Last updated: 28 May 2026
You spend a weekend pulling apart the form. You find a filter that looks beautiful. Back-tested over five years, it shows 14% ROI and an A/E ratio above 1.15. You bet it for six weeks. You lose money.
Welcome to overfitting. It is the single most common reason horse racing betting systems quietly stop working the moment you start betting them with real cash.
This piece explains what overfitting is, why backtests lie to you about it, and how Race Advisor has built and tested its rating signals so they survive contact with fresh data instead of folding under it.
Quick answer: what is overfitting in a horse racing betting system?
Overfitting is when a system is tuned so tightly to past races that it captures noise instead of signal. The rules look profitable on the backtest, then the edge collapses on fresh races because the patterns it found were specific to that sample, not the sport. Bootstrapping and out-of-sample testing are the fix.
What overfitting actually looks like on a betting system
The story always feels the same. You take a clean window of data, say 2013 to 2016, and you start filtering. Course type, going, class, field size, trainer, rank by your favourite rating. Each filter you add nudges the profit curve a little higher. After enough nudges, the curve looks like a staircase to the moon.
It looks that way because the filters have memorised the sample. Every time you tighten a rule to lift the past profit number, you are baking in a coincidence that was specific to those races. None of those coincidences are obliged to repeat in 2024.
Here is the cleanest demonstration of this we can run on the Race Advisor master dataset. Take every UK trainer with at least 500 runners between 2013 and 2016. Filter to the ones with an A/E ratio above 1.10 across those runners, which is a strong-looking edge. Thirty trainers cleared that bar.
Now ask the obvious question. If you had built a system in 2017 that backed those thirty trainers, how would it have performed across 2020 to 2026? Across the runners those same trainers actually sent out in that later window, just over 26,000 runners in total, the combined A/E collapsed to 1.01. Not above 1.10. Not above 1.05. Bang on the market price. The apparent edge was gone.
This is overfitting in its purest form. The “edge” was a property of the sample, not a property of the sport.
Why a perfect backtest is the warning sign, not the proof
Most bettors treat a clean backtest as the green light. It is closer to a red flag.
A backtest is a question the data already knows the answer to. The runners are settled, the prices are recorded, the wins are stamped. Any rule you point at that data can be tuned until the curve looks lovely, because the rule is allowed to see the answer while it is being chosen.
What matters is the next race, where the answer is not yet known. A system earns the right to your money by holding its A/E on races it has never seen. That is out-of-sample performance, and it is the only number that pays.
A useful rough rule: if your in-sample A/E is significantly higher than your out-of-sample A/E, the difference is overfit. The bigger that gap, the more of your “edge” was noise.
The bootstrapping method that stress-tests a system
The original Race Advisor approach to this was deliberately mechanical. Take your historical data, draw a random 60% slice of it, and build your first rule against that slice. Before you add the second rule, draw a fresh random 60% from the same data and build the second rule against that. Keep rotating the slice every time you touch the system.
Why it works: every rule has to survive a different sample. A coincidence that lived inside one slice has to also live inside another, and another, before it earns its place in the system. Most coincidences do not survive the rotation. Most genuine signals do.
Bootstrapping is the simplest discipline you can apply. It is not the only one. Out-of-sample testing, where you hold back the most recent year of data and never touch it during the build, is the other core technique. The principle is the same in both: never let the rule see the answer to the question it is being asked.
What 13 years of UK racing tells us about signal durability
Race Advisor’s PR and VDW ratings were not built to fit a sample. They were built to rank ability and value in any race, then tested on data the build never saw. The 13-year master dataset gives us a clean way to check whether that worked.
Across all UK runners between July 2013 and March 2026, the top-ranked horse by Race Advisor’s PR rating wins 32.3% of its races, against just under 160,000 runners. The A/E ratio sits at 0.99. Split that same signal into four temporal windows and the picture barely moves.
- 2013 to 2016: 32.0% win rate, A/E 0.99 across 44,401 runners
- 2017 to 2019: 32.8% win rate, A/E 0.99 across 37,032 runners
- 2020 to 2022: 32.3% win rate, A/E 0.99 across 36,898 runners
- 2023 to 2026: 32.2% win rate, A/E 0.99 across 41,440 runners
VDW Rank 1 shows the same flat profile, with win rates between 25.3% and 26.6% in every window and A/E hovering at 0.99 to 1.00 throughout. When PR Rank 1 and VDW Rank 1 agree on the same horse, the win rate lifts to around 35% in every period, on samples between 10,000 and 21,000 runners per window.
Notice what the A/E is telling you. These signals are not magic value flags, they sit at fair market price. What they are telling you is that the ranking has remained meaningful through every chunk of the last 13 years. The signal is durable. That is what a system needs from its inputs before any further filtering is allowed to touch them.
The contrast with the trainer experiment earlier is the entire point. A signal that captures structural truth about a race survives the temporal split. A signal that captures historical performance of a name does not.
The five-minute overfitting check before you bet a system
Before you trust any system, your own or someone else’s, run it through these five questions. They will not give you a yes or no answer, but they will tell you how much of the backtest you should believe.
- How was the data split? A system built and tested on the same data has not been tested. Look for an explicit out-of-sample window or a bootstrapping protocol. If neither is present, treat the backtest as a sales pitch.
- What does the A/E look like on the held-out slice? Not the win rate, not the profit, the A/E to Betfair SP. If it is meaningfully above 1.05 on a sample of at least 1,000 runners that the system has never seen, you have something. Below that, you have a story.
- How many filters does the system stack? Each additional filter multiplies the chance of overfitting. A system with eight conditions in a row is almost always memorising a sample. Two or three robust filters are usually enough if the underlying signal is real.
- Does the sample size survive the filters? If your final filtered sample contains 80 winners over five years, you are looking at noise. The Race Advisor working minimum for trusting an angle is at least a few hundred winners across multiple years.
- Has the edge been re-tested on recent data? Markets adapt. Even a genuine edge can decay. The Race Predictor’s top probability quintile sat at A/E 1.05 across 2013 to 2016 and has drifted to A/E 1.00 across 2023 to 2026 on similar sample sizes. That is not overfitting. It is the market catching up. Either way, the system has to keep proving itself.
The bottom line on building a system that survives
You cannot stop yourself from finding patterns in past races. The brain is too good at it. What you can do is refuse to bet any pattern until it has earned the right by surviving a sample it was not built on. That single discipline separates systems that pay from systems that pose.
If you want to start the next ranking exercise on data that has already done the durability work for you, see today’s PR and VDW ranks on the Race Advisor racecards. They are the same signals that produced the flat A/E across every temporal slice above, applied to runners going to post this afternoon.
