How accurate is it, really?
We believe in showing our working. Here is the model's scorecard in full, tested against the Met Office's own observations of where snow has actually fallen across Britain (1991–2024, 37,523 locations), the flattering numbers and the honest ones.
The headline number
Across 34 years, the model called the right outcome (white Christmas or not) for 90.4% of UK place-Christmases. That sounds excellent, and it is true, but it needs one honest caveat, which is why this page exists.
Our model
90.37%
correct, across every UK location, over 34 years
A model that always says “no snow”
90.36%
scores almost exactly the same, because snow is rare
A white Christmas happens at only ~10% of UK locations in a given year, so simply betting “no” is right ~90% of the time. Plain accuracy mostly measures how rare snow is, not how clever the model is. So we judge it on the metrics that actually reward skill, below.
The metrics that show real skill
| Measure | Score | What it means |
|---|---|---|
| Plain accuracy | 90.4% | Looks great, but a model that just says "no snow" every time scores 90.36%, see below. |
| Discrimination (ROC-AUC) | 0.72 | Given a snowy place-year and a non-snowy one, it ranks them in the right order ~72% of the time. |
| Ranking skill (Spearman ρ) | 0.73 | Its ordering of which places are snowiest matches the real record closely (~75% pairwise). |
| Balanced accuracy | 50–66% | When snow and no-snow are weighted equally, the rare-event difficulty shows (best 65.7%). |
| Day-level forecast skill | ≈ 0 (far out) | Months ahead, no one can call snow on a specific day, and we do not pretend to. |
The honest takeaway: the model is genuinely good at where it tends to snow (a coin-flip would score 50% on ranking; it scores ~75%), and deliberately humble about which exact day snow falls, because that is close to unpredictable months ahead.
It is well calibrated
Group every location into ten bands by the chance we predict, then look at how much snow each band actually got. It climbs every single step, higher predicted chance really does mean more real snow.
| Predicted chance (band avg) | Observed snow days (Dec avg) |
|---|---|
| 1.1% | 0.71 |
| 3.8% | 1.07 |
| 5.7% | 1.2 |
| 6.7% | 1.42 |
| 9% | 1.44 |
| 9.4% | 1.56 |
| 11.6% | 1.67 |
| 13.1% | 1.9 |
| 16.3% | 2.62 |
| 25.6% | 4.07 |
Region-by-region scorecard
How closely the model's ranking matches the independent record, by region (Spearman ρ; 1.0 is perfect). It is strongest where snow has real structure, the north and the hills, and weakest in flat, marginal regions where a few hundred metres or a single cold snap decides everything.
| Region | ρ | Locations |
|---|---|---|
| Scotland | 0.78 | 5,539 |
| Wales | 0.75 | 3,223 |
| North West England | 0.73 | 2,992 |
| North East England | 0.71 | 1,384 |
| London | 0.66 | 3,075 |
| Yorkshire & the Humber | 0.62 | 2,514 |
| Northern Ireland | 0.58 | 644 |
| South East England | 0.48 | 2,973 |
| West Midlands | 0.43 | 5,474 |
| South West England | 0.42 | 3,939 |
| East of England | 0.26 | 3,255 |
| East Midlands | 0.15 | 2,511 |
What it can and can't do
- Can: tell you which places are reliably snowier, and give a calibrated long-run chance for your exact spot.
- Can't (yet): tell you months ahead whether snow will fall on the 25th specifically, long-range daily skill is near zero for everyone, us included.
- Sharpens up: inside about a month, the live forecast folds in and the number tightens as Christmas nears.
Validated against the Met Office HadUK-Grid 1 km observational record (1991–2024) across 37,523 locations, using an independent snow-lying field. For entertainment and general interest, not an official Met Office forecast.