalisingly close to break-even — but that is just 361 bets on a small, jumps-heavy sample, well within the normal swings of luck. It is not a profit and not an edge: take real-world prices and pay commission and it slides back into a clear loss. A model that ranks horses well is still not the same as one that out-prices the market — and ours was actually less accurate than the market (Brier 0.102 versus 0.093).

One honest correction on the jumps favourite

It is tempting to think the favourite is safer over jumps, where class and stamina tell. It isn't — it loses 7.8% there, because so many jumpers fall or pull up that you lose those stakes outright. But it is not the worst favourite bet, as an earlier version of this study wrongly suggested: a settlement bug once counted winnerless races as losses and made the jumps favourite look like the worst of the lot. Corrected, the jumps favourite is one of the better-holding favourite bets — the Flat favourite (-9.4%) and the handicap favourite (-8.9%) are both worse.

Read the error bars

The confidence intervals are the honesty check. Where the sample is large the verdict is firm: the favourite at -8.79% [-10.2,-7.3] is confidently loss-making — the whole interval sits well below zero. Where a system is sliced thin the band widens and you should trust it less: backing the favourite in big 16-plus-runner fields reads -7.0%, but on just 611 bets its confidence interval [-19.8,7.3] runs from a heavy loss all the way into positive territory, so you cannot read any edge into it. A point estimate without its error bar is exactly how a losing system gets sold as a winning one. None of these intervals, on a meaningful sample, crosses into real profit.

The single most useful thing this study settles: no betting system we tested beats the market. Twenty systems, 27,676 races, and the scoreboard reads zero. Every rule loses — from about -4.8% for the bet that holds up best to -62.1% for a four-fold of random horses — and even a purpose-built AI model, backed blind in every race, loses about as much as the favourite.

That isn't bad luck or too small a sample. It's structure. The bookmaker's overround — about 12% per race, climbing towards 30% in big fields — is baked into every price, the selection rules all bet into that same loaded book using information the market has already priced, and the consistency across every time window we cut confirms it. A system reshuffles which races you lose; it cannot change that you lose.

So what is the data actually good for? Not a winning formula — there isn't one to sell, and anyone selling one should be read in light of this table. What it's good for is calibration: knowing that the favourite is a 35%-not-two-thirds proposition, that the outsider's big price flatters a 3% strike rate, that the each-way "insurance" is still a 30.0% cost on a longshot, changes how you read a price. That's the only honest use of a model or a system — as a lens on a horse's true chance, so you can occasionally spot when a price looks wrong, never as a promise that a rule will grind out a profit. It's also why we publish our own model's losing backtest in the open: if a system genuinely worked, no one would be giving it away.

For exactly how every bet is settled — the conventions we hold fixed, and the three backtesting bugs that can fake a winning system — see the full methodology.

For the staking-system version of the same lesson — that doubling your way out of a losing edge just changes how you go broke — see does the Martingale system work.

Bet responsibly. Every system here loses money over time; none is a route to profit, and nothing on this page is a tip or betting advice. Only ever stake what you can afford to lose, and set deposit and time limits. Free, confidential help: GamCare · GambleAware · GAMSTOP. 18+.