Methodology

Technical Methodology

This page provides technical details about the data sources, models, and calibration methods behind snowpredictor.ca’s probability outputs. For a non-technical explanation, see our How It Works page.

Weather Data Pipeline

Primary Data Source

Weather forecast data is retrieved from the Open-Meteo API (open-meteo.com), which integrates multiple numerical weather prediction (NWP) model outputs. For Canadian locations, our primary model is Environment Canada’s Global Deterministic Prediction System (GDPS), which runs at approximately 10 km horizontal resolution and updates twice daily (00Z and 12Z runs). Secondary verification uses the NOAA Global Forecast System (GFS) model.

Variables Retrieved

For each location query, we retrieve the following forecast variables for the 9 PM to 7 AM decision window:

  • Hourly snowfall rate (mm water equivalent)
  • Hourly precipitation type (snow / rain / freezing rain / ice pellets)
  • 2-metre temperature (°C)
  • Apparent temperature / wind chill (°C) — calculated using Environment Canada’s wind chill index formula
  • 10-metre wind speed (km/h)
  • Snow depth — current existing snowpack (cm)
  • Visibility (km) — where available

Probability Calibration

Base Scoring Model

Raw weather variables are converted to a 0–100 point score using piecewise linear functions derived from historical analysis of observed Canadian school board closure decisions. Each function maps a weather variable value to a point contribution representing its marginal contribution to closure probability, controlling for the other variables.

Location Calibration

Raw scores are calibrated against location-specific historical baselines using a logistic regression model trained on five years of Canadian school board closure records. Calibration parameters include: historical closure rate at a given raw score for that board; school board region type (urban/suburban/rural); board geographic position relative to lake-effect snow source regions; and province-level policy context (extreme cold protocols, minimum temperature thresholds, etc.).

Probability Output

The calibrated logit output is transformed to a 0–0.98 probability using a standard sigmoid function capped at 0.98 to reflect irreducible forecast uncertainty. Outputs are expressed as percentages for user-facing display.

Accuracy Measurement

Prediction accuracy is assessed using a held-out test set of observed school closure events from the 2023–24 and 2024–25 winter seasons. Accuracy is measured using the Brier Score (mean squared error of probability forecasts) and the Area Under the ROC Curve (AUC) across binary closure/open outcomes, stratified by event type and region. See our How It Works page for published accuracy rates by event category.

Limitations and Known Biases

  • Lake-effect snow underestimation: NWP models at 10 km resolution systematically underestimate peak lake-effect snowfall rates within tight snow bands. Our calibration partially corrects for this in affected regions but cannot fully compensate for model resolution limitations.
  • Rare event underrepresentation: Our training data contains relatively few extreme events (>40 cm accumulations, wind chill below -45°C). Predictions in these ranges may be less well-calibrated than for more common events.
  • Board policy drift: School board closure policies can evolve over time as administrations change or policies are updated. Our calibration reflects historical behaviour and may lag actual current policy.