Claims Development and Reserving

Build a defensible claims-development data contract before projecting a reserve.

Claims Development and Reserving

Reserving estimates the future cost of events that have already occurred. A run-off triangle is a data arrangement, not a method: rows identify origin periods, columns identify development age, and diagonals identify calendar periods.

1. Fix the data contract

Before calculating a factor, document:

QuestionHarbour Mutual answerWhy it matters
What is one row?accident yearpolicy year and report year answer different questions
What is one column?annual development ageunequal intervals invalidate a simple factor comparison
Incremental or cumulative?cumulative paid losslink ratios require a consistent basis
Gross or net?gross of reinsurancemixing bases changes development mechanically
Currency and price level?nominal £minflation can create calendar effects
Information date?31 December 2023determines which cells were genuinely observable
What is “ultimate”?selected value after Dev 3oldest observed does not automatically mean complete
Zero means “observed and no amount occurred.” A blank upper-right cell means “not observed by the valuation date.” Confusing them biases every projection.

2. See all three time axes

For origin year ii, development age jj, calendar year is i+ji+j. Therefore cells on a diagonal share an economic and operational period.

AYDev 0Dev 1Dev 2Dev 3
2020100180240280
2021120220300
2022140260
2023160

The bold diagonal is the latest observation. A claims-system change in calendar 2023 can affect all four bold cells, even though they have different development ages. A pure development model may mistake that diagonal effect for ordinary maturation.

3. Convert carefully

If Xi,jX_{i,j} is incremental loss and Ci,jC_{i,j} cumulative loss,

Ci,j=k=0jXi,k,Xi,j=Ci,jCi,j1.C_{i,j}=\sum_{k=0}^{j}X_{i,k}, \qquad X_{i,j}=C_{i,j}-C_{i,j-1}.

For AY 2020, cumulative (100,180,240,280)(100,180,240,280) becomes incremental (100,80,60,40)(100,80,60,40). Negative paid increments usually indicate recoveries, corrections, or data errors; negative incurred increments can also reflect case-reserve releases. Investigate rather than deleting them automatically.

4. Diagnose before modelling

Use this order:

  1. Reconcile triangle totals to the source system or ledger.
  2. Check duplicates, currencies, gross/net basis, and reopenings.
  3. Plot age-to-age factors by origin year.
  4. Plot incremental proportions by development age.
  5. Inspect calendar diagonals for inflation, legislation, operational changes, or catastrophes.
  6. Separate large losses or materially different claim types when justified.
  7. Record every adjustment and retain an unmodified extract.

Mini-case: faster settlement or lower severity?

Suppose Dev-0 paid loss rises from 140 to 160 while ultimate claim cost is unchanged. A chain ladder may interpret the higher early payment as a higher ultimate. An operational explanation—claims are being settled earlier—implies a lower later factor instead. Paid triangles cannot distinguish these explanations without claim counts, incurred data, or operational evidence.

5. Choose the question, then the method

MethodMain evidenceUseful whenMain vulnerability
Chain ladderobserved development patterntriangle is sufficiently mature and stablerecent years inherit historical pattern
Frequency–severitycount emergence and average costcount and cost processes explain the movementinconsistent count definitions; interaction between processes
Bornhuetter–Fergusonprior ultimate plus observed maturityimmature periods need external structureprior may be weak or stale
Uncertainty and diagnosticsresiduals, backtests, stochastic assumptionsa point reserve is not enoughassumptions can be hidden behind one error number

These are not competing buttons. A reserve committee compares their evidence, understands differences, and selects a range or central estimate.

Practice: classify the problem

For each observation, identify the most relevant axis and one response.

  1. All claim types settle faster after a new payment platform goes live in 2022.
  2. Bodily-injury claims have much larger and later payments than vehicle-damage claims.
  3. One catastrophe affects several origin years through reopened claims in the same calendar year.
  4. A recent year has almost no development data, but pricing supplied an expected loss ratio.
Suggested answers
  1. Calendar/operational effect: inspect diagonals and paid-versus-reported speed; do not extrapolate old link ratios blindly.
  2. Portfolio heterogeneity: segment if definitions and volume support it.
  3. Calendar/event effect: isolate and model the event where appropriate.
  4. Immaturity: BF or another prior-informed method may be more stable than chain ladder alone.

Public data for extended work

The Casualty Actuarial Society provides updated Schedule P loss-reserve data and associated research resources. They are useful for reproducing methods on real triangles, but Schedule P line definitions and US reporting conventions must be documented before transferring a conclusion to another market.

Copyright © 2026