Two From Three at the Top: Scoring a Melbourne Cup Ratings Model

Nine dollars.

That's what Half Yours paid to win the Melbourne Cup, and it was the second-highest rating in a model I'd built the week before the race. The model's top pick, Buckaroo, finished twenty-fourth. Its third pick, Middle Earth, ran third at $26. Two of the top three finishers came from the model's top three selections. The model's number one pick ran dead last.

That's about as honest a scorecard as horse racing gives you.

The model

The approach was simple and deliberately narrow: profile the last twenty Melbourne Cup winners by a set of observable traits (sire origin, dam-sire origin, age, lead-up race, distance jump, days between runs) and weight each variable by how frequently it appeared in winning profiles. No speed figures, no sectionals, no form-cycle analysis. Just pattern frequency across a small but consistent sample.

The field was scored against those weights. Five runners stood out:

Buckaroo topped the model at 14.4 and was around $10 in the market. Half Yours scored strongly across every metric at $8. Middle Earth rated highly and looked possible overs at $41. Onesmoothoperator shaped as a quiet achiever at $31. Presage Nocturne carried a balanced profile at $8.50.

The model agreed with the market on Half Yours and Presage Nocturne. It disagreed most on Middle Earth and Onesmoothoperator, long-priced runners whose profiles matched past winners more closely than their odds suggested.

The result

Half Yours won for Tony and Calvin McEvoy under Jamie Melham, beating Goodie Two Shoes ($31) and Middle Earth ($26) in a result that rewarded the patient and punished the favourites. Presage Nocturne, the market's second-elect, ran nineteenth. Buckaroo, the model's top-rated runner, finished last of twenty-four.

Scoring the model: two of its top three selections finished in the first three placings. Its number-one pick was a complete miss. Onesmoothoperator ran sixteenth. Presage Nocturne was worse.

If you'd backed the model's top three each-way, you collected on two of three. If you'd backed its top pick to win, you watched it trail the field. That's the nature of a frequency-based model applied to a twenty-four-horse handicap on a wet track. It identifies profiles, not certainties.

What it learned

Three takeaways worth banking for next year.

First, the model's trait-matching works as a filter, not a selector. It correctly narrowed a twenty-four-horse field to a shortlist that contained the winner and the third placegetter. It couldn't rank within that shortlist. Buckaroo's profile scored highest but his race was worst.

Second, wet track form was the variable the model didn't weight and should have. The original preview flagged it as a watch factor but didn't build it into the scores. Half Yours and Middle Earth both handled the conditions; Buckaroo didn't. A rain-adjusted layer would have reshuffled the top three.

Third, the model has no opinion on the race itself: barriers, tempo, rider tactics, weight-for-age dynamics. A profile model tells you which horse fits the template. It doesn't tell you which horse fits the race. That gap between fitting the history and fitting the day is where the model's ceiling lives.

The Cup remains the race that stops the nation and reliably donates my money to the bookmakers. This year, at least, the numbers gave some of it back.