Actuarial Modelling

The global actuarial profession faces a talent shortage that is reshaping how insurers operate. Qualified actuaries take years to train, and demand outstrips supply in every major market. If actuarial modelling continues to require the same ratio of qualified actuaries to model complexity, something has to change: either the models get simpler, or the tools get better.

What is actuarial modelling?

Actuarial modelling is the application of mathematical and statistical methods to quantify risk and uncertainty in insurance and finance. Actuaries build models that estimate future claim costs, set reserves, price policies, assess capital requirements, and test the financial resilience of an insurer under adverse scenarios. The outputs feed directly into underwriting decisions and claims processing reserves, making actuarial modelling foundational to an insurer's financial position and regulatory capital.

The work spans multiple timescales. Pricing models estimate the expected cost of a single policy over its term. Reserving models estimate the ultimate cost of claims already incurred but not yet fully settled. Capital models estimate the probability and severity of extreme events over a one-year horizon. Each requires different data, different techniques, and different validation approaches, but they share a common foundation: the actuary's judgement about which assumptions are reasonable and where uncertainty lies.

IFRS 17, the international accounting standard for insurance contracts, took effect in January 2023 and represents the largest change to insurance accounting in a generation. It requires insurers to measure insurance contract liabilities at current fulfilment value, with explicit recognition of risk adjustment and contractual service margin. The modelling requirements are substantially more complex than under the previous standard, and many insurers are still working through the operational implications.

The landscape

The talent shortage is not a recruitment problem; it is a structural one. The IFoA qualification takes three to six years, and the pipeline is not keeping pace with demand driven by regulatory complexity, climate risk modelling requirements, and the growth of new insurance markets. Insurers are responding by augmenting actuarial teams with data scientists, but the two disciplines have different training, different vocabularies, and different professional standards. Managing this interface is an organisational challenge.

Climate risk is widening the uncertainty band in catastrophe models. Historical loss experience, the traditional foundation of actuarial pricing, is becoming less predictive as climate patterns shift. The gap between modelled and observed losses has widened for perils including wildfire, convective storm, and flood. Actuaries are being asked to price risks for which the historical data may not represent the future, which is a fundamentally different challenge than statistical estimation from stable distributions.

The PRA's expectations on model risk management, articulated in SS1/23, apply to actuarial models as much as to any other model class. Validation, ongoing performance monitoring, and clear documentation of assumptions and limitations are regulatory requirements, not best practices. The governance overhead for actuarial models is increasing, and this is another pressure on already-constrained actuarial resources.

How AI changes this

The most immediate application is the acceleration of routine modelling tasks. Data preparation, assumption setting for standard products, sensitivity testing, and report generation consume a substantial portion of actuarial time. AI tools that automate these tasks, from data cleaning pipelines to automated sensitivity analysis, free actuaries to focus on judgement-intensive work: setting assumptions for novel risks, interpreting model outputs for the board, and challenging model limitations.

Machine learning models are being used as supplements to traditional actuarial models, particularly for pricing. Where a traditional GLM might use thirty to fifty rating factors, a gradient-boosted tree can incorporate hundreds, capturing non-linear interactions that the GLM structure cannot represent. The trade-off is interpretability: the GBM may predict more accurately, but explaining why a specific policyholder received a specific price is harder. The hybrid approach, using ML to identify factors and interactions that are then encoded into an interpretable model, is gaining traction.

Generative AI for scenario construction is emerging. Risk assessment traditionally relies on predefined scenarios: what happens if interest rates rise by 200 basis points, if a pandemic closes borders, if a Category 5 hurricane hits Miami. GenAI can generate novel scenarios that a human team might not consider, stress-testing the portfolio against a wider range of futures. The actuarial judgement shifts from constructing scenarios to evaluating whether AI-generated scenarios are plausible and relevant.

Predictive analytics applied to reserving enables earlier identification of deteriorating claim development patterns. Rather than waiting for quarterly reserve reviews to identify adverse trends, AI monitors claims development continuously, flagging segments where actual development is diverging from expected. This gives the reserving actuary earlier visibility into potential reserve strengthening and reduces the surprise factor in financial reporting.

What to know before you start

Actuarial models are regulatory artefacts. Any AI augmentation or replacement must satisfy the same governance, validation, and documentation requirements as the traditional model. Engage your model validation function and your regulatory relationship team before deployment. The IFoA's guidance on data science and AI, published in 2023, provides a professional framework that actuaries using AI tools should follow.

The IFRS 17 rebuild is a once-in-a-generation opportunity to modernise your modelling platform. If you are still running calculations on legacy actuarial platforms, the IFRS 17 implementation is the moment to invest in infrastructure that supports both traditional actuarial methods and machine learning. Retrofitting AI into legacy platforms is expensive and fragile; building on a modern data and computation platform from the start is significantly more cost-effective.

Climate risk modelling is not a data science project; it is a multidisciplinary one. The actuarial expertise to translate climate scenarios into financial impact, the climate science expertise to define plausible scenarios, and the data engineering expertise to integrate climate data into modelling pipelines are all required. Build the team before you build the model.

Start with reserving analytics. The data is available (claims history), the validation is straightforward (compare predictions against actual development), and the business impact is visible (earlier identification of reserve movements). Pricing applications are higher value but require more careful regulatory navigation. Reserving analytics builds the organisational confidence and the data infrastructure for more ambitious actuarial AI applications.