Estimating Our Tail Fire Risk

In many disasters sheltering in place can make a lot of sense, but what about fire? As someone who lives in one of the densest municipalities in the country, with 3-story wooden buildings packed close together, is fire likely enough to rule out any kind of "stay put" approach? After having Claude do some rough calculations I think probably not.

Very few houses are lost to fire anymore, but this is mostly downstream from putting a lot of effort into suppression. From when a fire is reported until firefighters are on the scene is typically an impressive three minutes, and they can very quickly be delvering massive amounts of water onto a fire anywhere in the city. If an emergency were bad enough to break or overwhelm this system, a fire anywhere could lead to city-wide distruction.

Still, I really didn't have a good sense of scale here. If we had a month-long outage, are we talking 10%, 1%, 0.1% chance of fire? Historically, I would have tried to do a BOTEC, but these days I set Claude on the problem. I started with Opus 5.5, but after it didn't seem to be doing a great job I spent $29.93 to put Fable 5.1 on the problem (prompts). I don't know if Claude would beat an expert here, but after skimming what it did I think it's well above my historical level of modeling.

It estimated from historical fires, and thought our monthly risk was ~1% (but could easily be 5x higher or lower), and if our block worked together in a disaster we might reduce risk by 4x. It also thought that my dad's house in Medford (less dense) was ~4x safer, as was my sister's house in Brattleboro VT (even less dense). If you want to play with it for your house, feel free to point an LLM at the code.

