You can backtest a Betfair strategy, but only the part that depends on outcomes and starting prices — selection systems, lay-the-favourite, dutching by rule. You cannot faithfully backtest in-play scalping or any edge that relies on getting matched, because historical data records what traded, not what you personally could have got filled. Test what the data supports; forward-test the rest on paper.
This page contains affiliate links — if you open an account through them we may earn a commission at no cost to you. It never changes our verdict.
This is a sub of our software & automation pillar, and it tackles the single most over-claimed idea in exchange trading. “Backtested and profitable” is the phrase that sells the most bots and the most strategy PDFs — and most of the time the backtest behind it is measuring something that has nothing to do with what you’d actually achieve live. Backtesting on the Betfair Exchange is genuinely useful, but only if you are brutally clear about which questions the historical data can answer and which it physically cannot.
The short version, before the detail: if your strategy is a rule for selecting bets and holding them to a result, you can backtest it well. If your strategy is a rule for getting in and out of a price before the result, the historical record does not contain the information you’d need to know whether your orders would have filled, so the backtest is a guess dressed up as evidence.
Is backtesting actually possible?
Yes — with a hard caveat. Backtesting is possible and worthwhile for any strategy whose outcome is decided by the result of the event and whose entry price is a published, observable number like the Betfair Starting Price. For those strategies the historical data tells you everything: what you’d have backed or laid, at what price, and whether it won. You can compute a real profit-and-loss curve and a real strike rate.
The caveat is that a large share of what people call “Betfair trading” is not outcome-based at all. Scalping, swing trading and most in-play trading close out before the result, so their profit depends entirely on whether you got matched at your chosen prices — and that is precisely the thing historical data can’t confirm. So “is backtesting possible?” splits into two answers: a confident yes for selection systems, and a heavily qualified “not really” for execution-dependent trading.
What Betfair historical data does and doesn’t contain
To understand the limit you have to know what the data actually is. Betfair’s historical data — whether the official Historic Data service or the third-party archives covered in our historical data guide — records, at intervals, the available-to-back and available-to-lay prices, the sizes queued at them, and the total amount traded at each price. After the event it records the result. That is a rich record, and it is enough for a lot of analysis.
Here is what it does not contain, and why it matters. It does not record your order. So when the data shows £40,000 traded at 3.5, it cannot tell you whether your £50 lay would have been matched there, where you’d have sat in the queue, or whether the price touched 3.5 for half a second on a single tick that you’d never have reacted to in time. It records that a price was available, not that you specifically could have transacted at it after the market saw your order. For outcome bets this gap is irrelevant — you take BSP or a clearly liquid price and hold. For scalping it is the entire game.
The other thing the data flattens is time granularity. Even one-second snapshots miss the sub-second flicker where in-play prices actually move, so any in-play backtest is reading a blurred photo of a fast event and pretending it’s a video.
Strategies that backtest cleanly
Some strategies map perfectly onto what the data records, and you should backtest these properly before risking a penny.
Lay-the-favourite and back-the-favourite systems
A rule like “lay the SP favourite in UK handicaps over 12 runners” is pure outcome plus published price. Every variable is in the data: the favourite, its Betfair SP, the result. You can replay thousands of races and get an honest P&L, drawdown and strike rate.
Dutching and rule-based selection
Any dutching system that stakes across selections by a fixed rule at known prices backtests fine, because both the prices and the result are recorded. So does any model-driven selection system — the kind discussed in our quantitative trading and machine-learning pieces — provided you enter at SP or a clearly liquid published price.
Drift-and-steamer systems on observed moves
A rule based on the price at a fixed time before the off versus the SP — “back anything that shortened more than 20% in the last ten minutes” — is backtestable, because you’re comparing two recorded prices and a recorded result. You’re not assuming you got matched mid-move; you’re using snapshot prices the data actually holds.
Strategies that don’t — and why
Now the uncomfortable half. These do not backtest honestly, and any tool that claims they do is selling you a number that won’t survive contact with the live market.
Scalping. A scalp is two or three ticks of profit captured by getting matched on both sides. Whether you get matched depends on queue position, on whether your order moved the price, and on sub-second timing the data doesn’t hold. A backtest that assumes you filled every time the price was available is assuming away the only thing that determines whether scalping is profitable.
In-play momentum and overreaction fades. These depend on reacting inside the few seconds when a price gaps after a goal or a break of serve. Snapshot data smooths that gap; you cannot tell whether you’d have got the fill you’re crediting yourself with. The edge described in our overreactions piece is real but it can only be forward-tested, not backtested.
Anything in a thin market. Where liquidity is low, your own order is a meaningful fraction of the book, so you move the price you’re trying to hit. The historical data, which never saw your order, will always overstate how well you’d have done.
The idea: a simple, fully outcome-based rule — lay the Betfair SP favourite in UK & Irish handicaps with 14+ runners, level stakes to £10 liability. Exactly the kind of thing that should backtest cleanly.
The test: I pulled a season of qualifying races from historical data — just over 600 races. For each I recorded the favourite’s BSP and the result, then computed the lay P&L after 5% commission on winning lays.
