Forecast accuracy and test results

We do not have enough results to claim that the forecasts are accurate. The initial historical check holds out three days per park. That is useful for finding problems, but too small to establish performance across seasons or holidays.

What was tested?

Each test uses only completed days earlier than the day being estimated. The attraction basket is chosen from those earlier days too. We compare the predicted park-day queue index with the index calculated from the held-out day.

The index summarizes posted standby waits across a consistent group of rides. An error of ten minutes refers to that index, not to every ride or a visitor’s actual time in line.

Compare against a simple baseline

The baseline is the mean of earlier park-day indices. A more complicated model needs to do better than a simple alternative before it earns a stronger claim.

What has not been established?

The annual weekday, holiday, school-break and event rules have not been backtested against a full year of observations. The short-range historical check does not validate annual ratings. We have not measured a success percentage for the three crowd bands.

Results from forecasts made in advance

Model runs are saved before future dates occur. We need to match those saved predictions with later observations and accumulate enough results before publishing a prospective accuracy claim. No prospective accuracy figure is available yet.

Read the calculations or report a problem. Visitor reports are reviewed as context; they do not automatically become measured attendance or wait data.