Do our lowest-confidence picks have any edge?

No. They run 7 points behind the market they were backed at. MARGINAL −5.7 points vs market (369 picks) · GOOD +1.9 (199 picks). Two thirds of our output is MARGINAL — and since the prompt changed on 1 August, more like five in six. Measured against the prices those picks were actually available at, that tier loses to the market. It is excluded from Top Picks entirely, which is the main reason our published board is often small or empty, and on some days empty outright. We would rather show you nothing than pad the list. The gap between the two tiers is smaller than the numbers suggest once you account for GOOD picks sitting at shorter prices, and it is not yet established: a forward test with its decision rule fixed in advance is running now.

Is there a corner of the market that's softer than the rest?

Not once you account for field size. 538 races · no segment soft on both measures · Group and Listed races the tightest, as expected. We measured how much margin bookmakers charge and how far prices move before the off, by course, class and field size. The first version of this found big fields looked soft. Then we realised books take margin per runner, so that was arithmetic, not opportunity. Adjusted for field size, nothing stands out.

When our model is asked the same race twice, does agreeing mean a better pick?

Two measures disagree. We treat it as unproven. win rate says yes (52.9% vs 40.2% expected) · closing-price test says no (+3.3%, margin of error −6.1% to +13.1%). We ask the model each race more than once and only publish a Top Pick when the answers agree. On strike rate that looks strong. On the closing-price test, which is the harder and more honest measure, it's indistinguishable from nothing, and it doesn't survive correction for the number of angles we tested to find it. We still use it, because a stable answer is worth having on its own terms. We don't count it as an edge.

← Back

How we measure ourselves

Q: Does a favourable draw make a horse worth backing?

The bias is real. It is also already in the price. favourably drawn 12.1% winners vs 8.2% · returns −33.8% vs −37.2% · 1,209 runners each side. Well-drawn horses do win considerably more often, a 48% higher strike rate, which is a big effect. They return almost exactly the same, because draw bias has been in the form guides for decades and the market priced it long before we saw the card. This is the clearest example on the site of a factor being genuinely predictive and completely worthless as a bet.

Current Power rules · Earlier AI research retained below

How the current Power picks work

French courses are excluded from 6 September 2026. We remove the bookmaker margin using the Power method across the full captured field. Eligible races have 8+ morning runners; eligible prices are strictly above evens and no higher than 6/1. We take the highest-ranked eligible horse per race, then the best five across the day. From 19 September 2026 we then back only those five that are in a stated Class 4, 5 or 6 race, renumbering what is left so the headline is the strongest surviving read; unclassified cards, mostly Irish, do not qualify. That usually leaves two or three picks rather than five, and some days none. There is no extra probability cutoff.

Selections and ranks are frozen before publication. Non-runners are void with no replacements. AI provides commentary on the selected horses; it cannot choose, veto or reorder them.

Live returns use 1pt win stakes with simulated Best Odds Guaranteed: the higher of the frozen morning price and official SP, without exchange commission. From 19 September 2026 that morning price is the best captured bookmaker price, with the bookmaker named against each pick on the record; before then it is the median across books. Dead heats divide the payout first; non-runners are void. Confirmed winners await official SP before entering profit and ROI. This assumes BOG eligibility and no Rule 4 deductions; the named bookmaker is not verified as having offered BOG that morning, and earlier medians were not verified as available at any bookmaker. Adopted 9 September 2026 and applied to every live pick from launch. The historical panel remains at BSP less 2% commission. This is a paper trial, not a proven profitable strategy.

The October–July panel backtest replays the current rules, Class 4–6 included, settled at simulated BOG off the panel’s 08:00 best book, so it is directly comparable with the live record. It stays exploratory and is shown separately: the class rule was found on this same panel, so its return is in-sample and will overstate what new races produce. Historical 08:00 prices and field coverage can differ from the live morning capture. The 5 September reconstructed preview is excluded from the live record.

Previous AI strategy: research archive

The short version

Our own data says our tips do not beat the betting market. We've tested it repeatedly and the answer keeps coming back the same. We publish that here because a tipping service that can't show you its failed tests isn't showing you anything.

The previous strategy offered a written verdict on every race and a published record, and tests that were designed before we saw the answer. If you want someone claiming a system that beats the bookmakers, there is a great deal of that available elsewhere.

How the testing works

We write the test down first. Before looking at any result we record what we're measuring, on which races, and what number would count as success. That stops us finding a pattern after the fact and calling it a discovery.

We look once. Checking repeatedly until a run of luck makes the numbers look good is the most common way betting records are fooled, including by the people keeping them.

We compare against the price, not against zero. A 40% strike rate is meaningless without knowing what odds those horses were. Backing favourites produces a high strike rate and loses money. Every figure below is measured against what the market expected from the same horses at the same prices.

We publish it whichever way it lands. Everything on this page is a test we ran hoping for a different answer.

What we've tested

Does our model know anything about a horse that the price doesn't?

No. Its own read is worse than the market's.