For three New England houses, the modeled chance of burning during 30 days without fire suppression is about 1.2% in dense Somerville, 0.30% in Medford and 0.24% in Brattleboro if nobody defends the house. Preparing one house cuts that by between a fifth and two fifths, and a whole block preparing cuts it four to eight times. The estimates are rough: for Somerville, changing one assumption at a time moves the figure between about 0.2% and 6.5%. Results Chance that the house burns in a 30-day period with no hydrants or fire department, using annual-average weather and the central ignition rate. House Do nothing Attentive, no prep Attentive and prepared Whole block attentive Whole block prepared Somerville MA (dense two- and three-family streets) 1.2% 1.1% 0.80% 0.87% 0.32% Medford MA (less dense) 0.30% 0.25% 0.17% 0.16% 0.04% Brattleboro VT (small town) 0.24% 0.23% 0.19% 0.11% 0.03% Attentive means someone is home, fighting embers and small fires with whatever is at hand. Prepared adds stored water, several extinguishers and cleared combustibles. Whole block means every building within 100 m is at the same level. That is about 55 houses around the Somerville house, 35 in Medford and 20 in Brattleboro, not counting garages and sheds: roughly 17, 11 and 7 houses per hectare. The Somerville houses also hold two or three households each, so the difference in people is larger than the difference in buildings. The Somerville whole-block figure assumes defenders can hold their houses as a wind-driven fire arrives. If block defense works no better than single-house defense on red-flag days, it is 0.51%. Risk scales close to linearly with the length of the outage: the Somerville do-nothing figure for 90 days is 3.4%. Somerville and Medford share the same weather, and both are seasonal: about half of each house's risk comes from red-flag days, and dry gales at Boston Logan fall almost entirely between November and April. Brattleboro is not seasonal, because its nearest airport record has no such days. 30 days Somerville, do nothing Somerville, attentive Somerville, prepared Medford, do nothing March weather 3.0% 2.7% 2.1% 0.65% December–April weather 2.1% 1.9% 1.5% 0.48% June–September weather 0.31% 0.25% 0.17% 0.13% Red-flag days matter for both Massachusetts houses, but more in Somerville. On a red-flag day a fire can reach the Somerville house from a median of about 7,400 buildings and the Medford house from about 1,900, because the streets around the Medford house are less tightly packed. Red-flag days are 48% of the do-nothing risk for each; windy-and-dry days add another 36% in Somerville and 12% in Medford. In Brattleboro nearly all the risk is fire in the house's own cluster of buildings on ordinary days. The Keene record has no red-flag day in 20 years, and if one came, a fire could reach the house from about 150 buildings. How the model works The model treats every building as a node and asks, for one house, which other buildings' fires would reach it. Buildings. Footprints come from OpenStreetMap, within 5 km of the Somerville and Medford houses and 2.5 km of the Brattleboro house. Short-range spread. The chance that a burning building ignites a neighbor falls off with the gap between them. From a typical 110 m² house on a red-flag day it is 95% at half a metre, 74% at 3 m, 64% at 4.3 m, 43% at 8 m, 20% at 15 m and 4% at 30 m. These are per burning neighbor: with a burning house 4.3 m away on each side the chance is 87%. A source with twice the footprint reaches about 20% further, and garages and sheds reach less far. Large, commercial and institutional buildings are treated as 35% as ignitable. Less wind, shorter reach. Without strong wind the curve is cut off beyond about 15 m, the limit of radiant heat from one house. On an ordinary day the chance is 61% at 4.3 m, 8% at 15 m and 2% at 20 m. Embers. A burning building can also ignite buildings up to 250 m away within 35° of straight downwind, with probability falling exponentially with distance. Weather. Each day is one of four types, defined on hourly airport observations from 2005–2024. Wind direction is drawn from that type's observed wind rose. Travel time. On red-flag and windy-and-dry days a fire can only reach the house from as far as it could travel during a ten-hour wind event: 480 m/h downwind on red-flag days, 180 m/h on windy-and-dry days, a quarter of that sideways and 30 m/h against the wind. Ignitions. Fires that grow to involve a whole building start at a rate of 6 per million per day for each 100 m² of footprint, with a range of 2 to 20. Defense. The three levels scale the chance of fire reaching a defended house and of its own fires escaping. Attentive: ember ignition ×0.45, spread from 8 m or more ×0.8, spread from under 3 m ×0.95, own fires ×0.55. Prepared: ×0.12, ×0.45, ×0.85 and ×0.25. Forest fire. Buildings within 30 m of forest can also be ignited by a wildland fire. This is included for Brattleboro and adds nothing measurable there; see below. Simulation. For each day type the model samples the set of buildings whose fire would reach the house and adds up their footprint, W. The house survives a day with probability (1 − q·a·own) × exp(−q·W), where q is the ignition rate and a the house's own footprint. Days are independent. Day type Definition Share of days, Boston Logan Share of days, Keene NH Distance multiplier Share of the curve beyond 15 m kept Ember strength, length scale Calm/damp Rain that day or the day before 43.0% 37.6% 0.8 none none Ordinary Everything else 54.4% 62.2% 0.9 10% 0.002, 40 m Windy and dry No rain, humidity ≤ 50%, wind 32–40 km/h 2.3% 0.2% 1.0 50% 0.010, 60 m Red-flag No rain, humidity ≤ 50%, wind ≥ 40 km/h 0.32% none in 20 years 1.0 all 0.030, 90 m "Wind" is the fourth-highest hourly sustained wind of the day, so at least four hours at that speed. Somerville and Medford use Boston Logan; Brattleboro uses Keene, the nearest long airport record. Calibration and checks One part of the model is fitted to data, the short-range spread curve. The rest is checked against historical fires where that was possible. Short-range spread, fitted to Los Angeles 2025. The data are LA County building footprints joined to CAL FIRE damage inspections for the Eaton and