Claims Development and Reserving

Uncertainty and Diagnostics

Distinguish process, parameter, and model uncertainty; backtest reserving assumptions.

Uncertainty and Diagnostics

A reserve estimate is conditional on data and assumptions. A complete analysis asks both how variable can future emergence be? and what if the model structure is wrong?

1. Separate four uncertainties

Source	Meaning	Example	Typical treatment
Process	future claims vary even with known parameters	next year's payments differ randomly	stochastic model or simulation
Parameter	factors and distribution parameters are estimated	Dev 2→3 uses one origin year	standard error, bootstrap, sensitivity
Model	another plausible structure gives another answer	calendar inflation breaks age-only development	alternative models and scenarios
Operational/data	recorded information is incomplete or unstable	reopened claims or system migration	reconciliation, controls, data-quality allowance

One confidence interval rarely captures all four.

2. Mack chain ladder: what it adds

Mack's distribution-free framework keeps chain-ladder mean structure while estimating conditional variance under assumptions such as:

E[C_{i,j+1}\mid C_{i,j}]=f_jC_{i,j},

\operatorname{Var}(C_{i,j+1}\mid C_{i,j})=\sigma_j^2C_{i,j},

with origin years independent. It yields standard errors for projected reserves without assuming a full claim distribution.

It does not automatically cover changing inflation, a new claims process, dependence across origin years, an omitted tail, or uncertainty in manually selected factors. Those are model and operational questions.

3. Bootstrap: preserve the workflow

A reserving bootstrap typically:

fits an incremental mean model;
computes and scales residuals;
resamples residuals to create pseudo-triangles;
refits development parameters;
simulates future process outcomes;
stores total reserve and origin-year results.

Refitting in each replicate captures parameter variation. Simulating future increments captures process variation. Merely sampling final point estimates captures neither correctly.

4. Backtest by valuation date

For each historical diagonal $t$ :

hide all observations after $t$ ;
fit the method using only information available at $t$ ;
predict the next diagonal or later ultimate;
compare with subsequently observed values;
repeat over several $t$ .

Useful measures include signed error, absolute percentage error where denominators are stable, interval coverage, and error by development age. Reserve backtesting must avoid look-ahead bias: factor selections and exclusions should reflect what could have been known at the historical valuation date.

Tiny backtest example

Using only AY 2020 and 2021 to estimate $f_{0:1}$ gives

\widehat f_{0:1}^{(-2022)}=\frac{180+220}{100+120}=1.8182.

It predicts AY 2022 Dev 1 paid of $140\times1.8182=254.5$ , versus £260m observed. The next-diagonal error is £5.5m, or 2.1% of observed. One good prediction is not validation; the example shows the information boundary.

5. Residual patterns diagnose structure

Pattern	Interpretation to investigate
residuals trend across origin years	portfolio mix, exposure, or underwriting change
residuals align on calendar diagonals	inflation, legislation, catastrophe, or operations
variance rises faster than fitted mean	heteroscedasticity or large-loss mixture
paid and incurred residuals diverge	settlement and case-reserve processes changed
youngest years fail repeatedly	prior or development pattern is stale

Residuals are prompts for investigation, not mechanical deletion rules.

6. Scenario table for the reserve committee

Scenario	Change from baseline	Question answered
factor selection	use trimmed or credibility-weighted link ratios	sensitivity to influential cells
tail	add plausible tail factors	cost beyond observed development
inflation	increase future severity/calendar effect	sensitivity to claims inflation
large loss	model separately or cap for factor selection	concentration risk
prior	vary ELR in BF	dependence on external expectation
settlement speed	shift payment timing without changing ultimate	paid-pattern instability

Present the rationale and probability status: a deterministic stress is not a percentile unless linked to a probabilistic model.

7. Current research: richer time and claim information

Recent work develops two complementary directions. State-space models make aggregate reserving dynamics and sampling distributions explicit. Individual-claim approaches use transaction histories, claim status, covariates, and survival methods. Examples include Selukar's state-space workflow (2025), the CAS Claim Life Cycle Model (2024), the ReSurv claim-count framework (2025), and a 2026 preprint connecting chain ladder with individual-claim prediction.

The teaching conclusion is measured:

individual data can explain reporting, closure, reopening, and payment timing;
richer data do not remove selection bias, censoring, or operational drift;
aggregate triangles remain valuable for reconciliation and benchmarking;
recent preprints are research-frontier evidence, not settled professional standards.

8. Minimum reserve report

A reviewer should be able to reproduce:

data scope and valuation date;
triangle transformations and exclusions;
factor/prior/tail selections;
central estimate by origin year;
process and parameter uncertainty where modelled;
alternative methods and scenarios;
backtest results and known failure modes;
judgement, ownership, and limitations.

Practice

Classify “future claim amounts vary around a correct expected value.”
Classify “a repair-cost shock affects all cells on the latest diagonal.”
Why is fitting on the full triangle and then pretending to backtest an earlier diagonal invalid?

Answers

Process uncertainty.
Primarily model/calendar uncertainty; parameter and process effects may accompany it.
The selected factors have already seen the held-out observations, so the error is optimistically biased by information leakage.

Sources and further reading

Mack, T. (1993), “Distribution-Free Calculation of the Standard Error of Chain Ladder Reserve Estimates,” ASTIN Bulletin, 23(2), 213–225.
England, P. D. and Verrall, R. J. (2002), “Stochastic Claims Reserving in General Insurance,” British Actuarial Journal, 8(3), 443–518.
CAS — The Development and Use of a Claim Life Cycle Model (2024)
CAS — Claim Counts Prediction Using Individual Data with ReSurv (2025)
Selukar — Stochastic Claims Reserving Using State Space Modeling (2025 preprint)
Richman and Wüthrich — From Chain-Ladder to Individual Claims Reserving (2026 preprint)

Edit this pageorReport an issue

Bornhuetter–Ferguson Method

Blend an expected-loss prior with observed maturity without letting early random payments dominate.

Frequency and Severity Models

Select loss models from the insurance question, observation process, and tail—not from a distribution list.