How Storm Ledger estimates damage

Calibration

Spatial pattern checked against emergency calls (P2000, NL). Loss level is a model estimate; with your claims data it becomes a calibrated figure. That is the pilot. Fire brigade calls catch only part of the damage: people call insurers, not 112, for a leaking roof or a flooded shed. Expected loss is therefore shown as a band for reporting rates of 2% to 20% of damaged buildings producing a call, around the central 10%.

Unit: street segment within a buurt (4119 segments, 107,923 buildings, 312 with a storm call). P(call) = 1 − exp(−ρ·Σ P_i) with ρ = 0.1 fixed (calls per damaged building; not identifiable from calls alone). AUC is pooled out-of-fold, leave-one-buurt-out. Ciara-only AUC 0.603; logistic benchmark (log Σ MDD_adj, log n) 0.639. P2000 coverage: see data/p2000/SOURCES.md. treePer was fitted to 0 but is kept at its 0.06 prior: Overpass was unavailable for many study buurten (58 mirror failures in this run), so most buurten outside Amsterdam had no tree data and the fit could not see a tree effect. P2000 windows: Ciara 9 Feb 2020 00:00–21:00 (1,378 geocoded incidents); 18 Jan 2018 10:00–18:00 (2,459), so the morning landfall in the west is missing and calls there are under-counted. Some regions are barely present in the archive (Noord-Holland Noord in 2020; the northern provinces in 2018); control buurten drawn there look undamaged when they were only unobserved, which pulls k down and the AUC with it. Half of the study buurten were chosen because they had calls (case-control sampling), which inflates the absolute call rate and so k; the AUC (ranking) is not affected. Treat absolute P(damage) as indicative.

Check against insured losses

The probability ranking is calibrated; the euro amounts are not. A rough outside check uses what Dutch insurers reported for the two calibration storms:

• Storm of 18 January 2018: first Verbond van Verzekeraars estimate about €90 m for homes and cars, later reported at about €436–450 m in total; together with the storm of 3 January insurers handled close to 300,000 claims (Hart van Nederland, NOS, ESB).
• Storm Ciara, 9 February 2020: first Verbond estimate about €150 m, mainly homes and cars (report).

That puts the average claim at roughly €1,500 (including cars). In the five demo areas the model's typical repair given that a building is damaged is €180–1,300 (median per area), and a building with many dwellings would usually produce several claims. Two separate questions follow.

How many buildings are damaged is set by k = 1147, fitted on fire-brigade calls with a fixed share ρ = 0.1 of damaged buildings producing a call. A few thousand P2000 calls against several hundred thousand claims suggests ρ is nearer 0.01, which would mean many more damaged buildings than shown, each with a smaller loss: with P = 1 − exp(−k·MDD_adj) the loss of a damaged building averages about 1/k of rebuild cost (0.09% now). That split is a modelling assumption, not data.

How much money in total is Σ MDD_adj × rebuild cost. It does not depend on k: it comes straight from the Schwierz et al. (2010) curve and the €/m³ rule, and has not yet been compared with the national insured totals above. The next calibration step is to run the model for the whole country for these two storms and scale the curve to the insured totals. Until then, read euro amounts as indicative and use P(damage) for ranking.

Equations

1. Gust at the building. g10 = inverse-distance-weighted (power 2) event peak of hourly max gust FX at the 4 nearest KNMI land stations. g(h) = g10 · (h/10)^α, α = 0.22 where built-up share within 100 m exceeds 0.15, else 0.14. h is the 70th-percentile roof height above ground (min 3 m).
2. Exposure E = clamp(0.6, 1.4, 1 + 0.25·log₂(clamp(h/h̄_up, 0.5, 2)) + 0.5·(fetch/200 − 0.5) + 0.2·corner + 0.05·detached). h̄_up is the mean roof height of buildings within 60 m upwind (±30°). Fetch is the upwind distance to the nearest building; water and parks hold no buildings, so they count as open.
3. Vulnerability V = roof type × face orientation (area-weighted: windward 0.85, edge 1.15, leeward 1.3, flat 1) × slope band × build-year band × floors band × (1 + 0.06·min(n_trees, 6)), where n_trees counts trees taller than their distance to the facade.
4. Mean damage degree MDD = min(1, a·max(0, g − g0)^b) with a = 1.3415e-13, b = 5.2874, g0 = 72 km/h, fitted to the Schwierz et al. (2010) curve (PAA × MDD). MDD_adj = min(1, MDD · E · V).
5. Damage probability P = 1 − exp(−k · MDD_adj). Given damage, the damage ratio is exponential with mean MDD_adj/P, which splits P into minor (< 1% of reconstruction cost), moderate (< 8%) and severe. Expected state: none if P < 5%, else severe if MDD_adj ≥ 0.03, moderate if ≥ 0.004, otherwise minor.
6. Value impact repair = MDD_adj × reconstruction cost, reconstruction = 3DBAG LoD 2.2 volume × €/m³ by class. Value decrease % = expected repair ÷ value × 100. P10 and P90 are quantiles of the same damage distribution. This is repair cost relative to value, not a market price effect.
7. Value from the CSV when given; otherwise CBS mean WOZ × dwellings in the buurt, shared by floor area (area view), or, in the portfolio, the buurt mean WOZ scaled by the addressor the buurt mean WOZ per address (portfolio, which carries its floor-area share of the building's repair cost)apos;s usable area relative to a 100 m² dwelling; the address carries its floor-area share of the building's repair cost.

Every parameter

Defined in model/params.ts; fitted values come from model/calibration.json, written by scripts/calibrate.ts.

Sources and licences

What is approximated, and what would replace it