Palisades fires: about 45,000 buildings, of which about 22,000 inspected residences are the targets. Each residence's chance of burning is modeled as a background term, set by how much burned 50–250 m away, combined with a per-neighbor spread probability for each destroyed building within 40 m. The data show no downwind stretch and no effect of building height. Lost, or merely damaged. Where little else nearby burned, houses within 1 m of a destroyed building were lost 95% of the time. At 2–3 m, 51% were lost but 76% were damaged; at 6–8 m, 40% and 58%. Many exposed houses caught fire or were scorched and then survived, most plausibly because someone put the fire out. With nobody to do that, damage is the better guide to what would be lost, so the model uses the fit to any damage. Its 50% distance is 5.2 m (90% interval 4.9–5.6 m from a spatial block bootstrap) for a 120 m² source, and it falls off roughly exponentially. An earlier version of this model used buildings lost (50% at 4.3 m) and a curve shape that could not rise above 75% even for touching buildings. A physics cross-check, and taller houses. Treat a fully burning house as a flat radiator at 84 kW/m², the figure behind building-code separation tables, and ask where the heat reaching the neighbor exceeds what ignites it. For a one-storey Los Angeles house facing stucco, the 50% distance is 3.7 m, close to the 3.5 m fitted for houses lost. For heat enough to ignite wood or vinyl siding, or to break windows, it is 6.0 m, close to the 5.2–6.2 m fitted for any damage. For a 2.5-storey New England house with wood or vinyl siding next door it is 7.5–9.5 m, that is 1.25 to 1.6 times the Los Angeles figure. The model stretches the fitted curve by 1.3. That is a judgment: the physics supports it and the Los Angeles data, with little variation in height, do not show it. The same calculation puts ignition at 4 m near 95%, well above the 64% used. Real fires are less regular than a flat radiator: walls face different ways, fences and trees intervene, and some houses burn down before they radiate much. The fitted curve keeps that irregularity. The calculation also shows that radiant heat from one house cannot ignite another beyond about 15 m, so spread at that range in Los Angeles was wind-driven, and the model removes it on days without strong wind. Does the fit measure spread, or only correlation? Fires were simulated on the real LA footprints with known spread parameters, then fitted the same way. Across five parameter sets and three background levels, the recovered curve was within about 0.05 of the true probability at 2, 5 and 10 m. This test was run on the earlier curve shape. Chelsea, Massachusetts, 1908. A gale reported at 40–45 mph burned about 2,800 buildings over 492 acres. Nearly everything between the railroad and the waterfront burned, and one building was lost north of the tracks. Run on today's Chelsea footprints with red-flag settings, a fire front arriving at the upwind edge burns 90% of the downwind district on average, and a single ignition inside it burns 100 or more buildings in 82% of runs. The model also burns 13% of the buildings just north of the railroad, where in 1908 firefighters held the line. Without the adjustment for taller houses the front burns 85%. Today's Chelsea is a stand-in: it was rebuilt after the fire, and 1908 roofs were largely wood shingle. Chelsea 1973 sets the red-flag threshold. On October 14, 1973, a fire burned 18 blocks despite a very large mutual-aid response. Boston Logan recorded sustained wind of 35–50 km/h from the west and humidity of 23–28% that afternoon. The red-flag class starts just below that day's wind. Spread rates. The 1908 fire reached its far end, 2.4 km away, in five hours: 480 m/h. In Kobe in 1995 fires spread at 20–30 m/h in light wind, and about 20 m/h faster for each extra metre per second of wind, which gives roughly 180 m/h on a windy-and-dry day. These set the travel-time limit. Kobe 1995 as a calm-day check. Kobe had light wind (2.6 m/s), humidity near 70% and no hydrant water. Of 150 surveyed fires, 49% stayed in one building, 29% burned 1,000 m² of site area or more, and 4.7% burned 33,000 m² or more. In dense Chelsea streets the model's calm days give 16%, 69% and 3%, and its ordinary days 13%, 76% and 26%. So the model now spreads fire further than Kobe saw, most of all on ordinary days. Kobe's brigades put out many fires early and others stopped at wide streets and concrete buildings, so some gap is expected, but this is the least settled part of the model. Around the three houses themselves, ordinary-day fires stay small, with a median of 3 to 8 buildings able to reach each house. Defense levels against bushfire evidence. At Mount Macedon in 1983, residents who stayed had about a 90% chance of saving the house, against about 44% losses among those who left. On Black Saturday in 2009, 77% of defended houses survived against 44% of unattended ones. Those imply defended houses were lost 20–40% as often. The model's "prepared" level gives 60–80%, weaker because most exposure here is heat from a burning neighbor a few metres away. No data were found on defending against that. Forest fire in Brattleboro. Vermont averages about 80 wildland fires and 157 acres burned a year, out of about 4.5 million forested acres. The model assumes 30 times as much would burn with nobody putting fires out, in line with 1903, when over 5,000 acres burned before the state had fire wardens. When a patch of forest burns, buildings within 30 m of it ignite with a probability of up to 20%. National land-cover data show forest over 54% of the area within 2.5 km of the Brattleboro house, but the house is a few hundred metres from the nearest patch, and none of the buildings that can carry fire to it stand at the forest edge. The forest term adds less than 0.001 percentage points. A house right at the forest edge would gain about 0.002 points, and assuming a hundred times today's burning in place of thirty would still leave that below 0.01. In this model, the buildings next door matter far more than the woods. What could not be fitted. An attempt to fit ember spotting on the LA data failed: the burn pattern there looks like a regional ember storm, not building-to-building spotting. How much of the wind-driven part of the curve applies on windy-and-dry days is also a judgment, and those days are 36% of the Somerville figure. Two Massachusetts fires that beat a full fire-department response (Lawrence 2008 and Lynn 1981) happened on days the model classes as ordinary. What moves the answer Each row below changes one assumption and leaves the rest alone. 