Every Betting System, Replayed at the Ebor Festival

Seven of the twenty betting systems we track made money at the Ebor Festival.

None of them makes money.

Both of those sentences are true, they are computed from the same settled prices by the same code, and the distance between them is the single most useful thing we can show a punter. It is how every winning betting system that has ever been sold to anybody came to exist: somebody ran a strategy over a short window, found a profit, and stopped looking.

We ran all twenty over York's 28 races and put each one's four-day figure directly next to its all-time figure. The four-day column has winners in it. The all-time column does not, and never has, in three years of measurement.

This is not a criticism of anyone who thought they had found something. Over 28 races, a strategy finishing ahead is unremarkable. Chance produces it routinely. The mistake is not noticing a profit; the mistake is treating a fortnight as evidence.

Every system, both windows

A few rows deserve pointing at.

Backing the favourite in a big field topped the meeting and is one of the more modest losers over the long run. Big-field favourites won repeatedly at York last week. That is what a good week looks like for a strategy that loses slowly.

The favourite variants all landed on the same figure. Backing the favourite, backing it on the Flat, and backing it on soft ground produced identical returns, because at this meeting every race was on the Flat and every race was on ground soft enough to qualify. Three rows, one underlying bet. Anyone counting them as three pieces of evidence is counting the same evidence three times.

Everything built on outsiders returned nothing. Backing the longest price in every race, and backing it each-way, both lost the entire stake: not one outsider won across four days. Over the full sample those strategies lose heavily but not totally, which is the difference between a bad week and a bad idea.

Two rows should be ignored entirely. Odds-on favourites qualified twice and small-field favourites once. A single bet is not a return, and the table marks them rather than letting them sit there looking like results.

Why a good festival proves nothing

The reason a short window produces winners is arithmetic, not luck in any mysterious sense.

A system that loses slowly over the long run still wins a large minority of any short stretch you care to sample. Backing the favourite gives up a few pence in the pound across tens of thousands of races. Over 28 races, the spread of possible outcomes around that small average loss is far wider than the loss itself. Finishing ahead is not surprising; it is the expected experience a good fraction of the time.

Which means the test "did it make money over this meeting?" carries almost no information about the question anyone actually cares about, which is "will it make money next year?"

There is a second trap in the same table, and it is subtler. With twenty systems, you are not asking one question, you are asking twenty. If each has a decent chance of finishing a short window ahead, then getting several winners out of twenty is close to guaranteed even if every single one is a long-run loser. Picking the best row afterwards and calling it a strategy is how backtested systems get sold. The row is real. The selection of the row is the problem.

This is why our board publishes an all-time figure for every system and refuses outright to publish any system showing a gain. It is also why this festival record is deliberately kept out of that board: a short-window positive has to be allowed to exist somewhere, or we would be hiding a true number, but it must never sit in the place where readers go to find out what works.

The useful way to read the table is not top to bottom. It is left to right, one row at a time, asking whether the four-day number would have told you anything about the all-time number. For every row here, it would not.

What we could not replay, and why

Four of the twenty-four systems on the board are absent from this replay. We would rather name them than quietly publish a shorter list that looks complete.

Back the old stagers needs each runner's age. Our published race results carry the horse, the price, the finishing position, the jockey, the trainer and the draw, but not the age.

Festival favourites needs prize money, to decide which races count as festival class. That figure exists in our racing database but has never been exported to the published files.

Follow the AI and the AI's most-confident picks settle from the model's own ledger rather than from a universe of races, so replaying them over an arbitrary subset would not mean the same thing as the board figure it would sit beside.

Two further rows appear in the table but should be read as absent. Backing the favourite over jumps had no qualifying race, because York in August is a Flat meeting. Odds-on favourites and small-field favourites qualified once or twice.

None of this is a limitation we intend to leave alone. The first two are a matter of adding two columns to the published results, and when that happens the record for this meeting will not be rewritten to include them. A published record is frozen; the improvement will show up in the next festival and the difference will be visible.

Betting systems at a festival FAQ

Seven systems made money. Why shouldn't I use one?

Because the same seven lose over every longer window we have measured, and there is no way to know in advance which handful will win the next four days. Choosing the row that won last week is choosing on the basis of the one thing that carries no predictive information.

Isn't a festival a special case where systems might work better?

It is a fair question and the honest answer is that we cannot yet test it. Answering it properly means comparing many festivals against many ordinary weeks, and our published race-by-race results only reach back to June. This record is the first entry in the dataset that would make that comparison possible.

Why do three favourite systems show the same number?

Because at this meeting they were the same bet. Every race was on the Flat and every race was on qualifying ground, so "back the favourite", "back the Flat favourite" and "back the favourite on soft ground" selected the same horse every time. At a mixed meeting they diverge.

How are the systems settled?

One unit per qualifying race at starting price, no commission, with the stake split when two horses share favouritism. Each-way bets use quarter odds throughout rather than the real per-race terms, which is the same convention our season-long board uses so the two columns are comparable. All of that is recorded in the meeting's data rather than left to be inferred.

Does the Lab have any system that works?

No. Twenty-four tested, none in profit over the full sample, and we publish that continuously rather than only when asked. If one ever does show a durable gain we will have a great deal more checking to do before we believe it.

Bet responsibly. Nothing on this page is a tip or betting advice. If you bet, stake only what you can afford to lose, and if it stops being fun, stop. Help is at BeGambleAware.org. 18+.