Pondview Pool Forecaster
SourcePredicted family arrivals per hour, from the day’s weather.
Hold the day fixed and change the weather, to see what the model thinks the weather is actually worth.
The shaded stretch is the temperature range the training season actually reached.
Working it out…
Ask about the crowd
Try one of these, or type your own question. The last one is deliberately out of scope — it shows you where the answers stop.
Answers are built by calling the same forecast endpoints this page uses, so every number traces to a model output rather than the assistant’s own arithmetic. It covers arrivals per hour — it cannot see the pool.
How this works
A gradient-boosting regressor predicts how many families arrive each hour, from the hour, the day type, and the day’s weather. It was trained on a season of pool sign-in sheets. The figures below are read live from the deployed model.
How it was validated
Leave-one-day-out cross-validation. Splitting by row would leak: hours in the same afternoon share weather and the same crowd, so a random split lets the model train on part of the day it is being scored on and reports a flattering number.
The baseline is reported next to the model every single time. Beating a sensible lookup table by a modest, honest margin is the result — and that margin narrowed as the season grew, because more days make the lookup table better too.
What it uses
Hour of day and weekend flag, plus the day's weather.
An ablation settles whether the weather earns its place: a weather-only model already beats the baseline, and a calendar-only model does not. The gains are largest exactly where a lookup table has nothing to say — cool hours (1.21 arrivals/hour better) and weekends (1.09 better).
What it can't do
It predicts arrivals, not occupancy — people leave whenever they like, and the sign-in sheets never recorded that. For how full the pool is right now, the live dashboard counts people, on Tuesdays, Wednesdays, Thursdays and Saturdays.
One pool, one season, and temperatures between 67 and 90°F — outside that range it is extrapolating and says so. The line through each bar is the model’s typical error for that hour, not a calibrated confidence interval. Past the 15-day weather horizon it falls back to a typical day rather than inventing one.
Weather by Open-Meteo (CC BY 4.0). Data, notebook, training code and evals are on GitHub.