30 days, do nothing Somerville Medford Brattleboro Central estimate 1.2% 0.30% 0.24% Ignition rate 2 to 20 per million per day per 100 m² 0.39% – 3.7% 0.10% – 0.98% 0.08% – 0.79% Adjustment for taller houses, none to 1.6× 0.47% – 1.9% 0.12% – 0.58% 0.16% – 0.41% Windy-and-dry days keep 25% to 75% of the curve beyond 15 m 0.96% – 1.4% 0.28% – 0.35% 0.25% No travel-time limit on wind-driven days 4.6% 0.52% 0.25% Firefighters save exposed houses as often as in Los Angeles 0.17% 0.05% 0.10% Which windy days count as dangerous. Gale-force wind occurs on 3.5% of days at Logan, but 64% of those days have rain. The model counts a gale day as red-flag only if there was no rain that day or the day before and humidity was 50% or lower. That rule is a judgment call, and it matters: Gale days counted as red-flag Share of days Somerville, do nothing All of them 3.5% 6.5% No rain that day 1.3% 2.8% No rain that day or the day before 0.63% 1.7% Also humidity ≤ 50% (the rule used) 0.32% 1.2% Also humidity ≤ 35% 0.16% 0.89% Rain mostly stops embers from lighting roofs, debris and fences; it does little to stop a burning house from igniting one three metres away. Half of the red-flag days are below freezing, and snow cover, which the model ignores, would reduce spotting further. How the estimate got here. The Somerville figure has moved a long way as assumptions were replaced with data. Version Somerville, do nothing All parameters guessed 5.7% Spread fitted to buildings lost in the LA fires (50% at 4.3 m, not 5–16 m) 1.6% Day-type frequencies measured at Logan (red-flag 0.32% of days, not 5%) 0.18% Footprint window widened from 2.5 km to 5 km 0.21% Spread refitted to any damage, with a curve that reaches 95% for touching buildings, and stretched 1.3× for taller houses 4.6% Travel-time limit on wind-driven days 1.2% What is fitted, what is judgment, and known gaps Status Parameters Fitted or measured Short-range spread curve for one-storey stucco houses in extreme wind (LA 2025); day-type frequencies and wind roses (airport observations); forest cover (national land-cover data) Checked against history, not fitted Spread on red-flag days (Chelsea 1908); the red-flag wind threshold (Chelsea 1973); downwind spread rates (Chelsea 1908, Kobe 1995) Judgment The 1.3× adjustment for taller wood-sided houses; how much of the curve beyond 15 m applies on days with less wind; ember strength; the rain and humidity rule; lower ignitability of large and commercial buildings; the defense multipliers; how much a disaster raises ignitions; the forest-fire settings other than Vermont's burn rate The ignition rate. US fire departments attend about 358,000 home structure fires a year, roughly 7 per million housing units per day. With a fire department responding, 24% spread beyond the room of origin and 38% are small confined fires, so between 24% and 62% might grow to involve the building with no response. At about two housing units per 100 m² of footprint, as in Somerville, that gives 4 to 9 per million per day, which brackets the central value of 6. On single-family streets the same reasoning gives 1 to 3, so the Medford and Brattleboro figures may be two to three times high on this count. No measured figure was found for how much a long outage raises ignitions through candles, generators and improvised heating, which pushes the other way. Known gaps. On ordinary days the model spreads fire further than the Kobe survey found, and nothing settles how much of that gap is Kobe's firefighting. The travel-time limit applies only on wind-driven days. On other days a reach set is everything the fire would eventually burn. No snow cover and no drought term. No building materials beyond the few that OpenStreetMap tags. Ignitions scale with footprint, not with the number of dwellings. Keene's weather stands in for Brattleboro's. The forest-fire term is crude: one burn rate for the whole state, and a guessed chance that a burning forest edge ignites a building. The spread curve comes from one-storey stucco houses in Los Angeles in extreme wind; its transfer to New England wood-frame streets rests on the radiant-heat calculation and the Chelsea comparison above. Code and sources The code is at github.com/jeffkaufman/fire-risk-without-suppression. It runs for any address given coordinates and a nearby airport weather station, and includes the scripts for the fit and for each check above. The three houses' locations are not in it. Data: Building footprints: OpenStreetMap via the Overpass API. Los Angeles calibration: LA County 2023 building footprints joined to CAL FIRE damage inspections for the Eaton and Palisades fires (ArcGIS feature service), and CAL FIRE's statewide inspection points (ArcGIS feature service). Weather: hourly airport observations from the Iowa Environmental Mesonet, 2005–2024, plus single days in 1973, 1981 and 2008. Forest cover: USGS National Land Cover Database 2021, from the MRLC map service. Historical fires, read in full: Chelsea 1908: Pratt, The Burning of Chelsea (1908); Fire Engineering's account; Wikipedia. Kobe 1995: Hokugo, The Performance of Fire Protection of Buildings against the Fires following the Great Hanshin-Awaji Earthquake, Fire Safety Science 5:947–958 (1997). Figures taken from abstracts or summaries, not the full papers: Mount Macedon 1983: Wilson and Ferguson, Fight or flee? A case study of the Mount Macedon bushfire (1984). Black Saturday 2009: Whittaker and colleagues, Household preparedness and responses to the Black Saturday bushfires. Home fire counts and spread beyond the room of origin: NFPA, Home Structure Fires. Vermont wildland fires: Vermont Department of Forests, Parks and Recreation, Fire Statistics for 2024; the 1903 figure is from the department's 2025 legislative testimony. Lawrence 2008: Firehouse. Lynn 1981: the fire of November 28, 1981.

