Estimating reserves for claims incurred but not yet paid (or not yet reported) is one of the core tasks of the property and casualty actuary. For decades the tool of choice has been the chain-ladder, a deterministic, transparent and surprisingly robust method. However, the growing availability of granular data and the regulatory requirement to quantify uncertainty have driven a new generation of techniques, from the classic stochastic models to machine learning applied to claims triangles.
The chain-ladder method and its limits
The chain-ladder arranges claims in a development triangle, where each row represents an origin year (accident year) and each column a development period. From the cumulative amounts, development factors (age-to-age) are estimated as ratios of column sums, and the lower triangle is projected to obtain the reserve.
Its appeal is simplicity: it requires no explicit distributional assumptions and produces a point estimate that is easy to audit. But its limitations are well known. It assumes development patterns are stable over time, it is highly sensitive to the factors on the latest diagonals and to outliers, it does not use external information (exposure, inflation, changes in business mix) and, above all, it does not by itself provide a measure of the uncertainty of the estimated reserve. In a risk-based solvency framework, a point estimate without an error band falls short.
Stochastic methods: Mack and the bootstrap
The first major step was to acknowledge variability. The Mack model (1993) provides a distribution-free framework that reproduces the chain-ladder estimates and also derives an analytical formula for the mean squared error of prediction, separating process error from estimation error. This makes it possible to report not only the central reserve but also its standard deviation.
The bootstrap, for its part, takes a resampling approach. From the Pearson residuals of an over-dispersed Poisson (ODP) model, consistent with the chain-ladder, thousands of pseudo-triangles are generated which, once projected again, yield a full predictive distribution of the reserve. From it one obtains percentiles, value at risk and other tail metrics. These methods are now standard practice and are widely documented in the literature of the Institute and Faculty of Actuaries (IFoA) and the Casualty Actuarial Society (CAS).
The new wave: machine learning for reserving
Predictive analytics proposes going beyond the aggregate triangle. Instead of working with summed cells, machine learning approaches can operate at the level of the individual policy or claim, incorporating covariates such as region, line of business, initial severity or reporting delay.
- Tree-based models (gradient boosting, random forests) that capture nonlinear interactions between development variables and risk characteristics.
- Neural networks for claims triangles, which learn flexible development patterns. An influential line of research is that of Wüthrich and coauthors, who recast the chain-ladder as a neural network and propose extensions to incorporate additional information, as well as recurrent architectures to model the time dynamics of payments.
- Individual claims reserving models, which estimate the future development of each open claim, useful when the portfolio is heterogeneous or when there are structural changes in the business mix.
The promise is twofold: greater predictive accuracy and the ability to adapt to changing environments. The cost is a loss of transparency and a higher risk of overfitting, which calls for rigorous cross-validation and backtesting discipline.
Quantifying uncertainty
Whatever the method, today’s regulatory focus is the predictive distribution, not just the mean. Formally, what matters is the prediction error of the reserve $R$, usually decomposed as:
$$ \text{MSEP}(\hat{R}) = \underbrace{\text{Var}(R)}_{\text{process error}} + \underbrace{\text{Var}(\hat{R})}_{\text{estimation error}} $$
Machine learning methods address uncertainty through bootstrapping the residuals, Bayesian approaches or ensemble techniques that average multiple models. An open challenge is that many of these techniques quantify process risk well but underestimate model risk, that is, the uncertainty of having chosen the wrong structure.
Link to Solvency II
The link to regulation is direct. Under Solvency II, technical provisions are valued as the best estimate plus a risk margin, and the solvency capital requirement (SCR) for reserve risk is calibrated to a 99.5 percent value at risk over a one-year horizon. All of this requires precisely a distribution of the reserve, not a single figure. Stochastic and predictive methods are therefore the natural machinery to feed these calculations, whether through the standard formula or through internal models.
In the Mexican context, reserve valuation on a best estimate plus risk margin basis, as required by the local regulatory framework, shares the same philosophy of capturing uncertainty explicitly.
The role of the actuary
Predictive analytics does not replace actuarial judgment: it enhances it. A deep learning model can detect patterns the chain-ladder ignores, but it is the actuary who validates its reasonableness, checks results against classic methods and communicates uncertainty to management and the regulator. The practical recommendation is to tier the approach: use chain-ladder and Mack as the baseline and control, and reserve machine learning methods for large, granular portfolios where their marginal value justifies the complexity. Combining classic rigor with predictive flexibility is, today, the standard that modern reserving practice aims for.