Lightning Outlook

What the numbers are

Forecast is total cloud-to-ground strokes expected in that area over the period. Normal is what that area averages for the same calendar window. vs normal is the ratio between them, which is the part the model is better at than the raw count.

How much to trust it

Verified against Alaska Lightning Detection Network observations for 2013–2019, scored separately within each Predictive Service Area — so knowing that the Interior gets more lightning than the North Slope earns no credit. Roughly 280 forecasts per area per period.

Both figures below are medians across the 17 areas, and the spread behind them is wide: at one day, lightning-vs-none runs from 79% in the weakest area to 96% in the strongest, and more-vs-less from 58% to 75%. That makes the median a fair summary of the product overall and a poor guide to any one area — the figure for the area you care about is in its tooltip on the map.

• Lightning vs none answers “is anything coming at all”. Take one period in that area that saw lightning (above 25 strokes a day) and one that saw effectively none; this is how often the outlook put the higher number on the one that had it. Guessing scores 50%.
• Rank correlation answers the harder question, “how much”. It is the Spearman correlation between the forecast anomaly and the observed one across all periods: 1 would mean every period ranked in exactly the right order, 0 no relationship at all. Values near 0.5 are typical for medium-range forecasts of convection.

The two columns sit on different scales — one starts at 50%, the other at 0 — so they are not comparable to each other, only across periods and areas within a column.

Longer periods hold up better than their final day alone does, because summing across days averages out errors in exactly when storms arrive: day nine on its own scores about 70% on lightning-vs-none, while the nine-day total scores 80%.

Two different things are measured and they do not move together. Finding active periods is whether the outlook knows something is coming; ranking severity is whether it knows how far above normal. Across areas these are nearly unrelated (rank correlation 0.08). Koyukuk and Upper Kobuk finds active periods 96% of the time but ranks severity only moderately — a clean on-or-off signal without much gradation. The table column reports severity, because that is what the “vs normal” beside it claims; the tooltip gives both.

Reliability is not uniform across the state. It is highest in the Interior boreal areas, where lightning is common enough to verify against, and lowest on the North Slope, the Kenai and the Seward Peninsula, where storms are rare and the record thin. The reliability shown beside each area is that area’s own, and areas without enough observed activity to score are marked not verified rather than given a number.

If you treat it as a yes/no flag

The figures above are threshold-free: they describe the ordering, not a decision. If instead you act whenever the 1-day forecast for your area exceeds some level, the trade-off between missing storms and chasing false alarms is yours to set, and it is steep:

Statewide, 2013–2019, against days that saw more than 25 strokes. Missing a storm and standing up for one that never arrives carry different costs, and only you can weigh them — which is why this is shown as two numbers rather than combined into one.

What good, moderate and limited mean

The word beside each area summarises its rank correlation.

Hovering an area gives its own figure rather than the band range.

Known limits

• The model was fitted on June and July. May and August are outside that and run weaker — May in particular ranks the size of an anomaly less reliably, though it still separates active areas from quiet ones.
• It says nothing about ignition or holdover. Strokes are not starts; fuels, cure state and rainfall on the same day decide that.
• A ratio needs a normal worth dividing by. Areas averaging under two strokes a day for the date show none — the North Slope in May, for instance. Everywhere else gets one, including the quiet end: a forecast well under normal is a real signal and is shown as such.
• This is guidance from a statistical model, not an official forecast, and not a warning product. It does not replace NWS fire weather products or the Alaska Interagency Coordination Center.

Where it comes from

GEFS ensemble forecasts drive a gradient-boosted lightning parameterization trained on ERA5 and ALDN observations. Each member is run separately and the members averaged afterwards. The result is divided by a climatology built the same way from twenty years of GEFS reforecasts at the matching calendar window and forecast lead, so the model’s own biases largely cancel, then rescaled to observed stroke counts.

Method: Hostler, J., and coauthors: Seasonal predictability of June lightning in Alaska from the March–April PNA: evidence from a 67-year reconstructed record. Journal of Climate. (under review)