How Power picks work, how results are tracked, and what the research can tell us.
PostTime publishes up to five ranked Power picks each day from races with 8+ morning runners, at median odds above evens through 6/1. Power removes the bookmaker margin and ranks eligible horses; AI explains the selections afterwards.
The public record uses 1pt win singles at BSP after 2% commission on positive profit. The jumps favourites trial, historical backtest and previous AI strategy each have their own record. The live strategy is under evaluation, not proven profitable.
The material below describes the earlier AI selection system. Its methods, totals and results are historical and are not the current Power record.
PostTime analyses every race each morning, names one horse per race, and publishes the horse and the price before the race, not after. Every pick is win-only, and every pick and result is tracked forever, whether it wins or loses.
Most tipping accounts quietly delete their losers and screenshot their winners. PostTime doesn't delete anything. Every tip ever published, and its result, feeds the public track record. If the AI is bad at this, the numbers below will show it. That's the point of publishing them.
Recent form
Finishing positions, distances beaten, and how the horse has performed on similar ground and at similar distances.
Draw bias
Some courses heavily favour high or low draws, especially in large fields. PostTime applies known draw statistics for each track.
Pace scenario
Front-runners perform differently depending on how many other pace horses are in the race. Where running-style data exists for a field we map the likely pace shape, but honestly that data is missing far more often than it's there (currently well under 1% of runners), and when it's absent the AI is told it has no pace read and must not invent one.
Pedigree
Sire and damsire data helps identify horses likely to improve with age, handle soft ground, or excel at certain distances.
Honest probabilities
Every pick shows the market's own estimate of the horse's chance: the probability its price implies, with the bookmaker's margin taken out. It is the market's number, not ours. We used to fold a conviction adjustment into it; that adjustment tracked the price so closely (a correlation of 0.994) that it was the price restated, and it was removed in August 2026. What the AI contributes is which horse in each race we write about — not this percentage, and not a claim that it beats the market.
Our bet means the headline pick: one bet a day, the horse with the biggest de-vigged chance on the board, priced between Evens and 6/1. That single bet is what counts in the tracked record, out of about 37 races. Every other pick is published and scored, just not backed.
Not a bet means everything else: the AI named a horse but it didn't clear the bar, or it passed the race entirely and just told you which horse it makes most likely. Shown because the read is still worth having, but not counted in the record, and not a recommendation.
Every bet is tracked at flat 1pt level stakes, not a bankroll-proportional suggestion. That's a deliberate choice: level stakes is the standard, comparable way to prove an edge, since it can't be gamed by sizing up on "confident" picks after the fact.
Some picks are marked EW. This means you bet to win and to place. If the horse wins you get paid on both parts. If it finishes 2nd or 3rd (depending on field size) you get the place part back. We only suggest each way on longer-priced horses where the place safety net is worth it.
Not every pick makes the tracked Top Picks portfolio:
Everything except the headline. One bet is placed a day. Every other pick is published, settled and shown, but not backed and not in the record. (Three filters used to sit here and none survived: an arithmetic "edge bar", retired July 2026; a rerun-agreement check requiring the same horse on repeated reads, retired August 2026; and the conviction tier itself, retired August 2026 after shuffling its own labels produced the same result. The measured record showed none of them separated winners from losers, so we stopped pretending they did.)
Picks outside the price band. Anything shorter than Evens or longer than 6/1 is published but never backed. Over 6/1 is the one price effect large enough to act on in our own record; the floor is a product choice, taken against the evidence rather than because of it.
Prices outside Evens–6/1. Odds-on favourites are excluded as a product decision (a board of 4/9 shots tells you nothing you can't see in the market), and prices past 6/1 are excluded because the favourite–longshot bias makes long prices systematically poor value at level stakes regardless of any model. Both bounds are documented in the changelog and applied from their ship dates only, so the record is never rewritten.
Filters are reviewed as more data comes in. See the changelog for when they've changed.
Cards are fetched first thing each race day. AI tips are generated and published at 8am UK time each morning, giving you time to get on before the first race.
ROI = total profit ÷ total staked, across every settled bet at flat 1pt level stakes. It's the same number whether a pick wins at 2/1 or 20/1, so a big win at long odds and a string of short-priced winners both count exactly as much as they'd pay out for real, no cherry-picking which bets "count".
CLV (closing line value) compares the price we published against the market's final, most-informed price once all the money is in: the Starting Price, and where we have it the Betfair exchange close. If we consistently get a better price than that, it's evidence of a genuine edge that doesn't depend on luck: beating the closing line is one of the few betting metrics variance can't explain away. A positive CLV with a losing run is still meaningful; a negative CLV with a winning run is a warning sign. It's the number we take most seriously, including when it goes against us, as it currently does.
Even a genuinely profitable long-term edge loses regularly in the short run. That's how variance works in horse racing, where even a "correct" pick only wins a fraction of the time. So a losing streak on its own doesn't prove the model is wrong, but nor does a winning streak prove it's right, and that's the direction the warning usually needs pointing. The homepage shows the longest winning and losing streaks so far, deliberately alongside each other rather than just the flattering one.
3,181 races analysed · 1,383 bets settled so far.
Every pick is tracked and settled against real results. You can see our full record, win rate, ROI and profit/loss, on the Track Record page. No hiding the losers.
We also pre-register tests of our own tips and publish them whichever way they land, including the ones saying we have no edge. See how we measure ourselves.
PostTime is for entertainment only. Tips are AI-generated and not financial advice. Never bet more than you can afford to lose. BeGambleAware.org · 18+
