US racing picks — Thu 5 Nov 2026

The Turkish tote (TJK) takes bets on US racing; each runner's share of that win pool is the market's opinion, published before the off. The model starts from that share and adjusts it with sixteen form layers (trainer, jockey, sire, damsire, owner, last finish, form, days since, head-to-head, track record, draw, rating, weight, age, depth of history). Edge is the model probability minus the pool-implied probability, in points. A positive edge means the model likes the horse more than the crowd does.

Weights were fitted by conditional logit on 7,572 US races and tested walk-forward: each block scored with weights fitted only on earlier races, 6,972 races out of sample. On its own the model does not beat the pool (-6.0 pts top-3); combined with it, top-3 coverage moves +0.4 pts and likelihood improves. Verdict: passed (marginal). The gain is calibration, not coverage: the record page shows exactly how it performs.

Layer weights: piyasa +0.73 · antrenor +0.27 · aygir +0.24 · jokey +0.23 · form3 -0.21 · damsire +0.10 · hp +0.05 · adres -0.03 · son_bitis +0.03 · n_gecmis +0.02 · sahip +0.02 · siklet +0.02 · h2h +0.01 · kgs +0.00 · draw +0.00 · yas +0.00 · son_fark +0.00

A race is usable when at least 70% of runners have US form on file and the field averages two or more runs; thin at 40%; otherwise blind, where no verdict is given and the order shown is the market's. Before the pool opens (TJK opens it on race morning) every race is blind.