Methodology
Snow Day Predictor Canada Methodology
How our multi-model AI forecast engine evaluates Canadian winter weather conditions and tunes the result to your region’s school board behaviour.
We publish what we measure, how we cross-check it, and where the data comes from. We do not publish the specific point weights inside the algorithm — those evolve as we benchmark against each winter’s real closure outcomes — but every conceptual ingredient is documented on this page.
What we measure
The five conditions our forecast weighs
Canadian school boards do not make the morning call from a single weather number. Neither do we. Our engine evaluates the same five overnight signals superintendents and transportation directors look at when they decide between 5:30 and 6:30 am.
For every prediction we slice the 8 pm to 8 am window into hourly samples and extract five independent measurements. Each one is evaluated on its own merits — a 25 cm storm and a −45 °C wind chill night are very different events even though both can close school. Treating them separately is what lets the predictor explain its own reasoning back to you instead of returning a black-box percentage.
1. Overnight snowfall accumulation
The dominant signal across nearly every Canadian region. We pull hourly snowfall totals at your exact coordinates between 8 pm and 8 am — not a daily average, not a wide-area forecast — so lake-effect bands off Georgian Bay, Alberta clippers crossing the Prairies, and Colorado lows tracking through Ontario all show up at their true local intensity. The same 15 cm storm reads very differently in Vancouver, Toronto, and Calgary; reading the actual hourly accumulation lets the regional layer (below) do its job.
2. Wind and blowing snow
Wind gusts only matter when there is snow available to blow around, so we combine maximum overnight gust speed with both new snowfall and existing snow on the ground. The interaction is what closes rural Ontario, Prairie, and Atlantic Canada bus routes during whiteout conditions — sometimes with only modest accumulation. Wind without snow does not contribute to the closure score.
3. Freezing rain and ice pellets
Pound for pound, the most disruptive form of Canadian winter weather. Just a couple of millimetres of freezing rain creates an ice glaze that closes more schools than 10 cm of dry snow — the December 2013 GTA ice storm and the periodic St. Lawrence freezing-rain corridors that close CSSDM, EMSB, and Lester B. Pearson are the canonical examples. We weigh freezing rain and ice pellets separately because they behave differently underfoot and on the road.
4. Wind chill and extreme cold
The primary closure trigger across Saskatchewan, Manitoba, Northern Ontario, and the territories. We use the overnight wind-chill minimum — the metric Canadian boards themselves use — and compare it against a regional safety threshold. The February 2019 polar vortex closures across the Prairies are the clearest recent example: no snowfall required, just dangerous cold.
5. Active Environment Canada warnings
Blizzard, Winter Storm, Snowfall, Freezing Rain, Wind, and Extreme Cold warnings issued by Environment and Climate Change Canada are corroborating evidence rather than the headline signal. Different warning categories carry different weight — a Blizzard Warning is a much stronger closure indicator than a Snowfall Warning, which boards in Alberta routinely operate through. We treat the warning as one input among five, not the final word.
The consensus engine
Why we cross-check three forecast models
A single weather model has bad nights. Three independent models, queried in parallel and reconciled into one consensus, do not.
Every prediction is run independently through three forecast models before we combine them. The first is an international ensemble that ingests the European ECMWF model, the US NOAA GFS model, and Germany’s DWD ICON model — the same multi-model backbone meteorologists rely on for medium-range forecasts. The second is the Canadian Meteorological Centre’s GEM model, the operational forecast that Environment and Climate Change Canada itself runs. The third is an independent hourly-resolution forecast we use as a tie-breaker and freshness check.
Each model produces its own factor breakdown and its own school-closure probability against the same regional profile. We then take a weighted average that gives the Canadian GEM model meaningful pull — it is the forecast Canadian boards themselves look at — without letting any single model dominate. The independent third-party forecast keeps us honest when the ensemble and GEM drift in the same direction for the same reason.
The agreement between models is itself a signal. When all three land within a few points of each other we report higher confidence. When they spread apart by 20 percentage points or more we report a split forecast and explicitly lower the confidence label, because the underlying atmospheric situation is genuinely uncertain. The user sees a small transparency panel showing each model’s individual closure percentage so you can judge the agreement for yourself rather than trusting a single averaged number.
The regional layer
Per-province school board tuning
The same storm closes Vancouver and is a normal Tuesday in Calgary. Any Canadian snow day predictor that ignores that fact is wrong half the country at a time.
Canadian closure behaviour is not a national average. Vancouver SD #39, Surrey SD #36, and Greater Victoria SD #61 routinely close at 8 to 10 cm of overnight snow because coastal British Columbia has minimal salting and plowing infrastructure. The Calgary Board of Education and Edmonton Public Schools operate through 25 cm and consider it a normal winter morning. The Toronto District School Board almost never closes buildings, while the Toronto Student Transportation Group cancels buses several times each winter. A useful predictor has to know all of this.
We maintain fourteen distinct regional profiles to capture those differences — separate profiles for coastal versus interior British Columbia, separate profiles for the Greater Toronto Area versus rural and Northern Ontario, and dedicated profiles for Quebec, each Atlantic province, the Prairies, and each of the three territories. Each profile encodes a typical closure threshold, a regional wind-chill threshold, and a tolerance factor that captures how readily local boards translate weather into closures.
