Commentary on the Casualty Actuarial Society's 2026 Reserves Call Paper Program on improved methodologies and technologies for reserving, and on the expanded Schedule P loss reserving dataset released by the CAS in 2026. Details are available from the CAS.
The CAS has opened a 2026 call paper program on improved methodologies and technologies for reserving, noting that evolving technology has made enhanced modelling more practical. It has also updated its Schedule P-derived loss reserving dataset to include an additional ten accident years across major personal and commercial lines from US P&C insurers.
Those two things together are more significant than either alone. A methodological call without data produces theory. Data without a call produces scattered experiments. Both at once is how a discipline actually moves.
But it is worth asking a prior question that the call implies without stating: why did reserving lag pricing by roughly fifteen years in adopting machine learning? The answer is not conservatism, and understanding it properly tells you which new methods will work and which will not.
Four Structural Reasons Reserving Was Different
1. The unit of observation is not the unit of interest.
In pricing, you have hundreds of thousands of policies, each with an outcome. That is a supervised learning problem in its natural form.
In reserving, the traditional unit is the accident year — and you have perhaps twenty of them. A triangle is not a large dataset. It is twenty highly correlated observations of a process, presented as a matrix. Applying a method designed for large-n to something with n≈20 is not conservative caution; it is arithmetic.
The move to individual claim reserving changes this fundamentally, which is why it is the most important structural development in the field. Once the unit becomes the claim, you have hundreds of thousands of observations again.
2. The target is not observed.
Pricing models predict a loss that eventually occurs and can be checked. Reserving models predict ultimate — a quantity not observed for years, and for long-tail lines possibly a decade or more.
This breaks the machine learning development loop at its foundation. You cannot iterate rapidly against a validation set when the validation set will not exist until 2034. Backtesting against historical triangles helps, but every backtest is contaminated: the historical reserves influenced the claims handling, the settlement behaviour and the case reserve adequacy that generated the data you are testing against.
3. Reserving output is a financial statement item.
A pricing model that is wrong produces suboptimal rates and, eventually, a corrected experience. A reserving model that is wrong produces a misstated balance sheet, a restatement, and potentially an adverse regulatory finding. The Appointed Actuary signs an opinion on it.
The tolerance for methodological opacity is therefore genuinely lower, and appropriately so. Chain ladder's dominance is not inertia — it is that a chain ladder can be explained to an auditor in ten minutes and audited by re-performance in an afternoon.
4. Changes in the process invalidate the data.
Reserving methods assume the past predicts the future development pattern. Every change in claims handling, case reserving philosophy, settlement strategy, legal environment or product mix breaks that assumption. Experienced reserving actuaries spend much of their time detecting and adjusting for these breaks.
A machine learning model trained on historical development learns the historical operating environment, including all its handling practices. It will confidently extrapolate a settlement pattern that a change in claims strategy has already ended. Detecting that requires exactly the judgement that the model cannot supply — and, worse, the model's confidence obscures the need for it.
What Actually Works
Given those constraints, the methods that have earned their place share a common property: they add granularity or structure without discarding the actuarial framework.
Individual claim reserving. Modelling development at claim level using claim characteristics — injury type, litigation status, jurisdiction, claimant age, treatment path. This directly addresses the small-n problem and allows heterogeneous portfolios to be modelled without aggregation loss. It is the most substantive advance in the field and it pairs naturally with LLM-based extraction of claim characteristics from narrative documents, which the CAS has also been funding work on.
Mixed and hierarchical models. Partial pooling across segments — borrowing strength from related segments while allowing genuine differences to persist. This is the statistically principled solution to the perennial reserving problem of segments too small to model alone and too different to combine. The CAS's recent monograph work on practical mixed models for actuaries reflects a well-timed recognition that this belongs in the standard toolkit.
Stochastic reserving with credible ranges. Not new, but chronically underused. Boards and regulators increasingly ask for reserve ranges rather than point estimates, and a defensible range requires an explicit uncertainty model rather than a judgemental margin.
Machine learning for diagnostics rather than estimation. The application I would put first for most firms: use ML to detect changes in claims process, identify anomalous development patterns, and flag segments where the traditional method's assumptions are breaking down. The model does not produce the reserve — it tells the actuary where to look. This works with the grain of the four constraints rather than against them, and it is where I have consistently seen the best return in practice.
