What a week for the Pick Pony AI engine! Across 846 races covered, our algorithms delivered consistently strong results from top to bottom of the card. The headline performer was Algo 2 (AI Ensemble), which posted an outstanding 37.4% win rate on top picks — virtually matching the morning-line chalk baseline of 37.5% — while crushing it in the in-the-money department with a remarkable 70.9% ITM rate and 274 outright wins. Its morning update version was even sharper, hitting 38.5% wins and 71.1% ITM across 460 settled races, a clear beat on the chalk baseline. The flagship AI Exacta model also impressed, notching a 36.7% win rate and 69.5% ITM across 725 settled races, with 69 straight exacta hits and 71 boxed trifectas. Even our experimental models showed their teeth: Experimental Weight Set 2 delivered 252 wins on its morning update pass, while Weight Set 3 led all models with an outstanding 66.4% ITM rate on its base run and 53 boxed exacta hits. With 35 races still pending settlement across all models, those final numbers could tick even higher. All told, this was a week that demonstrated the depth and consistency of the Pick Pony AI suite across a high volume of races.
Standout Wins
The highlight reel this week is absolutely loaded with value. HARD LUCK PRAYER (ML 5-1) stormed home at IND on September 11, a top-pick win flagged by the AI Exacta, Experimental Weight Set 3, and Algo 2 — a powerful cross-model consensus that rewarded bettors who followed the signal. WINSOR GEORGE (ML 4-1) at LRC and SARAH'S RUNNER (ML 4-1) at EVD were both nailed by the AI Exacta and Algo 2, underlining the Ensemble model's ability to zero in on value plays that the public underestimates. On the longshot front, Algo 1 had an exceptional week identifying overlooked horses: POP D'ORO (ML 15-1) at EMD on September 6 is the pick of the week — a massive overlay that paid handsomely — joined by SERENAS GHOST (ML 8-1) and HOLLYWOOD ICON (ML 8-1), both landing as top picks on September 11. STUDY SESSION (ML 6-1) at DMR appeared across Algo 1, Experimental Weight Sets 2 and 3, and even the morning updates — a model consensus longshot win that was as satisfying as it gets. Algo 3 (AI Win Probability) chipped in with FROST MOUNTAIN (ML 5-1) at PRX, CANTORA (ML 5-1) at MNR, and GALA GOLD (ML 5-1) at SWF, demonstrating the model's nose for mid-price overlays throughout the week.
Key Insights
Several clear patterns emerged from this week's data. First, the AI models showed exceptional cross-track consistency, finding winners at a diverse mix of venues — from Del Mar (DMR) and Laurel (LRC) to smaller circuits like Elkton (ELK), Keendam (KDX), and Indiana (IND) — suggesting the algorithms are not overfit to any single track profile. Second, the 4-1 and 5-1 morning-line bracket was a sweet spot all week: horses in that odds range appeared repeatedly in the notable wins lists across every model, pointing to a persistent edge in identifying horses the public slightly undervalues relative to their true probability. Third, the morning update models consistently matched or improved on their base-run counterparts — Algo 2's morning version jumped to 38.5% wins, and Weight Set 2's morning run added three additional wins — confirming that late-breaking data integration is meaningfully sharpening our predictions. Finally, the Experimental Weight Sets continue to prove their value: Weight Set 3's industry-leading 66.4% ITM rate and strong boxed exacta count suggest those weighting adjustments are working exactly as intended and deserve close attention going forward.
Looking Ahead
The Pick Pony AI is firing on all cylinders heading into the next week, with multiple models beating or matching the chalk baseline and the Ensemble approach continuing to set the pace — keep your eyes on Algo 2 and the morning update signals as your primary edge plays. With pending results still to settle and the experimental weight sets showing real promise, next week's recap could be even more exciting — stay locked in and let the AI do the heavy lifting! 🏇
Weekly Performance Summary
How each pick source's top selection finished across every graded race this week.
Pick Pony's AI Winner models rank every horse in each race using a weighted combination of speed figures, pace analysis, trainer/jockey statistics, track bias, and dozens of other handicapping factors. Each algorithm uses a different set of optimized weights to capture different angles on the same race data. Top Pick Wins measures how often the model's #1 ranked horse won the race outright. Places and Shows count 2nd and 3rd place finishes respectively. ITM (In The Money) is the combined total of Win + Place + Show. Rates are calculated against settled races only (pending results excluded). The Exotic Wager Hits table shows how often the model's top 3 picks finished in the correct order (Exacta/Trifecta) or in any order within the top positions (Boxed Exacta/Boxed Trifecta). Morning Update rows are the same picks re-ranked after the official scratches and surface changes were applied that morning.
