AI/ML Catastrophe Modelling in Indian Commercial Property Underwriting 2026: Hazard Data, Exposure Accumulation, the India Model Gap and Pricing

Indian commercial property underwriting lives or dies on natural-catastrophe risk. The country's exposure is among the most varied in the world: the entire eastern and western coastline is cyclone-prone, the Indo-Gangetic plain and the coastal cities flood with the monsoon, large parts of the north and northeast sit in high seismic zones, and the same industrial cluster can carry flood, cyclone and earthquake exposure at once. A property underwriter pricing a manufacturing plant, a warehouse, a data centre or a hotel is, whether explicitly or not, taking a view on how that asset performs when a once-in-fifty-year or once-in-two-hundred-year peril strikes.

Catastrophe modelling is the discipline that turns that view into numbers. A cat model takes the hazard (how often and how severely a flood, cyclone or earthquake hits a location), the vulnerability (how much damage that intensity does to a given building type), and the exposure (what is insured there and for how much), and produces a loss distribution: the expected annual loss, and the probable maximum loss at defined return periods. From that distribution come the two outputs an underwriter most needs: a technical price for the catastrophe element of the risk, and a measure of how much loss the portfolio accumulates in a single event, against which accumulation limits are set.

The stakes are large. A single severe event can produce losses that dwarf a normal year's claims, and the way Indian insurers price and accumulate catastrophe risk feeds directly into their reinsurance treaties, their solvency and their results. Mispricing the cat element, or failing to see how exposure accumulates in a flood-prone city or along a cyclone-exposed coast, is how a property book that looks profitable in a quiet year produces an outsized loss in a bad one.

This post is about the modelling used in underwriting: the hazard data and exposure information that feed it, the geocoding and accumulation that make it usable, the gap in India-specific models, how machine learning augments traditional vendor models, climate-conditioned views, and how an underwriter actually uses the output for pricing and accumulation limits. It is not about catastrophe pools or post-event response; it is about the model that sits behind the underwriting decision.

A cat model is only as good as the data fed into it, and in the Indian commercial-property context the data quality is often the binding constraint. The three inputs (hazard, exposure and vulnerability) each have their own data problems.

Hazard data

The hazard layer describes the peril at each location: flood depth at given return periods, cyclone wind speed and surge, ground shaking intensity for seismic zones. Some of this comes from public sources (the Bureau of Indian Standards seismic zonation that underlies IS 1893, the IMD cyclone climatology, the central and state flood-hazard mapping), some from scientific and remote-sensing data (digital elevation models, satellite-derived flood extents, river-gauge records), and some from commercial hazard datasets. The Indian hazard data is patchier and lower-resolution than in mature markets, particularly for flood, where fine-grained inundation mapping that distinguishes one street from the next is the difference between a usable and a useless model.

Geocoding the exposure

The model can only place a risk against the hazard if it knows where the risk is, and that means geocoding: converting an address into precise coordinates. In India this is genuinely hard. Addresses are inconsistent, informal and often imprecise, the same location can be described many ways, and a geocode that lands a flood-exposed warehouse a few hundred metres from its true position can put it on the wrong side of a flood line. Poor geocoding is one of the largest and least-visible sources of error in Indian cat modelling, because it silently mislocates the exposure against the hazard.

The exposure record

The exposure data is the schedule of insured values, building characteristics, occupancy and location for every risk in the portfolio. Its quality is frequently poor: locations recorded only to city or pin-code level, missing or default construction and occupancy details, and sum-insured figures that may not reflect reinstatement-value. A model run on weak exposure data produces confident-looking numbers built on sand.

The output an underwriter uses as much as the price is accumulation: how much loss the portfolio suffers when a single catastrophe strikes a single area. Pricing each risk in isolation is not enough, because catastrophe risk is correlated, and many individually-sound risks in the same flood plain or the same cyclone track can produce a portfolio loss that breaks the book.

Accumulation analysis answers the question of concentration. The cat model, having geocoded the exposure and placed it against the hazard, computes how much of the portfolio is exposed to the same event: the total insured value and modelled loss within a flood footprint, a cyclone swath or an earthquake's affected radius. This reveals where the book is concentrated (a cluster of insured warehouses in a Mumbai or Chennai flood zone, a set of plants along the cyclone-exposed east coast, a portfolio weighted to a seismic city) and how large a single-event loss could be.