For three New England houses, the modeled chance of burning during 30 days without fire suppression is about 1.2% in dense Somerville, 0.30% in Medford and 0.24% in Brattleboro if nobody defends the house. Preparing one house cuts that by between a fifth and two fifths, and a whole block preparing cuts it four to eight times. The estimates are rough: for Somerville, changing one assumption at a time moves the figure between about 0.2% and 6.5%.

Chance that the house burns in a 30-day period with no hydrants or fire department, using annual-average weather and the central ignition rate.

Attentive means someone is home, fighting embers and small fires with whatever is at hand. Prepared adds stored water, several extinguishers and cleared combustibles.

Whole block means every building within 100 m is at the same level. That is about 55 houses around the Somerville house, 35 in Medford and 20 in Brattleboro, not counting garages and sheds: roughly 17, 11 and 7 houses per hectare. The Somerville houses also hold two or three households each, so the difference in people is larger than the difference in buildings.

The Somerville whole-block figure assumes defenders can hold their houses as a wind-driven fire arrives. If block defense works no better than single-house defense on red-flag days, it is 0.51%.

Risk scales close to linearly with the length of the outage: the Somerville do-nothing figure for 90 days is 3.4%.

Somerville and Medford share the same weather, and both are seasonal: about half of each house's risk comes from red-flag days, and dry gales at Boston Logan fall almost entirely between November and April. Brattleboro is not seasonal, because its nearest airport record has no such days.

Red-flag days matter for both Massachusetts houses, but more in Somerville. On a red-flag day a fire can reach the Somerville house from a median of about 7,400 buildings and the Medford house from about 1,900, because the streets around the Medford house are less tightly packed. Red-flag days are 48% of the do-nothing risk for each; windy-and-dry days add another 36% in Somerville and 12% in Medford. In Brattleboro nearly all the risk is fire in the house's own cluster of buildings on ordinary days. The Keene record has no red-flag day in 20 years, and if one came, a fire could reach the house from about 150 buildings.

The model treats every building as a node and asks, for one house, which other buildings' fires would reach it.

Buildings. Footprints come from OpenStreetMap, within 5 km of the Somerville and Medford houses and 2.5 km of the Brattleboro house.

Short-range spread. The chance that a burning building ignites a neighbor falls off with the gap between them. From a typical 110 m² house on a red-flag day it is 95% at half a metre, 74% at 3 m, 64% at 4.3 m, 43% at 8 m, 20% at 15 m and 4% at 30 m. These are per burning neighbor: with a burning house 4.3 m away on each side the chance is 87%. A source with twice the footprint reaches about 20% further, and garages and sheds reach less far. Large, commercial and institutional buildings are treated as 35% as ignitable.

