Our accuracy, in public
Every 5 minutes we forecast the next hour. Then the radar tells us whether we were right. Here is the score — updated weekly, never retro-edited, misses included.
Loading the latest scores…
What POD, FAR and CSI mean
Three standard scores from weather verification, in plain language. For each forecast we ask a yes/no question — "will it rain at this spot, at or above this intensity, at this lead time?" — and compare the answer with what the radar later observed.
| Score | Plain language | Perfect |
|---|---|---|
| POD probability of detection | Of the times it actually rained, how often did we warn you? POD 0.8 = we caught 8 rain events in 10. | 1.00 |
| FAR false alarm ratio | Of the times we said "rain", how often did it stay dry? FAR 0.3 = 3 warnings in 10 were false alarms. | 0.00 |
| CSI critical success index | The combined score: hits divided by (hits + misses + false alarms). Punishes both crying wolf and sleeping through rain. | 1.00 |
There is a real trade-off here: any app can hit POD 1.0 by warning constantly — its FAR would be terrible. Publishing all three together is what keeps us honest.
Methodology
Forecast vs. later radar. Every 5 minutes the engine issues a 60-minute point forecast from live Met Éireann radar (motion-tracked extrapolation). We store each forecast, wait, and compare it against what the radar actually showed at that spot 5 to 30 minutes later. A hit means rain was forecast and rain was observed at or above the threshold; a miss and a false alarm are exactly what they sound like.
- Lead times: 5, 10, 15, 20 and 30 minutes ahead.
- Thresholds: 0.3 mm/h ("any rain" — a drizzle you'd notice) and 2.5 mm/h ("real rain" — you want to be indoors).
- Spatial window: a ~3 km neighbourhood around the point, so a shower drifting 1 km sideways counts fairly for both the forecast and the observation.
- Window: scores aggregate a rolling 7-day window and are republished weekly. Past weeks are never restated.
Methodology v1 — August 2026. Changes will be announced and versioned here, not silently applied.
Honest limitations
- Radar-vs-radar. We verify against the same radar we forecast from, not rain gauges on the ground. Shared radar biases cancel out of the score, but what reaches your jacket can differ from what the beam sees.
- Drizzle below the beam. Far from the radar the beam sits high; very light drizzle underneath it can go unseen — by the forecast and by the verification alike.
- Rain that hasn't formed yet. Extrapolation moves existing rain; it cannot foresee a shower that initiates within the hour. This is the hardest case in nowcasting and it shows up as lower POD at 30 minutes. We publish that number anyway.
- Coverage edge. Scores cover the Dublin radar's ~240 km range. Spots near the edge get fewer, noisier samples.
- Radar interference. External emitters can paint narrow spokes of false echo for a scan. A filter now removes them before the forecast (after a false-alarm cluster on 23 September 2026); a new kind of interference could still slip through for a frame — how it works has the story.
Verified accuracy: Ireland. Preview areas (Europe, United States) get their own scores before they launch — see coverage.