From the accumulation analysis come accumulation limits: the maximum exposure or modelled loss the insurer is willing to carry in a defined zone or to a defined event. These limits are an underwriting control. When a zone approaches its accumulation limit, the underwriter has to decline, reduce or reinsure additional risk in that zone, regardless of how attractive the individual risk looks, because the marginal risk adds to a concentration that is already near the ceiling.

Accumulation also drives the reinsurance programme. The probable maximum loss the cat model produces for the portfolio at a chosen return period sets how much catastrophe reinsurance cover the insurer buys, and the accumulation by zone informs how the treaty responds. An insurer that models its accumulation well buys reinsurance that matches its real exposure; one that models it poorly is either over-paying for cover it does not need or, worse, under-protected against the event that finds its concentration.

The practical underwriting discipline is to run accumulation continuously, not annually. As each new risk is bound, its marginal contribution to the zonal accumulation should be visible, so the underwriter sees a risk that pushes a zone toward its limit before binding it, rather than discovering the concentration after a season of writing has built it up.

Mature insurance markets rely on vendor catastrophe models from a small number of established providers, and those models are detailed, peer-reviewed and embedded in underwriting and reinsurance. For India, the picture is thinner. Vendor models exist for Indian earthquake and cyclone, and flood models have improved, but the coverage is less complete, the resolution coarser and the validation weaker than for the markets the vendors built first. The Indian flood model gap in particular is real: India's flood exposure is enormous and complex, and the models that capture it at the resolution underwriting needs are still maturing.

This gap is where machine learning earns its place, and the right framing is augmentation, not replacement. The established vendor models encode decades of scientific and engineering knowledge about how perils behave and how buildings respond, and that physical grounding is valuable. Machine learning adds to it rather than discarding it.

The ways ML augments the traditional models include:

1. Better hazard data. ML applied to satellite imagery, digital elevation models and historical flood extents can produce finer-grained flood-hazard mapping than the public data offers, filling the resolution gap that hurts Indian flood modelling most.
2. Better geocoding and exposure enrichment. ML can parse messy Indian addresses into more accurate coordinates and can infer building characteristics (construction, height, occupancy, roof type) from imagery and other data, improving the exposure record the model runs on.
3. Vulnerability refinement. ML can help calibrate how Indian building stock actually responds to flood, wind and shaking, using claims and damage data, where the standard vulnerability curves were derived from other markets' building types.
4. Filling gaps in sparse vendor coverage. Where a vendor model is thin for a peril or region, ML-based approaches trained on available hazard and loss data can provide a usable view rather than leaving a blind spot.

Catastrophe models were traditionally built on the assumption that the historical climate is a reasonable guide to the future, and for Indian perils that assumption is increasingly questionable. The monsoon is shifting, extreme rainfall events are becoming more frequent and intense, sea levels are rising along the coasts where so much industrial exposure sits, and cyclone behaviour is changing. A model calibrated to the past underprices the future where the future is more hazardous than the past, and Indian flood and coastal exposure is where this matters most.

A climate-conditioned view of catastrophe risk adjusts the modelled hazard to reflect the changing climate rather than the historical baseline alone. Instead of asking what the flood or cyclone risk was over the historical record, it asks what the risk is under present and near-future climate conditions, conditioning the hazard frequencies and severities on the climate signal. For an underwriter pricing a multi-year exposure or an asset with a long life, the climate-conditioned view is the more honest basis, because the historical view systematically understates a risk that is trending up.

This matters in concrete underwriting terms. A coastal industrial asset, a flood-exposed warehouse cluster, or a low-lying data-centre site may price acceptably on a historical-climate model and unacceptably on a climate-conditioned one, and the difference is the part of the risk the historical view misses. An underwriter who prices and accumulates on the historical view alone is building a book that looks adequate now and is increasingly exposed as the climate moves, while one who incorporates a climate-conditioned view sees the trend in the price and the accumulation.

