I spent several years pricing excess-of-loss reinsurance treaties before I founded Morildsen. The work was technically demanding in a particular way: you were not pricing individual risks, you were pricing the aggregate behavior of an entire cedant portfolio under stress scenarios. The question was never "what is the expected loss on this risk?" It was "what does the cedant's loss distribution look like above the retention threshold, and what is the risk load for taking that tail?"
That experience shapes how I think about ML applications in reinsurance. The opportunities are real. The obstacles are specific. And the adoption pattern — I believe reinsurers will move faster than primary carriers, despite appearances — is driven by structural factors that are often misread from the outside.
Why Treaty Pricing Is Hard to Change
Reinsurance treaty pricing has historically been a practitioner art form with quantitative scaffolding. The core inputs to a quota share or excess-of-loss treaty are exposure data (the cedant's portfolio composition), loss experience (historical claims development patterns), and market cycle context (current ROE expectations for the risk class). A skilled treaty underwriter can integrate these into a price without a formal model — indeed, many still do.
The challenge with introducing ML into this workflow is not the modelling itself. It is the audit trail. Treaty pricing decisions carry long-tail liability: a treaty written today may still have open reserves from events occurring in years 2-3 of a three-year cover. The underwriter's decision must be defensible to actuarial review, to management, to the reinsurance regulators, and eventually to cedants disputing reserve estimates. A gradient boosting model that produces an accurate expected loss rate but cannot produce an interpretable attribution of the rate to specific risk factors is not useful in this context — regardless of its predictive accuracy.
The ML applications that are making inroads in reinsurance are the ones that address this explainability constraint directly: models where the output can be decomposed into a clean attribution showing how each portfolio characteristic contributes to the expected loss rate. Generalized additive models (GAMs) with interaction terms, SHAP-value frameworks applied to tree-based models, and structured factor models are finding adoption specifically because they are technically competitive with black-box approaches on accuracy while producing outputs that traditional underwriters can interrogate.
Why Reinsurers Will Adopt Faster Than Primary Carriers
The conventional view is that reinsurers are conservative institutions that adopt technology slowly. This is true for some dimensions of the business — claims management, policy servicing, distribution — where reinsurers have fewer economic incentives than primary carriers because the volume is lower. But it is wrong for pricing and portfolio analytics, and for structural reasons.
Reinsurers already operate quantitative modelling teams. The catastrophe modelling function — the group responsible for pricing natural catastrophe exposure using vendor models like RMS or AIR — has been running computationally intensive stochastic simulations for decades. The technical infrastructure, the data science culture, and the model validation protocols are already in place. Adding ML-based portfolio analytics is an incremental capability extension for a team that already runs stochastic models, not a paradigm shift.
Primary carriers, by contrast, frequently need to build the quantitative infrastructure from scratch. The actuarial function exists, but the data engineering stack, the model deployment infrastructure, and the ML-aware product management capabilities are nascent at most mid-size carriers. The organizational lift is larger.
Reinsurers also have a stronger motivation to improve pricing accuracy on individual cedant portfolios. A primary carrier mispricing one account is a local problem. A reinsurer mispricing a programme that contains concentrations from multiple cedants in the same geography or product line is a portfolio problem — it shows up in aggregate results in ways that are harder to attribute and harder to correct after the fact. The incentive to get treaty pricing right is higher.
The Data Sharing Problem at the Reinsurance Layer
The significant practical obstacle for ML in reinsurance treaty pricing is data quality and availability at the cedant portfolio level. Reinsurers receive exposure data in formats that vary by cedant, by broker, and by treaty class. The data arriving for a property catastrophe treaty might come as a Bordereau in one format from one cedant and in a completely different schema from the next. Normalizing this data into a consistent feature space for model training is an engineering challenge that is often underestimated.
The insurtech companies building in this space — and this is genuinely nascent at seed stage — are addressing it at the data standardization layer: building ingestion pipelines that normalize diverse cedant exposure data into a consistent schema before any modelling occurs. The boring infrastructure problem is the bottleneck. The modelling, once the data is clean, is relatively tractable. We look for founding teams who understand that and invest in the data engineering proportionally.
What ML Does Not Change in Reinsurance
It is worth being clear about the limits. ML does not change the fundamental nature of reinsurance: transferring tail risk between capital pools. The catastrophe bond market, the ILS market, and the reinsurance treaty market price tail risk that is, by definition, sparsely observed. No ML model trained on 20 years of loss data is going to make substantially better predictions about the loss distribution above the 1-in-100 year level than a well-constructed stochastic CAT model. The data is too thin for ML to dominate.
Where ML adds genuine value in reinsurance is in the less extreme loss layers: aggregate stop-loss pricing for frequency lines, motor excess-of-loss treaties where the loss development pattern is complex, workers' compensation programmes with long-tail medical development. These are data-rich domains where ML can extract signal from historical patterns that actuarial methods handle imprecisely. The catastrophe tail belongs to catastrophe modelling. The frequency layers are where ML earns its place.
That is a narrower claim than many insurtech companies in this space make. We prefer founders who have a clear map of where their model adds value and where it does not — the honest articulation of model limitations is a signal of technical maturity that correlates, in our experience, with production reliability.