Hong Kong · Paper test
Horse Racing Predictions
Model probabilities for every runner. Every pick settled in public.
Paper test, no stake placed. Two model configurations, neither converged nor selected. The probabilities below are a forward test sealed before the meeting, not a tip.
Race 1 · ADMIRALTY HANDICAP
Both configurations pick the same horse12 runners · Field entropy 2.45| # | Runner | Jockeys | Trainers | Model win | Model place | Fair odds | SP | Edge at SP | Finish |
|---|---|---|---|---|---|---|---|---|---|
| 6 | ALWAYS FLUKE★ Model pickDraw 4 | 何澤堯 | W Y So | 12.3% | 33.7% | 8.1 | 8.9 | +9% | 5 |
| 10 | TELECOM POWERDraw 6 | 班德禮 | Y S Tsui | 11.8% | 32.7% | 8.5 | 3.2 | -62% | 4 |
| 8 | SOLAR RIVERDraw 5 | 楊明綸 | K H Ting | 9.8% | 28.5% | 10 | 11 | +8% | 11 |
| 12 | TEAM HAPPYDraw 10 | 田泰安 | C W Chang | 9.6% | 28.1% | 10 | 8.2 | -21% | 1 |
| 5 | AUTUMN VIBESMarket favouriteDraw 9 | 潘頓 | B Crawford | 9.6% | 28.0% | 10 | 5.4 | -48% | 3 |
| 7 | NOBLE FANSDraw 3 | 艾兆禮 | J Size | 9.3% | 27.3% | 11 | 11 | +2% | 7 |
| 1 | FLASH STARDraw 11 | 金誠剛 | C S Shum | 6.7% | 21.4% | 15 | 27 | +82% | 12 |
| 11 | VON BAERDraw 12 | 潘明輝 | J Richards | 6.7% | 21.2% | 15 | 11 | -27% | 2 |
| 9 | DASHDraw 7 | 梁家俊 | D J Whyte | 6.6% | 21.1% | 15 | 15 | -1% | 10 |
| 3 | FORTUNE SUPERNOVADraw 2 | 周俊樂 | D J Hall | 6.6% | 21.1% | 15 | 25 | +65% | 9 |
| 2 | MARVEL AND GOLDDraw 8 | 奧爾民 | C H Yip | 5.7% | 18.7% | 18 | 39 | +121% | 6 |
| 4 | STORMY KNIGHTDraw 1 | 艾道拿 | D A Hayes | 5.4% | 18.0% | 19 | 18 | -3% | 8 |
Track & conditions
- TracksHappy Valley · A
- Distance1200m
- ClassClass 5
- GoingGOOD
- PrizeHK$875,000
- Market overround23.5%
Settled ledger · 2026-09-09
WIN
1/8
-74.4%
PLA
4/8
-14.4%
ALL
5/16
-44.4%
Hypothetical $50 per leg on published dividends. No money was staked. Net: $-355
Go deeper
How these numbers are made
Four features per runner — career starts, win rate, place rate and days since the last run — go through a same-race softmax fitted on 1317 development races up to 2026-03. Place probabilities come from a Plackett-Luce model. Two configurations differ only in how missing form is encoded. Neither reached convergence, so nothing here is a signal: the pick is the highest model probability, the edge is that probability against the official starting price, and the ledger is what $50 a leg would have returned.