What the Expanded Schedule P Data Enables
The addition of ten accident years to the CAS Schedule P dataset matters for specific reasons.
Longer development windows for long-tail lines. Testing a reserving method on commercial liability requires observing development to something approaching ultimate. Ten additional accident years converts several lines from "insufficient to evaluate" to "evaluable."
Cross-cycle testing. Reserving methods behave differently in hardening and softening markets, and in low versus high inflation. A dataset spanning more cycles allows methods to be tested for robustness across regimes rather than only within one — which is precisely the failure mode that undoes reserving models in practice.
Genuine benchmarking. A common public dataset lets competing methods be compared on identical data, which is the mechanism by which methodological fields actually make progress. Without it, every paper reports strong results on proprietary data nobody else can access.
One caution worth stating: Schedule P is aggregate industry data. It is excellent for testing method behaviour and comparing approaches. It cannot be used to develop individual claim reserving methods, which need claim-level data. The most important methodological frontier is the one this dataset cannot directly support.
Five Recommendations for Reserving Teams
Start with diagnostics, not estimation. The fastest defensible value from ML in reserving is process-change detection. Build that first, prove it, then consider extending.
Treat individual claim reserving as a data programme first. The method is not the constraint — claim-level data quality, consistency and history are. Most firms need eighteen months of data work before the modelling question becomes live.
Keep the traditional method running in parallel, indefinitely. Not as a transition arrangement. As a permanent control. A material divergence between a granular model and a chain ladder is information, and you lose it the moment you decommission the benchmark.
Document assumption breaks explicitly and durably. Every change in claims handling, case reserving philosophy or settlement strategy should be recorded with its date and expected effect, in a register the reserving team maintains. This is the institutional knowledge that makes any method work, and it walks out of the door with experienced staff.
Make the range, not the point, the primary output. A single number invites false precision and offers no basis for discussing the uncertainty that actually matters to the board. This is a communication change more than a methodological one, and it is overdue.
The Thread
The through-line in all of this is that reserving's resistance to new methods was never conservatism. It was a correct response to four genuine structural features of the problem.
The methods now succeeding are the ones that address those features directly — granularity to solve small-n, hierarchical structure to solve segmentation, diagnostics to solve assumption breaks — rather than the ones that import a technique from pricing and hope the constraints do not apply.
That is a useful test for any new reserving method: which of the four does it actually solve?
Key Takeaways
- Reserving lagged pricing in adopting machine learning for four structural reasons: small effective sample size, unobserved targets, financial statement consequences, and process changes invalidating historical data.
- Individual claim reserving is the most substantive advance because it directly solves the small-n problem, and pairs naturally with LLM-based extraction of claim characteristics.
- Mixed and hierarchical models provide a principled solution to segments too small to model alone and too different to combine.
- The highest-return early application of ML in reserving is diagnostics — detecting process change and assumption breakdown — rather than estimation.
- The expanded CAS Schedule P dataset enables longer development windows, cross-cycle robustness testing and genuine benchmarking, but cannot support individual claim reserving development.
Frequently Asked Questions
Why is machine learning less established in reserving than in pricing? Because reserving's traditional unit of observation gives a very small effective sample size, the target quantity is unobserved for years, the output is a financial statement item signed by an actuary, and changes in claims handling invalidate the historical patterns that models learn from. These are structural constraints, not professional conservatism.
What is individual claim reserving? Modelling loss development at the level of the individual claim, using claim characteristics such as injury type, litigation status and jurisdiction, rather than aggregating into accident year triangles. It substantially increases the effective sample size and allows heterogeneous portfolios to be modelled without aggregation loss.
Should firms replace the chain ladder with machine learning methods? Generally not replace — run in parallel. The traditional method remains a valuable permanent control, and divergence between a granular model and a chain ladder is diagnostically informative. Most firms get more early value from using machine learning to detect assumption breakdown than from using it to produce the estimate.
Jonas Osman Abdelghafour is the founder of Quantica Risk, which builds reserving, capital and pricing models for insurers and provides independent model validation. This article is commentary on publicly available CAS research programmes.