249 races · market alone scored best · blending our model in made the forecast worse at every weight

We replayed 249 archived races with every price, fair-odds figure and market-movement note stripped out, and asked the model to rate every runner blind. Its blind opinion correlates only 0.62 with the market, so it genuinely has a mind of its own, but that opinion is worse than the price. Mixing even a tenth of it into the market's own probabilities made predictions no better (a shift of +0.0001, with a margin of error of ±0.008); mixing in more made them clearly worse. Normally the model correlates 0.94 with whatever price it's shown. That isn't laziness. It's the model correctly deferring to better information than its own.

Do our picks get better prices than the market settles on?

No better than same-priced runners picked at random.

+0.4% edge over price-matched runners · margin of error −5.2% to +12.1% · beat the closing price on 36% of picks

Beating the closing price is the fastest honest signal that tips have real value, because it strips out which bets happened to win. Compared against runners at the same prices we didn't pick, our edge is +0.4%, which is statistically indistinguishable from zero. An earlier, larger-looking figure turned out to be our own rule about backing horses whose price was already shortening: following the market, not anticipating it. A full pre-registered test on a larger sample is still running.

Does a favourable draw make a horse worth backing?

The closing-price test has reached its look

The "better prices than the market" finding above is measured on our full history under many different versions of the selection prompt. We've also pre-registered a cleaner version — fixed rules, one population, one look — that only counts picks made from 2 Aug onward, under the prompt running when we wrote the test down. We wrote down what would count as a pass or a fail before collecting any of this data, and we haven't published a result here yet — that write-up happens separately from this counter.

400 of 400

100%

No interim number was ever shown here, before or after the look — that was the rule.

One thing the model is good at, and why it isn't enough

Our picks tend to be horses the market backs after we publish them. They shorten between our 8am price and the off 56% of the time, against 43% for all the other runners on the same cards. Compared like for like — against other horses at the same morning price, because short-priced horses shorten more often anyway — that is 12 points more often than the field.

So the model is not picking at random, and it is not simply copying the price. It has real information about which way a price is going to move.

It still is not enough, and here is the arithmetic. Getting on before the market moves means we take a better price than the starting price. Settled at SP those same picks return −19.6%; settled at the prices we actually took, −12.5%. So this skill is worth 7 points a bet — it is real, it is measured, and the book still loses twelve points after it, because the bookmaker's margin is bigger than the edge.

Measured over 441 settled picks, 11 July to 21 August 2026. It is an observation we made while looking, not a pre-registered test, so treat it as a description of what the model does rather than a claim about what it earns you. The number that matters to you is the record, and the record is on previous AI record page.

Courses we've stopped backing, and how they get back

From 27 August 2026, a pick at a course where our own record is bad enough to rule out breaking even is published but not backed. The AI still reads the race, still names a horse, still gives its reasoning, and the result is still settled and still counted in that course's record. What it does not do is join the track record as a bet, and it can never be the day's stake. Those rows are marked Withheld on the board.

The bar is deliberately hard to clear: a course needs at least 20 settled picks, and the optimistic end of a 99% range on its return has to still be below zero. It is not “this course is losing” — most of them are, most of the time, and on 30 picks that means nothing. It is “we can rule out this course being fine”. Two courses clear it today: Musselburgh and Salisbury.

A suspension is not permanent, and that is why we keep counting the withheld picks. The same test decides both directions: the moment that range no longer sits below zero, the course is backed again. A course cannot earn its way back if we stop scoring it, so we score it.

How much do we think this is worth? Very little, and we would rather say so. Run over the record as it stood, the rule withheld eleven picks, and those eleven lost — against about three and a half winners expected on their prices. It moves the track record by roughly a point. Eleven picks is not proof of anything, and with 68 courses being checked you would expect three to look this bad by chance alone. That is exactly why the bar is 99% rather than the 95% we quote elsewhere, and why we wrote down what would make us drop the rule before we switched it on.

The same test runs in the other direction, and it is how we will ever claim to be good somewhere. A course where even the pessimistic end of that range is a profit — same 99% standard, same 20-pick minimum — earns a “Proven” badge on the course list. No course has earned it yet. That is the honest state of the record, and we show it rather than pointing at whichever course has been lucky lately: on our sample sizes, a course showing a 50% strike rate over 25 picks is usually luck, and the interval is what says so. When a course does clear the bar, the badge appears automatically, and it lapses the moment the record stops supporting it — proof is a state, not a trophy.

The full rules, the numbers behind them, and the dates we review them are in the pre-registrations committed to the repository before the first row changed. Nothing already published has been removed: every pick made before 27 August 2026 is in the record exactly as it went out.

So why publish tips at all?

Because a well-argued read on a race is worth something even when it isn't a licence to bet. Our picks come with the actual reasoning: the form, the conditions, the case for the horse, so you can disagree with it. That is the product.

What we won't do is tell you we have an edge we can't demonstrate. Every gambling business in this market has an incentive to overstate its record. The only defence a reader has is evidence, so we keep publishing ours, including the parts that argue against us.

Bet only what you can afford to lose. Nothing on this site is a prediction of a result.

Findings on this page are dated results from specific tests, not live counters — the only number here that moves is the sample count on the test still running, and a count of picks is not a result. A published finding that quietly moves is one nobody can hold us to. For the always-current record, see our results.