Methodology

For US locations the model uses National Weather Service gridpoint forecasts — the same hourly data the NWS publishes for forecasters, resolved to roughly a 2.5 km grid. Active NWS watches, warnings and advisories are read separately and applied on top of the hourly scoring.

For Canadian locations, and as a fallback whenever the NWS is unavailable, the model uses Open-Meteo. Open-Meteo reports snowfall using a fixed snow-to-liquid ratio of about 7:1, which under-reports accumulation compared with NWS output, so its snowfall values are normalised upward before scoring to keep the two sources comparable.

Forecasts are cached briefly per location so that a storm affecting one area does not generate a separate upstream request for every visitor.

The prediction is a weighted multi-factor score, not a single snowfall threshold. Snowfall accumulation is the largest single input, but timing matters as much as total: snow falling between roughly 5 AM and 8 AM weighs more heavily than the same amount arriving mid-afternoon, because it lands directly on the morning bus run.

Ice and freezing rain are weighted more heavily than snow at equivalent amounts. A thin glaze closes districts that would run normally through several inches of snow, because it makes both roads and sidewalks unsafe at once.

Wind chill is scored against a regional threshold rather than a fixed one, alongside wind speed and visibility — blowing snow can close a district on a day with little new accumulation. Forecast confidence is also folded in, so a marginal event with high model disagreement produces a less confident number than a well-agreed one.

Closure thresholds are not national. Two inches in Atlanta closes schools that four inches in Minneapolis would not, so the model applies per-state and per-province multipliers derived from regional closure behaviour. Canadian and Prairie-province thresholds are meaningfully higher than the US average.

School type shifts the threshold again. Public districts decide system-wide and have to account for bus safety across their whole catchment. Private and charter schools decide independently and often close at slightly lower thresholds. Colleges and universities rarely close for snow at all and generally do not issue two-hour delays.

Where we know a district is rural, that lowers its effective threshold: rural districts run the longest routes over the least-treated roads and close on forecasts that leave nearby city districts open.

Scoring changes are checked against a backtest suite of real historical closure and non-closure events rather than by eye. The suite replays archived forecast data for each case through the current model and reports where the prediction and the real outcome disagree.

This matters most for the negative cases. It is easy to build a model that fires on every winter storm; the harder test is whether it stays low on the days that looked threatening and still had school.

A probability is not a promise, and there are specific cases where this model is weaker than others:

• We do not have access to any district's internal decision-making, staffing situation, or superintendent. The model infers from weather, not from announcements.
• Forecast quality degrades quickly beyond about 24 hours. A prediction made two days out is a much weaker signal than one made the evening before.
• Marginal temperature events — rain versus snow near freezing — are the least reliable case, because a one-degree forecast error changes the outcome entirely.
• Districts sometimes close for reasons unrelated to the forecast: a power outage, a water main break, or accumulated conditions from a previous storm the model no longer sees.

Treat the number as a well-informed estimate of how a typical district in your area would read the same forecast. Always confirm an actual closure with your school or district directly.