A weather forecast is a chain of five links: observe, transmit, model, correct and decide. The first three are done well by the big agencies and global models, more and more with artificial intelligence. The fourth, correcting for the exact place where decisions are made, is the one that's usually missing, and it's where a forecast's usefulness is won or lost.
The grid problem
Global models resolve the atmosphere in cells several kilometres wide. An operation, on the other hand, is a point: a pit, a substation, a work front. Between the cell and the point there can be a valley, a slope or a ravine, and in the mountains that changes everything. A model can be right that there will be storms in the region and wrong about whether lightning will strike within 16 km of that point, which is what matters for safety.
Rain isn't lightning
The first temptation is to use the chance of rain as a storm warning. It doesn't work: many afternoons it rains without a single strike nearby. The second temptation is to use convective available potential energy (CAPE), the classic storm indicator. In the Andes that fails too: at a high-altitude site in northern Peru, in hours with lightning within 16 km, median CAPE was 65 J/kg, and only 2 of 156 hours exceeded the 1,000 J/kg that classic thresholds look for. A warning based on that threshold would have missed almost every storm.
Correcting with the site's history
The technique is half a century old and has an unglamorous name: Model Output Statistics (MOS). You take the variables global models deliver in the previous day's run and train a correction against what actually happened at the site, in this case lightning observed from space. The output isn't "it will rain" but a calibrated probability of lightning within 16 km in the afternoon.
How well it works
We tested it the right way: one day ahead, and always on months the model hadn't seen. The correction caught 9 in 10 afternoons with lightning within 16 km of the site, and 3 in 4 warnings were right. In the dry season, when storms are rare and catch people off guard, it anticipated 9 of 15 afternoons.
It helps that mountain storms keep time like a clock. Over ninety days of wet season, activity started around noon, peaked between two and four in the afternoon, and was gone by six. That makes the risk something you can plan around: a blast, a lift or an outdoor shift can be moved by a few hours, a day in advance.
A forecast isn't an alarm
There's a line worth not crossing. A calibrated forecast is a planning tool; the safety alarm is triggered by observed lightning, not by the forecast. In Pulse, the forecast can bring an alert level forward, but never lower it or shorten the all-clear time set by IEC 62793. That way, a wrong forecast never leaves anyone unprotected.
Limits
These results come from one site and a sample that keeps growing every season. The correction has to be trained for each place, because the whole point is that it's local. And in El Niño years, when history stops being a guide, it's worth watching whether the calibration still holds.
References
Glahn, H. R. and Lowry, D. A. (1972). The use of Model Output Statistics (MOS) in objective weather forecasting. Journal of Applied Meteorology 11, 1203–1211. · Wilks, D. S. (2019). Statistical Methods in the Atmospheric Sciences, 4th ed. Elsevier. · IEC 62793:2020. Thunderstorm warning systems.