When you enter a postal code, we resolve it to coordinates and then to the right regional profile — using longitude to split coastal from interior BC, and latitude to split GTA-style boards from rural and Northern Ontario. The same forecast inputs flowing through the Coastal BC profile and the Alberta profile produce very different probabilities, which is exactly the point. The regional layer is what makes the prediction reflect how your specific board actually behaves rather than a meaningless national mean.
Two outputs
School closure vs bus cancellation
In the Canadian system, buses and buildings are separate decisions. Our predictor returns both probabilities because both matter to families.
The Toronto Student Transportation Group cancels school buses on mornings when the TDSB keeps every one of its buildings open. The same pattern holds for the Ottawa Student Transportation Authority, Halton Student Transportation Services, Student Transportation of Peel Region (STOPR), and Student Transportation Services of Waterloo Region. Buses cancel for conditions that would not fully close a school — long rural routes, exposed concession roads, and the cumulative risk of transporting thousands of children through marginal weather.
For every prediction we return a school-closure probability and a separately calculated bus-cancellation probability. The bus probability is always at least as high as the closure probability and is capped at 97 percent. Families who walk to school care about the closure number; families who rely on buses care about the bus number; both numbers are the actual answer to the question parents ask in the morning.
The window
The 8 pm to 8 am forecast window
School boards do not decide based on what happens at noon. They decide based on what happens overnight.
We target the 8 pm to 8 am window because that is exactly the window Canadian school boards evaluate before making the 5:30 to 6:30 am call. Transportation directors check overnight accumulation, road conditions, and the morning forecast — not yesterday’s daytime weather and not tomorrow afternoon. By aligning the predictor to the same window the boards themselves use, the probability reflects what decision-makers will actually be looking at when they pick up the phone before sunrise.
Freshness
Update frequency
Every prediction runs against live forecast data at the moment you ask for it.
The predictor pulls live forecast data on every page load. We do not cache predictions between users and we do not store your queries. Underlying weather models refresh on an hourly cycle, so each time you click Predict the engine reflects the freshest available data — a 6 pm check and a 10 pm check on the same evening will routinely disagree as the overnight forecast stabilises.
Attribution
Data sources and attribution
Every input that flows into a prediction comes from a publicly documented source. We cite them all.
- Open-Meteo — international forecast ensemble
Combined ECMWF, NOAA GFS, and DWD ICON output, accessed at open-meteo.com. The medium-range backbone of our consensus.
- Canadian Meteorological Centre GEM model
The Canadian Meteorological Centre’s Global Environmental Multiscale forecast — the operational model Environment and Climate Change Canada runs. Queried via Open-Meteo’s GEM endpoint and weighted toward Canadian conditions.
- WeatherAPI.com — independent cross-check
A third-party hourly forecast service used as a tie-breaker. Available at weatherapi.com.
- Environment and Climate Change Canada
Official winter weather warning categories — Blizzard, Winter Storm, Snowfall, Freezing Rain, Wind, and Extreme Cold — as defined and issued by ECCC.
- Publicly documented school board policies
Closure, bus-cancellation, and extreme-cold criteria published by TDSB, CBE, EPSB, CSSDM, EMSB, HRCE, NLESD, Winnipeg SD, Vancouver SD #39, and every other major Canadian district. These are the source of the per-province thresholds.
- zippopotam.us — Canadian postal code geocoding
Forward Sortation Area centroids for every Canadian postal code, used to resolve your input into latitude and longitude before the forecast is fetched.
Honest disclosure
Limitations and what the predictor cannot do
A useful forecast is one that knows its own edges. Here is where ours stops.
- We are a forecast, not an official decision. The authoritative call is made by your school board between 5:30 and 6:30 am — TDSB, CBE, EPSB, CSSDM, HRCE, NLESD, and every other district run their own decision process. Always confirm with your local board’s app or website before you change your morning plans.
- Probabilities are capped at 95 percent for school closure and 97 percent for bus cancellation. No forecast deserves certainty — there are always scenarios where the storm under-performs, a board makes an unusual call, or warnings expire before the morning.
- Weather model accuracy degrades beyond 24 hours. We are most accurate when checked the evening before, ideally between 6 pm and 10 pm once the overnight forecast has stabilised. Predictions for two or three days out are useful as a watch signal, not a commitment.
- Regional profiles are smoothed approximations. They capture how the typical board in each region behaves; your specific board may close earlier or later than the regional norm. Catholic, French, and independent boards in the same geography sometimes split.
- Edge cases can exceed model resolution. A localised snow squall band off Lake Huron, a sharp-gradient ice storm tracking along the St. Lawrence, or a chinook arriving mid-storm in Calgary may all behave differently from what the gridded forecast suggests. We flag low confidence when the models disagree, but a 5 km wide snow band is hard for any model to nail.
Continuous improvement
How we improve over time
Methodology pages should age well. Ours does because we keep updating the underlying model.
We benchmark predictions against actual closure outcomes every winter and document where the engine got it right and where it missed. When a Canadian school board updates its closure policy — Vancouver SD #39 revising its snow threshold, a Quebec service centre changing how it announces, an Ontario consortium tightening rural bus criteria — we revise the corresponding regional profile. We add more cities and sub-regional profiles based on user requests sent to hello@snowdaypredictorcanada.com.
Last methodology update: May 2026. Corrections, policy updates, and requests for additional regional profiles are welcome.
Ready to see the methodology in action?
Run the predictor for your postal code and watch the five factors, three-model consensus, and regional tuning resolve into a single probability.
Run the predictor →