Algo 1
Our weighted rating: every horse is scored on speed, pace, class, form and connections, and the top score is the pick.
Algo 1 Winner
The overnight top pick in every graded race, scored the night before from the official past performances.
Metric
Count
Rate
Total Races
846
Top Pick Wins (1st)
229
31.4%
Top Pick Places (2nd)
137
18.8%
Top Pick Shows (3rd)
96
13.2%
Top Pick ITM
462
63.4%
Pending
35
Exotic Wager Hits (Algo 1 Winner)
How often the source's top three finished 1-2-3, counted as straight and boxed exactas and trifectas.
Wager Type
Hits
Description
Exacta
55
Picks 1 & 2 finished 1st & 2nd in exact order
Boxed Exacta
43
Picks 1 & 2 both finished in top 2, reversed order
Trifecta
17
Picks 1, 2 & 3 finished 1st, 2nd & 3rd in exact order
Boxed Trifecta
65
Picks 1, 2 & 3 all finished in top 3, any order
Algo 1 Morning Update
The same picks re-ranked that morning after the official scratches and surface changes were applied.
Metric
Count
Rate
Total Races
846
Top Pick Wins (1st)
229
31.2%
Top Pick Places (2nd)
137
18.7%
Top Pick Shows (3rd)
99
13.5%
Top Pick ITM
465
63.4%
Pending
35
Exotic Wager Hits (Algo 1 Morning Update)
How often the source's top three finished 1-2-3, counted as straight and boxed exactas and trifectas.
Picks 1 & 2 both finished in top 2, reversed order
Trifecta
17
Picks 1, 2 & 3 finished 1st, 2nd & 3rd in exact order
Boxed Trifecta
66
Picks 1, 2 & 3 all finished in top 3, any order
AI Longshot Picks (Kelly Edge Plays)
Longshot picks are individual horses flagged at publication time: their morning-line odds were 8-1 or greater — prices the betting public considers unlikely to win — and the model's calibrated win probability said the price overpaid. Unlike the winner models above, a longshot is not a race's top-ranked horse, so these tables count selections, not races: every flagged horse is one bet, graded on its own finish, and a single race can contribute more than one. (For the same reason there is no exotic-hits table here — an exacta or trifecta needs a ranked 1-2-3 ticket, which a set of longshot flags does not define.) There is also one longshot table rather than one per algorithm: the flag depends only on the horse's price and its calibrated edge, never on where an algorithm ranked it, so all three models publish exactly the same longshot list. Because longshots win far less frequently by definition, win rates in this section are expected to be substantially lower than the winner models above. That is normal and by design — even a modest longshot win rate of 5-10% can be highly profitable because the payoffs are dramatically larger. A $2 win bet on an 8-1 longshot returns $18. Focus on ITM rate and on the realised ROI in the Longshot Betting Strategy Simulator below rather than on raw win percentage when evaluating longshot model performance.
The horses listed in each day's Kelly Edge Plays section: 10-1 or higher on the morning line and rated a value play by our models; counted per horse, not per race.
Metric
Count
Rate
Longshot Selections
419
Races Covered
231
Wins (1st)
17
5.0%
Places (2nd)
31
9.2%
Shows (3rd)
26
7.7%
ITM
74
21.9%
Pending
12
Longshot Betting Strategy Simulator
What a flat win bet on every flagged longshot would have returned this week at each minimum-edge threshold.