Less wind, shorter reach. Without strong wind the curve is cut off beyond about 15 m, the limit of radiant heat from one house. On an ordinary day the chance is 61% at 4.3 m, 8% at 15 m and 2% at 20 m.

Embers. A burning building can also ignite buildings up to 250 m away within 35° of straight downwind, with probability falling exponentially with distance.

Weather. Each day is one of four types, defined on hourly airport observations from 2005–2024. Wind direction is drawn from that type's observed wind rose.

Travel time. On red-flag and windy-and-dry days a fire can only reach the house from as far as it could travel during a ten-hour wind event: 480 m/h downwind on red-flag days, 180 m/h on windy-and-dry days, a quarter of that sideways and 30 m/h against the wind.

Ignitions. Fires that grow to involve a whole building start at a rate of 6 per million per day for each 100 m² of footprint, with a range of 2 to 20.

Defense. The three levels scale the chance of fire reaching a defended house and of its own fires escaping. Attentive: ember ignition ×0.45, spread from 8 m or more ×0.8, spread from under 3 m ×0.95, own fires ×0.55. Prepared: ×0.12, ×0.45, ×0.85 and ×0.25.

Forest fire. Buildings within 30 m of forest can also be ignited by a wildland fire. This is included for Brattleboro and adds nothing measurable there; see below.

Simulation. For each day type the model samples the set of buildings whose fire would reach the house and adds up their footprint, W. The house survives a day with probability (1 − q·a·own) × exp(−q·W), where q is the ignition rate and a the house's own footprint. Days are independent.

"Wind" is the fourth-highest hourly sustained wind of the day, so at least four hours at that speed. Somerville and Medford use Boston Logan; Brattleboro uses Keene, the nearest long airport record.

One part of the model is fitted to data, the short-range spread curve. The rest is checked against historical fires where that was possible.

Short-range spread, fitted to Los Angeles 2025. The data are LA County building footprints joined to CAL FIRE damage inspections for the Eaton and Palisades fires: about 45,000 buildings, of which about 22,000 inspected residences are the targets. Each residence's chance of burning is modeled as a background term, set by how much burned 50–250 m away, combined with a per-neighbor spread probability for each destroyed building within 40 m. The data show no downwind stretch and no effect of building height.

Lost, or merely damaged. Where little else nearby burned, houses within 1 m of a destroyed building were lost 95% of the time. At 2–3 m, 51% were lost but 76% were damaged; at 6–8 m, 40% and 58%. Many exposed houses caught fire or were scorched and then survived, most plausibly because someone put the fire out. With nobody to do that, damage is the better guide to what would be lost, so the model uses the fit to any damage. Its 50% distance is 5.2 m (90% interval 4.9–5.6 m from a spatial block bootstrap) for a 120 m² source, and it falls off roughly exponentially. An earlier version of this model used buildings lost (50% at 4.3 m) and a curve shape that could not rise above 75% even for touching buildings.

A physics cross-check, and taller houses. Treat a fully burning house as a flat radiator at 84 kW/m², the figure behind building-code separation tables, and ask where the heat reaching the neighbor exceeds what ignites it. For a one-storey Los Angeles house facing stucco, the 50% distance is 3.7 m, close to the 3.5 m fitted for houses lost. For heat enough to ignite wood or vinyl siding, or to break windows, it is 6.0 m, close to the 5.2–6.2 m fitted for any damage. For a 2.5-storey New England house with wood or vinyl siding next door it is 7.5–9.5 m, that is 1.25 to 1.6 times the Los Angeles figure. The model stretches the fitted curve by 1.3. That is a judgment: the physics supports it and the Los Angeles data, with little variation in height, do not show it.

The same calculation puts ignition at 4 m near 95%, well above the 64% used. Real fires are less regular than a flat radiator: walls face different ways, fences and trees intervene, and some houses burn down before they radiate much. The fitted curve keeps that irregularity. The calculation also shows that radiant heat from one house cannot ignite another beyond about 15 m, so spread at that range in Los Angeles was wind-driven, and the model removes it on days without strong wind.

