Planetary-scale answers, unlocked.
A Hands-On Guide for Working with Large-Scale Spatial Data. Learn more.
Authors
Probable maximum loss (PML) is the largest loss an insurer can reasonably expect from a property or a portfolio. For one building, PML is an engineering estimate of the worst likely event, such as a fire that defeats part of the fire protection. For a portfolio exposed to hurricanes or earthquakes, PML is the loss at a chosen probability, such as the 1-in-100 or 1-in-250 year loss, read from an exceedance probability (EP) curve. Underwriters, reinsurers, lenders, and regulators use PML to decide how much risk to accept and how much capital to hold.
The term has never had one fixed meaning. In a 1969 paper in the Proceedings of the Casualty Actuarial Society, John S. McGuinness called PML "one of the least clear concepts in all insurance." His four-year study collected PML definitions from over one hundred underwriters, and no two fully agreed.
Two definitions dominate today.
Single-site PML. IRMI defines PML as a property loss control term for the maximum loss expected at a location in the event of a fire, expressed in dollars or as a percentage of total values. Marsh describes a PML study as a financial estimate of the largest physical loss that can reasonably be expected from a single event, including time-related and indirect costs.
Portfolio PML. In catastrophe modeling, PML is a point on the exceedance probability curve: the loss at a stated return period for a peril and a portfolio. A reinsurer quoting a "1-in-250 hurricane PML" means the annual loss with a 0.4% chance of being exceeded.
Property engineers often place PML on a ladder of loss estimates for one site:
Underwriters use the single-site PML to set how much of a risk to write, and engineers revisit it when a site changes. Lenders on US commercial property ask for seismic loss estimates before closing a loan. The current ASTM E2026 guide for seismic risk assessment of buildings frames that work as probable loss (PL) and scenario loss (SL) assessments.
Portfolio PML answers a different question: how bad can a year get across thousands of locations? Primary insurers use it to size reinsurance programs, reinsurers to price layers, rating agencies to test capital, and catastrophe bond investors to judge the chance of losing principal.
An exceedance probability curve plots each loss amount against the probability that annual loss exceeds it. A catastrophe model builds the curve by running a stochastic event set of 10,000 or more simulated years against a portfolio and ranking the yearly results.
The return period is the inverse of the annual exceedance probability:
A return period describes probability. A 1-in-100 year loss can occur two years running, and over 30 years a 1% annual chance compounds to about a 26% chance of at least one exceedance.
The curve comes in two versions, and a PML figure needs to say which one it uses.
At any return period, the AEP loss is greater than or equal to the OEP loss, and the gap widens in active years with several large events.
NOAA's National Centers for Environmental Information (NCEI) publishes a list of US billion-dollar weather and climate disasters, with CPI-adjusted costs. We ranked the 45 years from 1980 to 2024 by their tropical cyclone losses in two ways, once by each year's costliest storm (OEP) and once by each year's total (AEP), and assigned each rank an empirical return period of 46 divided by the rank.
In 2017, Harvey ($160.0B), Maria ($115.2B), and Irma ($64.0B) made the costliest hurricane year on the list, while Katrina alone made 2005 the year with the costliest single storm. The mean across all 45 years, the empirical AAL, is $34.3B. Seventeen of the 45 years had no billion-dollar tropical cyclone.
The example illustrates the shape of an EP curve. Forty-five years is too short a record to estimate a 1-in-100 or 1-in-250 year loss, and NCEI's costs are economic losses from billion-dollar events, which differ from insured losses. Catastrophe modelers fill the tail with simulated years for that reason.
Average annual loss is the mean loss per year across all simulated years. PML comes from the tail of the same distribution. A portfolio with a modest AAL can carry a large 1-in-250 PML if its locations cluster in one hurricane path, and two portfolios with the same AAL can need different reinsurance.
Insurers watch both. AAL sets the technical premium. PML sets the capital and reinsurance the portfolio needs to survive a bad year.
A catastrophe model produces portfolio PML in four steps:
Steps 1 and 2 are spatial. Each location must be matched to the hazard footprints that touch it, which for a large portfolio and a large catalog means billions of location-event pairs.
WherobotsDB runs the spatial half of that workflow: geocoding and enriching exposure, then joining it to hazard rasters and footprints at portfolio scale. In one project, Wherobots scored every building in Colorado, 2,771,126 Overture buildings, against five perils and the parcels they sit on, and published the results as the Colorado property risk explorer. A spatial join attaches each building to its hazard values, and a raster zonal statistic such as RS_ZonalStats samples flood depth or wildfire grids under each footprint.