Each strategy below replays the past 60 days of AI longshot picks as actual $2 mutuel bets and reports the realised return. A longer window than the rest of this report is used here on purpose: longshots are rare and extremely high-variance, so a single week's results are dominated by luck — only over a larger sample do the calibrated edge strategies separate from simply betting every longshot. Bets are deduplicated by (track, race, horse) so a horse picked by all three algorithms generates one bet, not three. Edge strategies use a calibrated win probability (from the longshot meta-learner) and bet only when P(win) × decimal odds − 1 exceeds the listed threshold — these target overlays rather than long prices for their own sake. Hit% is the share of tickets that cashed any leg. Published longshots graded: 2825; wider 10/1-and-longer universe available to the edge rows: 14304, of which 10499 could be assigned a calibrated edge. Calibrator: out-of-sample tiered calibrator on the cross-race win probability (lgbm_win_prob_fwd), trained through 2026-07-30; edge rows replay only the 43 days after that date, so no graded race was ever seen by the calibrator. The edge filters use only leak-free forward scores (lgbm_win_prob_fwd — captured before each race ran), so the ROIs are a genuine out-of-sample estimate — never the in-sample re-scored values that previously inflated this section. The two families are measured differently — read them separately. The raw-filter rows grade the longshots we actually published, over the full 60 days. The edge rows sweep a different population: every thoroughbred runner priced at 10/1 or longer, published or not. That matters, because being published as a longshot already requires passing an edge test — so re-applying an edge threshold to the published picks would just re-filter an already-filtered group and tell you nothing. The edge rows also cover only the most recent 43 days (10499 runners), because everything before 2026-07-30 was needed to fit the calibrator and grading it would be marking its own homework. That window widens on its own as forward scores accumulate.
Strategy
Bets
Tickets Cashed
Stake
Payout
$2 WIN — all longshots
2825
155 (5.5%)
$5650.00
$3156.30
$2 PLACE — all longshots
2825
414 (14.7%)
$5650.00
$3901.60
$2 SHOW — all longshots
2825
677 (24.0%)
$5650.00
$3775.08
$2 WIN + $2 PLACE
2825
414 (14.7%)
$11300.00
$7057.90
$2 WIN — ML ≥ 15/1
1495
63 (4.2%)
$2990.00
$1569.64
$2 WIN — ML ≥ 20/1
911
27 (3.0%)
$1822.00
$727.36
$2 WIN — field ≥ 8
1832
91 (5.0%)
$3664.00
$2059.48
$2 WIN — edge ≥ 5%
5965
353 (5.9%)
$11930.00
$7818.78
$2 WIN — edge ≥ 10%
4806
289 (6.0%)
$9612.00
$6343.36
$2 WIN — edge ≥ 20%
4637
284 (6.1%)
$9274.00
$6192.94
$2 WIN — edge ≥ 30%
2947
213 (7.2%)
$5894.00
$3816.28
$2 WIN — edge ≥ 40%
1900
145 (7.6%)
$3800.00
$2500.30
$2 WIN — edge ≥ 50%
212
16 (7.5%)
$424.00
$235.70
$2 WIN — edge ≥ 60%
0
—
$0.00
$0.00
$2 WIN — edge ≥ 70%
0
—
$0.00
$0.00
$2 WIN — edge ≥ 80%
0
—
$0.00
$0.00
$2 WIN — edge ≥ 90%
0
—
$0.00
$0.00
$2 WIN — edge ≥ 100%
0
—
$0.00
$0.00
$2 WIN — edge ≥ 10% AND field ≥ 8
2717
151 (5.6%)
$5434.00
$3695.10
Trifecta Structure Plays (AI Exotics)
The Trifecta line from each day's AI Exotics section: key the top pick to win over the next three horses (six combinations per race); the tag shows whether the race shape looked playable.
Each AI Predictions post suggests one trifecta structure per race: the Algo 3 top pick to win, with its next three ranked horses boxed for 2nd and 3rd (6 combinations). Races our race-shape model rates most favorable are tagged ⭐ Best Bet in the post. This table grades last week's published plays against the actual trifecta payoffs at each track's real wager minimum — no assumptions, no hindsight.
Plays graded
427
Trifectas hit
58 (13.6%)
Total staked
$1,513.20
Total returned
$1,177.98
Profit / loss
−$335.22 (-22.2% ROI)
⭐ Best Bet races only
6 plays, 1 hit (16.7%), −$11.90 (-56.7% ROI)
397 race(s) skipped for scratches inside the top four; 22 skipped where the track's trifecta minimum wasn't recorded. Per-race wager minimums have been captured since July 4, 2026, so this ledger grows more complete each week.
Benchmark: The Morning-Line Favorite
What betting the morning-line favorite in every race would have returned: the baseline every pick source has to beat.
The simplest possible strategy is to always bet the morning-line favorite — the "chalk." This table shows how that baseline performed across the same thoroughbred races this week, so you can judge the AI's win rates against what the public's top choice delivered. Beating the chalk consistently is the real test of any handicapping model.