Does the fit measure spread, or only correlation? Fires were simulated on the real LA footprints with known spread parameters, then fitted the same way. Across five parameter sets and three background levels, the recovered curve was within about 0.05 of the true probability at 2, 5 and 10 m. This test was run on the earlier curve shape.

Chelsea, Massachusetts, 1908. A gale reported at 40–45 mph burned about 2,800 buildings over 492 acres. Nearly everything between the railroad and the waterfront burned, and one building was lost north of the tracks. Run on today's Chelsea footprints with red-flag settings, a fire front arriving at the upwind edge burns 90% of the downwind district on average, and a single ignition inside it burns 100 or more buildings in 82% of runs. The model also burns 13% of the buildings just north of the railroad, where in 1908 firefighters held the line. Without the adjustment for taller houses the front burns 85%. Today's Chelsea is a stand-in: it was rebuilt after the fire, and 1908 roofs were largely wood shingle.

Chelsea 1973 sets the red-flag threshold. On October 14, 1973, a fire burned 18 blocks despite a very large mutual-aid response. Boston Logan recorded sustained wind of 35–50 km/h from the west and humidity of 23–28% that afternoon. The red-flag class starts just below that day's wind.

Spread rates. The 1908 fire reached its far end, 2.4 km away, in five hours: 480 m/h. In Kobe in 1995 fires spread at 20–30 m/h in light wind, and about 20 m/h faster for each extra metre per second of wind, which gives roughly 180 m/h on a windy-and-dry day. These set the travel-time limit.

Kobe 1995 as a calm-day check. Kobe had light wind (2.6 m/s), humidity near 70% and no hydrant water. Of 150 surveyed fires, 49% stayed in one building, 29% burned 1,000 m² of site area or more, and 4.7% burned 33,000 m² or more. In dense Chelsea streets the model's calm days give 16%, 69% and 3%, and its ordinary days 13%, 76% and 26%. So the model now spreads fire further than Kobe saw, most of all on ordinary days. Kobe's brigades put out many fires early and others stopped at wide streets and concrete buildings, so some gap is expected, but this is the least settled part of the model. Around the three houses themselves, ordinary-day fires stay small, with a median of 3 to 8 buildings able to reach each house.

Defense levels against bushfire evidence. At Mount Macedon in 1983, residents who stayed had about a 90% chance of saving the house, against about 44% losses among those who left. On Black Saturday in 2009, 77% of defended houses survived against 44% of unattended ones. Those imply defended houses were lost 20–40% as often. The model's "prepared" level gives 60–80%, weaker because most exposure here is heat from a burning neighbor a few metres away. No data were found on defending against that.

Forest fire in Brattleboro. Vermont averages about 80 wildland fires and 157 acres burned a year, out of about 4.5 million forested acres. The model assumes 30 times as much would burn with nobody putting fires out, in line with 1903, when over 5,000 acres burned before the state had fire wardens. When a patch of forest burns, buildings within 30 m of it ignite with a probability of up to 20%. National land-cover data show forest over 54% of the area within 2.5 km of the Brattleboro house, but the house is a few hundred metres from the nearest patch, and none of the buildings that can carry fire to it stand at the forest edge. The forest term adds less than 0.001 percentage points. A house right at the forest edge would gain about 0.002 points, and assuming a hundred times today's burning in place of thirty would still leave that below 0.01. In this model, the buildings next door matter far more than the woods.

What could not be fitted. An attempt to fit ember spotting on the LA data failed: the burn pattern there looks like a regional ember storm, not building-to-building spotting. How much of the wind-driven part of the curve applies on windy-and-dry days is also a judgment, and those days are 36% of the Somerville figure. Two Massachusetts fires that beat a full fire-department response (Lawrence 2008 and Lynn 1981) happened on days the model classes as ordinary.

Each row below changes one assumption and leaves the rest alone.

Adjustment for taller houses, none to 1.6×

Which windy days count as dangerous. Gale-force wind occurs on 3.5% of days at Logan, but 64% of those days have rain. The model counts a gale day as red-flag only if there was no rain that day or the day before and humidity was 50% or lower. That rule is a judgment call, and it matters:

Rain mostly stops embers from lighting roofs, debris and fences; it does little to stop a burning house from igniting one three metres away. Half of the red-flag days are below freezing, and snow cover, which the model ignores, would reduce spotting further.

