Every August the same conversation happens. Who’s good this year? Who lost too much? Is that team as deep as it looks on paper?
Everybody answers it from memory and instinct. I wanted to see what happens when you answer it from the data.
So I built a model. It takes every California girl who ran the 2025 state meet and is back this fall, looks at every race she’s run at Woodward Park, the state course, plus the Clovis and Roughrider invitationals that are held there, adds her track results from the spring, and projects what she’ll run in November. Then it simulates the meet ten thousand times and scores it.
It’s called FRESNO — Forecast of Returning Entrants, Simulated Not Observed. Both halves of that are caveats, and they’re the two things worth knowing before you click the link at the bottom.
The first is what it can’t do. This model only knows athletes who ran the state meet last year. It has never heard of the incoming freshman who’s about to run 17:40, or the girl who missed her qualifying meet with a stress fracture and is now the best runner on her team. Transfers are in there when known. If you know of someone who’s changed schools and isn’t showing up on the right team, drop their name and the switch in the comments and I’ll add her. The clearest case is a team that graduated most of its scoring five. Whitney won Division 2 last year and returns three runners, so they don’t appear in these standings at all. A team needs five and they have only 3. Switch the scoring depth to 3 and they show up tenth. They’ll almost certainly reload, and when those athletes race this fall the model will see them. Right now it can’t.
The second is about the times. The fastest projection on the page is 17:13. Somebody will break 17:00 this fall. Somebody does every year, and the model isn't claiming otherwise. A projection is the middle of a range, and the middle is never the fastest thing that happens.
What this is attempting to do is paint the most accurate picture possible, given the information available right now, of how the state meet might go in November. As the season runs and more results go into the model, confidence in those projections should increase.
Does it work? I tested it the only honest way I know. I hid the 2025 season completely, built and tuned everything on the years before it, and then let the model predict 2025 exactly once. Ordering teams, it put the average team within 0.64 places of where it belonged, and 92% within one place.
I also checked it against a floor. If you assumed every returning athlete simply ran three percent faster than last year, how close would that get you? Nobody would actually forecast that way, but it’s the minimum a model has to beat to be worth building. FRESNO came out about 16% better.
That’s the whole test. One shot, on a season the model had never seen, and the results get published regardless of what they look like.
Every race an athlete runs at Woodward Park feeds her projection, so after Clovis and Roughrider this will know more than it does today. I’ll post updates as that happens.
The full methodology, including what I got wrong along the way, and there was plenty, is linked on the dashboard.
The boys’ model is next. Same approach, refit from scratch — boys and girls develop differently enough that nothing carries over. I’ll post it when it’s ready.