Metric
Count
Rate
Settled Races
595
Favorite Wins (1st)
223
37.5%
Favorite Places (2nd)
102
17.1%
Favorite Shows (3rd)
82
13.8%
Favorite ITM
407
68.4%
Consensus Expert Picks
Picks from more than twenty-five published handicappers, weighted into one consensus selection per race.
Consensus Expert Picks are derived from multiple independent handicapping sources. Each source submits their top picks per race, and the consensus ranking is built by scoring each horse: 4 points for a 1st-place pick, 3 for 2nd, 2 for 3rd, and 1 for 4th. The horse with the highest total score becomes the consensus top pick. Agreement % measures what share of experts included a horse in their picks — higher agreement signals stronger conviction across sources. The Win/Place/Show/ITM and Exotic stats below measure how the consensus top pick actually performed on race day.
Metric
Count
Rate
Total Races
683
Top Pick Wins (1st)
216
38.0%
Top Pick Places (2nd)
107
18.8%
Top Pick Shows (3rd)
86
15.1%
Top Pick ITM
409
72.0%
Pending
115
Exotic Wager Hits (Consensus Top 3 Picks)
How often the source's top three finished 1-2-3, counted as straight and boxed exactas and trifectas.
Wager Type
Hits
Description
Exacta
64
Picks 1 & 2 finished 1st & 2nd in exact order
Boxed Exacta
32
Picks 1 & 2 both finished in top 2, reversed order
Trifecta
14
Picks 1, 2 & 3 finished 1st, 2nd & 3rd in exact order
Boxed Trifecta
55
Picks 1, 2 & 3 all finished in top 3, any order
Performance by Breed
The same results split by thoroughbred and quarter horse races, which are handicapped differently.
Top-pick win rates vary dramatically by breed. Thoroughbred (TB) racing is the most form-driven and predictable, while Quarter Horse (QH) and Arabian (AR) sprints are decided in a few hundred yards and are far more chaotic — they naturally produce lower win rates. Because the summary above blends all breeds together, this table separates them so each breed's true performance is visible. W/P/S are wins, places (2nd), and shows (3rd), with the percentage of settled races in parentheses. ITM is the In-The-Money total (Win + Place + Show). EX/BEX/TRI/BTRI are exacta, boxed-exacta, trifecta, and boxed-trifecta hits on the top picks. Longshot rows count individual flagged selections rather than races (a race can contribute more than one), and show dashes for the exotic columns because those wagers need a ranked 1-2-3 ticket.
Thoroughbred
Source
Races
W
P
S
ITM
EX
BEX
TRI
BTRI
Algo 1 Winner
846
229 (31%)
137 (19%)
96 (13%)
462 (63%)
55
43
17
65
Algo 2 Winner
846
274 (37%)
158 (22%)
87 (12%)
519 (71%)
72
53
21
70
Algo 3 Winner
668
208 (36%)
104 (18%)
86 (15%)
398 (69%)
61
39
16
47
Experimental Weight Set 2 Winner
846
249 (35%)
114 (16%)
97 (14%)
460 (65%)
56
43
16
50
Experimental Weight Set 3 Winner
846
247 (34%)
133 (18%)
106 (15%)
486 (67%)
56
53
17
65
AI Longshot
419
17 (5%)
31 (9%)
26 (8%)
74 (22%)
—
—
—
—
Performance by Track
The same results split by track, to show where each source is strongest and weakest.
The tables below break down each pick source's performance at individual tracks. W/P/S show the number of wins, places (2nd), and shows (3rd) with the percentage of settled races in parentheses. ITM is the combined In The Money total (Win + Place + Show). EX (Exacta) counts races where the top 2 picks finished 1st and 2nd in exact order. BEX (Boxed Exacta) counts races where the top 2 picks both finished in the top 2 in either order. TRI (Trifecta) and BTRI (Boxed Trifecta) apply the same logic to the top 3 picks finishing in the top 3 positions. Longshot rows count individual flagged selections rather than races (a race can contribute more than one) and show dashes in the exotic columns, because those wagers need a ranked 1-2-3 ticket that a per-horse longshot flag does not define. Algo 1 is the weighted rating, Algo 2 the AI Ensemble and Algo 3 the AI Win Probability model — ordered by measured performance, so a higher number should be the stronger source; the Weight Set rows are experimental profiles graded as shadows. Comparing sources at a specific track helps identify which is strongest for that venue. The ML Favorite (chalk) row shows how the morning-line favorite performed in that track's thoroughbred races — the baseline any pick source needs to beat to be adding value at that venue.