The regulatory and disclosure direction reinforces this. Climate-risk disclosure expectations are rising for Indian corporates and insurers, and an insurer that can show it prices and accumulates catastrophe risk on a forward-looking, climate-conditioned basis is better placed than one relying on a backward-looking model. The climate-conditioned view is not a separate exercise from the cat model; it is a conditioning of the same model's hazard layer, and machine learning applied to climate and hazard data is one of the tools that makes producing such views tractable.

The reason an underwriter runs a cat model is to make two decisions better: what to charge for the risk, and how much of it to take in a given zone. The model output feeds both, but it informs the decision rather than making it, and how the underwriter uses it is where the modelling becomes underwriting.

Pricing the catastrophe element

The model produces an expected annual catastrophe loss for the risk, which is the technical cat-loss cost that the catastrophe element of the premium has to cover, on top of the attritional (non-catastrophe) loss cost, the expenses, the cost of capital and the cost of the catastrophe reinsurance that protects the risk. The underwriter uses the modelled cat-loss cost as the floor for the catastrophe element of the price: a risk in a high-flood or high-cyclone zone carries a higher modelled cat-loss cost and should carry a higher price, and a risk the model shows as benign carries less. Where market pressure pushes the achievable price below the technical cat-loss-inclusive cost, the model makes that visible, so the underwriter is declining or accepting a sub-technical price knowingly rather than blindly.

Setting and respecting accumulation limits

The accumulation output sets the zonal limits, and the underwriter uses them as a binding control on what to write. The discipline is to test each new risk's marginal contribution to the relevant zone's accumulation before binding, and to decline, reduce, sub-limit or reinsure additional exposure in a zone that is at or near its limit. This is where good underwriters part from bad ones: the individual risk may be attractively priced, but if it adds to a flood or cyclone concentration already near the ceiling, writing it is taking on correlated risk the portfolio cannot afford, and the accumulation control is what stops that.

Judgement over the model

The model output is an input to underwriting judgement, not a substitute for it. The underwriter has to weigh the model's known weaknesses (the geocoding quality, the exposure-data gaps, the resolution limits of the Indian hazard data, the uncertainty in the climate-conditioned view) and not treat a precise model number as a precise truth. A defensible underwriting practice uses the model to inform the price and the accumulation, records the basis of the decision (the model used, the data quality, the assumptions, the judgement applied), and keeps the underwriter accountable for the outcome. This matters for underwriting governance and for the model-risk expectations that increasingly apply to model-driven decisions in Indian insurance: the model is a tool, the underwriter owns the decision, and both have to be auditable.

Building a catastrophe-modelling capability that improves Indian commercial property underwriting is a matter of getting the data, the models, the accumulation discipline and the governance right together, rather than buying a model and trusting its output. The insurers and brokers that get value from cat modelling treat it as a system: better inputs, augmented models, continuous accumulation, climate-conditioned views and disciplined use of the output, all under proper governance.

The practical priorities are clear. Fix the inputs first, because geocoding and exposure-data quality usually limit accuracy more than the model does. Use machine learning to augment vendor models on hazard resolution, geocoding, exposure enrichment and vulnerability calibration, while keeping the physical model structure that anchors the output to how perils behave. Run accumulation continuously so concentration is visible before it is bound, not after. Incorporate a climate-conditioned view so the price and the accumulation reflect a changing climate rather than the historical record alone. And govern the whole thing so the model is a documented input to an accountable underwriting decision, with the model, the data quality and the assumptions recorded.

The modelling sits inside the wider underwriting of the property risk, and the model output is only one part of how the cover is structured and priced. The policy that responds when the modelled catastrophe actually strikes is defined by its wording: the perils covered, the flood, cyclone and earthquake sub-limits, the deductibles and franchise terms, the average-clause and valuation basis, and the exclusions that decide whether a given loss is paid. An underwriter or broker who models the catastrophe risk well still has to match that modelling to a wording that responds the way the pricing assumes, and the two have to be consistent.

That is where structured access to the market's property wordings supports the modelling. Sarvada gives commercial insurance brokers and underwriters structured, searchable access to insurer property policy wordings, so the catastrophe perils, sub-limits, deductibles and exclusions a risk is actually placed on can be compared across insurers and matched to the modelled exposure, rather than priced on a model and placed on a wording that does not align. Request Access to connect the catastrophe modelling behind a property risk to the wordings that decide how the cover responds.