Treaty Pricing

Reinsurance treaties are priced on assumptions that stretch decades into the future, yet the data behind those assumptions often looks backwards. When a long-tail liability treaty's claims can emerge fifteen or twenty years after inception, how confident is anyone in the burning cost calculation? Treaty pricing is where actuarial modelling meets genuine uncertainty, and AI is starting to change what that uncertainty looks like.

What is treaty pricing?

Treaty pricing is the actuarial process of determining the premium for a reinsurance treaty, the contract by which an insurer transfers a portfolio of risk to a reinsurer. Unlike facultative placement, which covers individual risks, treaty reinsurance covers entire books of business: all motor claims above a retention, all property losses in a geographic zone, all professional liability within a programme structure. The reinsurer is pricing not a single risk but a distribution of outcomes across thousands or millions of underlying policies.

Three methods dominate. Burning cost analysis uses historical loss experience to project future claims, adjusting for inflation, exposure changes, and development patterns. Experience rating refines this by weighting the cedant's own loss history against market benchmarks. Exposure rating bypasses historical losses entirely, estimating expected losses from the underlying risk profile using industry loss curves and catastrophe models. In practice, actuaries blend all three, and the blend itself is a judgement call.

The difficulty is sharpest in long-tail lines. A casualty treaty priced today will generate claims that settle in 2040 or later. Social inflation, the tendency for jury awards and litigation costs to rise faster than general inflation, has widened the gap between actuarial projections and actual outcomes. US casualty reserves across the market have been deficient by an estimated 100 billion dollars over the past decade, much of it attributable to social inflation that pricing models failed to anticipate.

The landscape

IFRS 17 has changed how reinsurance contracts are measured and reported. The standard requires reinsurers to measure contract liabilities at current fulfilment value, with explicit risk adjustments and contractual service margins. This increases the actuarial workload per treaty and creates new demands for granular data from cedants. Pricing actuaries now need to consider not just the expected loss but the accounting treatment of the treaty over its lifetime.

The PRA and EIOPA have both intensified scrutiny of reserving adequacy in reinsurance, particularly for casualty lines. Supervisors want evidence that pricing reflects current claims trends, not just historical averages. The days of pricing a long-tail treaty purely on five-year burning cost without adjusting for social inflation, litigation funding, or regulatory environment changes are ending.

Climate risk is reshaping property treaty pricing. The gap between modelled and actual catastrophe losses has widened, driven by secondary perils like convective storms, wildfire, and flooding that vendor models historically underpriced. The 2023 and 2024 renewal seasons saw rate increases of 20 to 40 per cent on loss-affected programmes, but the question for pricing actuaries is whether those increases are sufficient given the non-stationarity of the underlying hazard.

How AI changes this

Machine learning models are augmenting traditional pricing methods by incorporating data sources that actuarial models historically ignored. Climate trend data, litigation funding activity, legislative changes, and macroeconomic indicators can be systematically integrated into pricing models rather than applied as ad hoc adjustments. For property catastrophe treaties, ML models trained on satellite imagery and real-time weather data provide exposure estimates that update continuously rather than annually.

Natural language processing applied to claims reserving data and legal filings extracts early signals of emerging loss trends. A shift in judicial interpretation of policy wordings, an increase in class action filings in a specific jurisdiction, or a new theory of liability can be detected from court records and legal publications months before it appears in aggregate loss statistics. This gives pricing actuaries leading indicators rather than lagging ones.

Scenario generation using generative AI produces a wider range of plausible loss scenarios than traditional stochastic models. Rather than sampling from a fitted distribution, AI-generated scenarios can incorporate tail dependencies, regime changes, and novel event combinations. The actuary's role shifts from constructing scenarios to evaluating whether AI-generated scenarios are plausible and pricing-relevant.

Portfolio optimisation models help reinsurers assess how a new treaty interacts with their existing book. The marginal capital cost of writing a treaty depends on its correlation with the rest of the portfolio, and AI-driven optimisation can evaluate thousands of portfolio combinations to identify the most capital-efficient mix of treaties at each renewal.

What to know before you start

Data from cedants is the binding constraint. Treaty pricing depends on bordereaux data, loss triangles, and exposure schedules that arrive in inconsistent formats with varying levels of granularity. No ML model compensates for incomplete or inaccurate input data. Invest in automated data ingestion and validation before investing in model sophistication.

Regulators expect transparency. The PRA's SS1/23 on model risk management applies to pricing models, including those that incorporate machine learning. A black-box pricing model that produces a number without an auditable rationale will not satisfy supervisory expectations. Favour hybrid approaches where ML identifies signals and traditional actuarial methods translate them into pricing adjustments that can be explained to the board and the regulator.

Social inflation is not a single variable; it is a collection of correlated trends across jurisdictions and lines. Modelling it requires legal expertise as much as statistical technique. Build a multidisciplinary team, actuaries, claims specialists, and legal analysts, before attempting to incorporate social inflation into an AI-driven pricing model.

Start with short-tail property treaties where the feedback loop between pricing and outcome is fastest. You can validate model performance within two to three years. Long-tail casualty, where the ultimate answer takes a decade or more, is a harder environment to demonstrate ML value. Use property treaties to build the infrastructure and the organisational confidence, then extend to casualty with appropriate humility about the validation horizon.