The raw result: the system showed a £74 profit across 600+ races — favourites in big handicap fields lose often enough that laying them at SP edged ahead. Tempting. A strike rate of roughly 72% of favourites failing to win.
The reality check: the maximum drawdown along the way was -£128 — nearly thirteen losing liabilities in a stretch — against a final profit of £74. The edge was real but tiny and the variance was brutal: a 14-bet losing run would have shaken out most people before the profit arrived. I also re-ran it taking SP minus realistic slippage, and the £74 shrank to roughly £31.
The lesson: this strategy genuinely backtests — every variable was in the data — and the test still earned its keep, not by saying “profitable” but by showing the drawdown was nearly twice the profit. That’s the real value of a backtest: not a green number, but an honest picture of how ugly the road is. I didn’t trade it; the edge was too thin to survive my own discipline lapses.
The five pitfalls that produce fake edges
Even on backtestable strategies, it’s easy to manufacture an edge that doesn’t exist. Five mistakes do most of the damage.
1. Assuming fills. Crediting yourself the available price as if your order filled instantly and in full. The number-one cause of backtests that don’t replicate live.
2. Ignoring commission. A strategy that’s flat before commission is a loser after it. Always net it, and if you’re a high-volume trader, consider the Premium Charge too.
3. Curve-fitting. Tuning “14+ runners” to “13+” to “11-15” until the curve looks best. You’re fitting noise. Set rules in advance and resist optimising on the same data.
4. Survivorship and selection bias. Testing only races/markets you can easily get data for, or quietly dropping the ones that broke your rule.
5. Too small a sample. 50 races prove nothing. Thin edges need thousands of events before the result means anything, a point we hammer in the realistic income piece.
A workflow that doesn’t lie to you
Here’s the process I actually use, designed to fail honestly. First, classify the strategy: is it outcome-based (backtestable) or execution-dependent (not)? If it’s execution-dependent, skip straight to paper trading — don’t fake a backtest. Second, define the rules in writing before you touch the data, including entry price assumption, stake, and commission. Third, run it on a large out-of-period sample and look at drawdown and the worst losing run, not just the final figure. Fourth, haircut the result — assume worse fills than the data shows and re-run. Fifth, forward-test on paper or tiny stakes for a meaningful number of bets before scaling, because the live market is the only honest judge of whether you actually get matched.
For anything you want to automate, the same discipline carries into the bot: the logic behind building trading algorithms and Betfair bots should be backtested where the data supports it and paper-traded where it doesn’t — never deployed on a backtest alone. Tools like Bet Angel can replay markets, but a replay still can’t prove your order would have filled.
The verdict
Backtesting Betfair strategies is possible and valuable for the outcome-based half of trading — lay/back selection systems, dutching, model-driven picks entered at published prices — and close to worthless for the execution-dependent half, because the historical data records what the market traded, not what you personally could have got matched. The mature position is to backtest what the data legitimately supports, paper-test everything else, haircut every result, and care more about drawdown than about the final green number. A backtest is a filter for bad ideas, not a guarantee of good ones. Build the habit alongside the automation pillar, paper trading and our data analysis guide.
A profitable backtest is not a profitable strategy. Live fills, slippage, commission and your own discipline all erode results, and most Betfair traders lose money overall. Past results never guarantee future returns. Treat every backtest as a hypothesis to disprove, not a promise. 18+ only; help at BeGambleAware.org.
Test what the data supports, paper-trade the rest. Start with the automation pillar.
Automation Pillar Open Betfair Account →FAQ
Can you backtest a Betfair trading strategy?
Yes for outcome-based strategies — lay/back selection systems, dutching, model picks entered at the Betfair SP or a clearly liquid published price — because the historical data records the prices and the result. No, not honestly, for scalping or in-play trading that depends on getting matched, because the data records what the market traded, not whether your specific order would have filled.
What data do you need to backtest Betfair strategies?
Betfair historical data (the official Historic Data service or a third-party archive) giving time-stamped available-to-back/lay prices, queued sizes, total traded at each price, and the result. That’s enough for outcome-based systems. It is not enough to prove fills for execution-dependent trading, which no historical dataset can do.
Why can’t you backtest scalping on Betfair?
Because a scalp’s profit depends entirely on getting matched on both sides at your chosen ticks, and the historical data never saw your order. It can’t tell you your queue position, whether your order moved the price, or whether a price you’re claiming existed for only a fraction of a second. A scalping backtest assumes away the one thing that decides if it works.
How many bets do you need for a Betfair backtest to be meaningful?
For thin edges, thousands of events. A few dozen races prove nothing — variance alone can make a losing system look profitable over 50 bets. Just as important as sample size is looking at the maximum drawdown and worst losing run, because a small edge with a huge drawdown is untradeable in practice even if the final number is positive.
Related reading
Go deeper with the software & automation pillar, learn what you can do without a backtest in paper trading, source the data in our historical data guide, and see the modelling side in quantitative trading. For execution, study scalping and building bots.