How the estimate got here. The Somerville figure has moved a long way as assumptions were replaced with data.

Spread fitted to buildings lost in the LA fires (50% at 4.3 m, not 5–16 m)

Day-type frequencies measured at Logan (red-flag 0.32% of days, not 5%)

Footprint window widened from 2.5 km to 5 km

Spread refitted to any damage, with a curve that reaches 95% for touching buildings, and stretched 1.3× for taller houses

The 1.3× adjustment for taller wood-sided houses; how much of the curve beyond 15 m applies on days with less wind; ember strength; the rain and humidity rule; lower ignitability of large and commercial buildings; the defense multipliers; how much a disaster raises ignitions; the forest-fire settings other than Vermont's burn rate

The ignition rate. US fire departments attend about 358,000 home structure fires a year, roughly 7 per million housing units per day. With a fire department responding, 24% spread beyond the room of origin and 38% are small confined fires, so between 24% and 62% might grow to involve the building with no response. At about two housing units per 100 m² of footprint, as in Somerville, that gives 4 to 9 per million per day, which brackets the central value of 6. On single-family streets the same reasoning gives 1 to 3, so the Medford and Brattleboro figures may be two to three times high on this count. No measured figure was found for how much a long outage raises ignitions through candles, generators and improvised heating, which pushes the other way.

On ordinary days the model spreads fire further than the Kobe survey found, and nothing settles how much of that gap is Kobe's firefighting.

The travel-time limit applies only on wind-driven days. On other days a reach set is everything the fire would eventually burn.

No snow cover and no drought term.

No building materials beyond the few that OpenStreetMap tags.

Ignitions scale with footprint, not with the number of dwellings.

Keene's weather stands in for Brattleboro's.

The forest-fire term is crude: one burn rate for the whole state, and a guessed chance that a burning forest edge ignites a building.

The spread curve comes from one-storey stucco houses in Los Angeles in extreme wind; its transfer to New England wood-frame streets rests on the radiant-heat calculation and the Chelsea comparison above.

The code is at github.com/jeffkaufman/fire-risk-without-suppression. It runs for any address given coordinates and a nearby airport weather station, and includes the scripts for the fit and for each check above. The three houses' locations are not in it.

Building footprints: OpenStreetMap via the Overpass API.

Los Angeles calibration: LA County 2023 building footprints joined to CAL FIRE damage inspections for the Eaton and Palisades fires (ArcGIS feature service), and CAL FIRE's statewide inspection points (ArcGIS feature service).

Weather: hourly airport observations from the Iowa Environmental Mesonet, 2005–2024, plus single days in 1973, 1981 and 2008.

Forest cover: USGS National Land Cover Database 2021, from the MRLC map service.

Chelsea 1908: Pratt, The Burning of Chelsea (1908); Fire Engineering's account; Wikipedia.

Kobe 1995: Hokugo, The Performance of Fire Protection of Buildings against the Fires following the Great Hanshin-Awaji Earthquake, Fire Safety Science 5:947–958 (1997).

Mount Macedon 1983: Wilson and Ferguson, Fight or flee? A case study of the Mount Macedon bushfire (1984).

Black Saturday 2009: Whittaker and colleagues, Household preparedness and responses to the Black Saturday bushfires.

Home fire counts and spread beyond the room of origin: NFPA, Home Structure Fires.

Vermont wildland fires: Vermont Department of Forests, Parks and Recreation, Fire Statistics for 2024; the 1903 figure is from the department's 2025 legislative testimony.

Lawrence 2008: Firehouse. Lynn 1981: the fire of November 28, 1981.

How high is 1% per month? It means in ~6 years it becomes more likely than not. But even a few months of a disaster bad enough to prevent fire suppression would mean widespread starvation, since most people keep very little food on hand (and the density is far higher than the land can support). There are also a lot of very serious disasters where we still have fire hydrants: they're a pretty high priority for emergency response, and living in a low part of the city our water is gravity-fed from Quabbin.

Overall, I don't think I should be factoring fires heavily into thinking about what kinds of precautions and plans it makes sense for us to take. This is surprising to me, given how flammable and close the houses are, but mostly comes down to disasters that are bad enough to lose fire suppression are not very survivable.