Forecasts are in training.
Here's the test.
This page will show exactly how good our wait forecasts are — in minutes, per park, updated continuously. It's empty on purpose: we don't publish a forecast until it beats an honest baseline, and we don't publish accuracy claims without the receipts.
The claim we intend to make
"Our next-hour numbers beat trusting the sign." Posted wait times drift — signs freeze, lag, and pad. When our forecasts ship, this page will show, per park, how far off we were versus how far off you'd have been assuming the sign was right.
The test it has to pass first
- Beat the baseline. The forecast must predict next-hour waits at least 10% better (median absolute error) than the dumbest credible strategy: "assume the wait stays what it is now."
- On held-out data. Tested walk-forward — trained only on the past, scored only on days it never saw. No do-overs.
- Receipts kept. Every forecast we ever issue is archived as-issued, before the answer is known. Nothing gets revised after the fact.
Our recorder went live July 19, 2026. The evaluation needs roughly 30 days of verified data, so the first pass/fail call lands around mid-August. The full methodology, in plain language →
The misses policy
When the scoreboard goes live it includes a section called "where we were wrong this week", auto-populated, never omitted. If we have a bad week, you'll read it here first. A scoreboard that only shows the wins is an ad.