Weekly wait-time predictions for 5 California theme parks, 30-minute resolution.
Stale input - forecast onlyServing forecastAug 30 to Sep 06, 2026Modelprophet_fleetRun20260822T203750ZTrained2026-08-22 20:39:07 UTCPublished2026-08-30 21:49 UTC
7.0 min
MAE
11.5 min
RMSE
76.6 %
Within 10 min
13.6 %
Severe misses
1.2 min
Bias
10.3 min
High-wait MAE
7.8 min
Peak-hours MAE
Model comparison
Backtest on the held-out final 20 percent of history (78,110 observations). The serving model must beat the per-ride mean baseline.
Model
MAE
RMSE
Within 10 min %
Severe miss %
Bias
High-wait MAE
Peak MAE
prophet_fleet (serving)
7.0
11.5
76.6
13.6
1.2
10.3
7.8
xgb_global
8.3
12.9
71.8
17.6
3.8
12.1
9.4
xgb_local_fleet
8.5
14.7
71.9
19.3
3.7
12.9
10.0
baseline_ride_mean
9.5
14.5
64.5
23.5
1.0
14.4
9.1
Predicted vs actual, daily mean
Serving model backtest: mean wait across all rides per day. Hover for values.
ActualPredicted
Data table
Date
Actual (min)
Predicted (min)
Apr 15
9.3
12.3
Apr 16
8.5
13.0
Apr 17
13.6
15.0
Apr 18
19.1
18.8
Apr 19
14.1
15.1
Apr 20
11.0
14.6
Apr 21
12.4
13.1
Apr 22
12.2
12.6
Apr 23
12.5
13.0
Apr 24
15.4
15.0
Apr 25
20.0
18.8
Apr 26
14.5
15.1
Apr 27
11.4
14.6
Apr 28
10.2
13.1
Apr 29
11.6
12.6
Apr 30
12.4
12.9
May 01
15.3
15.0
May 02
20.2
18.9
May 03
15.0
15.2
May 04
10.6
14.6
May 05
13.3
13.2
May 06
11.0
12.7
May 07
11.9
13.0
May 08
12.5
15.1
May 09
17.6
18.9
May 10
11.9
15.2
May 11
14.3
16.0
Backtest error distributions by model
Predicted minus actual on the held-out test set, minutes. Blue under-forecasts,
red over-forecasts. Shared axes, so a wider spread means a genuinely worse model. The
baseline predicts each ride's historical average wait, so the difference between its
panel and the others is what the models actually learned.
prophet_fleet (serving)
MAE = 7.01 min | RMSE = 11.49
xgb_global
MAE = 8.26 min | RMSE = 12.88
xgb_local_fleet
MAE = 8.45 min | RMSE = 14.70
baseline_ride_mean (the floor every model must beat)
MAE = 9.51 min | RMSE = 14.46
Mean wait by local hour
The evening peak is the part a timezone mistake silently deletes.
ActualPredicted
Hardest rides to predict
Highest backtest MAE for the serving model. Long queues move in bursts.