Once a model writes a year event loss table, the EP statistics are SQL. This query reads AAL and the OEP and AEP losses at two return periods from a 10,000-year table:
WITH annual AS ( SELECT sim_year, SUM(loss) AS aep_loss, -- total of all events in the year MAX(loss) AS oep_loss -- largest single event in the year FROM my_catalog.cat_model.year_event_losses GROUP BY sim_year ), ranked AS ( SELECT aep_loss, oep_loss, ROW_NUMBER() OVER (ORDER BY aep_loss DESC) AS aep_rank, ROW_NUMBER() OVER (ORDER BY oep_loss DESC) AS oep_rank FROM annual ) SELECT COALESCE(SUM(aep_loss), 0) / 10000 AS aal, COALESCE(MAX(CASE WHEN oep_rank = 100 THEN oep_loss END), 0) AS oep_100yr, COALESCE(MAX(CASE WHEN aep_rank = 100 THEN aep_loss END), 0) AS aep_100yr, COALESCE(MAX(CASE WHEN oep_rank = 40 THEN oep_loss END), 0) AS oep_250yr, COALESCE(MAX(CASE WHEN aep_rank = 40 THEN aep_loss END), 0) AS aep_250yr FROM ranked
With 10,000 simulated years, the 1-in-100 year loss is the 100th largest year and the 1-in-250 year loss is the 40th. AAL divides by all 10,000 years, including years with no loss that never appear in the table. Those missing years rank below every year with a loss, so when fewer than 100 years have any loss, the 1-in-100 year loss is zero, and COALESCE returns 0 in place of NULL. The same applies to AAL when the table has no rows at all.
COALESCE
The models people use every day were trained on text, documents, databases, and the internet. They can define PML, and they cannot tell an underwriter how many insured buildings sit inside last night's hurricane wind field from a weather model such as HRRR. An AI agent connected to exposure and hazard tables through the Wherobots MCP server can answer that with a spatial query.
Join exposure to hazard layers with a Wherobots free trial at cloud.wherobots.com.
In insurance, PML stands for probable maximum loss: the largest loss an insurer can reasonably expect from a property or a portfolio. For a single building it is an engineering estimate of the worst likely fire or other event. For a catastrophe portfolio it is the loss at a chosen return period, such as the 1-in-100 year loss, read from an exceedance probability curve.
For a single site, engineers build a loss scenario, such as a fire with one protection system failed, and price the damage plus business interruption. For a catastrophe portfolio, a model simulates thousands of years of events against the insured locations, sums the losses for each simulated year, ranks the years, and reads the loss at the chosen return period.
AAL, average annual loss, is the mean loss per year across all simulated years. PML is a loss from the tail of the same distribution, such as the 1-in-250 year loss. AAL sets the long-run cost of the risk and feeds pricing. PML sets how much capital and reinsurance a bad year needs.
OEP, occurrence exceedance probability, is the probability that the largest single event in a year exceeds a loss. AEP, aggregate exceedance probability, is the probability that the total of all events in a year exceeds it. AEP losses are always at least as large as OEP losses at the same return period, because a year’s total includes its largest event.
A 1-in-100 year PML is the loss with a 1% probability of being exceeded in any one year. It does not mean the loss happens once a century. Over 30 years, a loss with a 1% annual chance has about a 26% chance of being exceeded at least once.
In property underwriting, maximum foreseeable loss (MFL) assumes every protection system fails, so it is the larger figure. PML assumes some protection works, such as fire walls holding while sprinklers fail. Normal loss expectancy (NLE) assumes all protection works as designed.
Seismic loss estimates for US commercial property follow ASTM E2026, the guide for seismic risk assessment of buildings. Its current edition frames building damage as probable loss (PL) and scenario loss (SL) assessments in place of the older seismic PML.
What is H3? Uber’s Hexagonal Spatial Index
H3 is Uber's open-source hexagonal grid for indexing locations on Earth. How the H3 index works, its 16 resolutions, H3 vs S2 and geohash, and H3 in SQL.
What is Geospatial Data? Types, Formats, Examples
Geospatial data describes things tied to a place on Earth. The two main types, vector and raster, plus formats, sources, and examples.
What is Geospatial Analysis? Methods and Examples
Geospatial analysis, or spatial analysis, examines where things are and how location shapes patterns. Methods, examples, GIS vs spatial analysis, and SQL.
share this article
Awesome that you’d like to share our articles. Where would you like to share it to:
