The largest liability on a P&C insurer's balance sheet: reserves for unpaid claims. A classical actuarial OR problem — chain-ladder, Bornhuetter-Ferguson, stochastic reserving.
In plain words
Sound familiar?
- Our actuarial team computes loss reserves by hand on a spreadsheet; quarter-end takes three weeks and the error risk is high.
- In our regulatory audit we were asked to provide a method breakdown and standard error; we only had a chain-ladder run.
- We write long-tail liability or workers comp; claims settle 5-10 years out and recent-year reserves are highly uncertain.
- In a high-inflation environment historical development factors no longer predict the future; we suspect under-reserving.
- Claim-handling policy (case reserve standards) changed in the past 2 years; chain-ladder keeps using the old pattern.
- Under Solvency II or equivalent we need a 75th percentile or risk margin; we need to move from point estimate to full distribution.
- We launched a new line; we have 2-3 years of data, chain-ladder is volatile — we are told to use Bornhuetter-Ferguson with an a-priori expected loss ratio but we don't know the method.
Why it matters
How it's solved
Technical depth
How it's solved
Technical depthOne-liner: Extract how past accident-year losses spread out over months or years; project new accident years forward along the same payment path. For mature years that’s reasonable (the pattern has settled); for young years, blend with an a-priori expectation (the pattern is still noisy).
This problem sits at the intersection of actuarial science and operations research (a discipline that uses math and computing to solve business-decision problems). Two sub-problems: reserve point estimate (expected ultimate loss) and reserve uncertainty (standard error, 75th percentile, risk margin). Solution in three stages:
1. Modeling. Build loss development triangles from the claims database: rows = accident year, columns = development period (months or years). Cell (i,j) = cumulative paid (or incurred) losses for accident year i, observed j periods after accident. The triangle is upper-triangular; the lower-right needs estimation. Homogeneity check: segment by line of business, claim type (attritional vs large), distribution channel, geography. Tail estimation: development beyond the last observed column. Method selection per cohort: CL for mature years, BF for green years, GLM for volatile patterns.
2. Solver-driven decision. Chain-Ladder age-to-age factors: ratios of cumulative amounts across development periods. BF blends an a-priori expected ultimate loss ratio (from pricing) with the CL development pattern. Stochastic: the Mack method gives a distribution-free standard error; bootstrap produces the full distribution; Bayesian/credibility approaches blend prior judgment with data. Munich Chain-Ladder uses paid and incurred triangles jointly. Open-source actuarial libraries implement all of these; commercial actuarial reserving software adds visual diagnostics and regulator-ready reports.
3. Field integration. Claims management system → triangles → reserving software → regulatory and internal audit reports. A quarterly reserving committee reviews method selection and backtest of past estimates against actual run-off. Audit trail: which cell came from which method, which development factor was selected, which a-priori was used. Annual actuarial function report (Solvency II) or local actuary report is standardized.
Alternatives
Manual spreadsheet + actuary's hand
FreeExisting spreadsheet software — license assumed sunk
Who it fits: Small insurer or single short-tail line
- + Zero additional software cost
- + Actuary in full control
- + Fast for small triangles
- − Stochastic methods — Mack/bootstrap — near impossible by hand
- − High error risk, weak audit trail
- − Does not scale across lines and segments
- − Cannot deliver 75th percentile when regulator asks
Local-market actuarial reserving software
Enterprise200K–800K local-currency license + 40K–120K/yr maintenance (TR market observation)
Who it fits: Mid-size insurer (100K-1M policies), multiple lines
- + Local language, regulator report templates
- + Basic chain-ladder and BF integrated
- + Local claims system integration
- − Stochastic methods (Mack, bootstrap) often limited
- − GLM and Munich CL may be missing
- − Diagnostics plots simple
International actuarial reserving software
Enterprise100K–500K EUR license + 25K–100K EUR/yr
Who it fits: Large insurer, multi-line, full Solvency II
- + Full stochastic (Mack, bootstrap, Bayesian)
- + Munich CL, GLM, Cape Cod, BF all included
- + Rich diagnostics and backtest
- + Solvency II SCR module integrated
- − Expensive
- − 6-12 month rollout, actuary training required
- − Local regulator templates customized
Open-source actuarial libraries + in-house build
Open SourceLicense free; in-house build 16-30 weeks or 400K-1.2M consulting
Who it fits: Insurer with strong tech team, special line structures
- + No license
- + All modern stochastic methods available
- + Custom reports possible
- − Need in-house actuary plus data scientist
- − Audit trail and regulator reports built by hand
- − Continuous maintenance
Recommendation
Ask in the meeting
- Are Mack 1993 distribution-free standard error and England-Verrall 2002 bootstrap fully implemented?
- Is Munich Chain-Ladder (paid + incurred jointly) and Bornhuetter-Ferguson in one tool or separate modules?
- In a high-inflation environment can you run inflation-adjusted CL (convert to real, then re-inflate to nominal at payment year)?
- Are diagnostic plots (residual plot, age-to-age trend, calendar year effect) standard output?
- Is there a backtest module — does it compare 3-5 year-old reserve estimates against actual run-off?
- Can the reserve distribution feed directly into the Solvency II SCR Non-Life Premium and Reserve risk module?
- Are local regulator actuary report templates ready, or do we adapt them?
- On contract end, in which format do we receive all triangles, method selections, output history, and audit trail?
Technical details
Editor’s note
This problem is colloquially called “IBNR reserve”, “outstanding claims provision”, or “incurred but not reported reserve”. The academic and actuarial literature calls it Claims Reserving, Loss Reserving, IBNR Estimation. Without these concepts you cannot tell the difference between an actuarial software that does reserve calculation (mechanical chain-ladder) and one that does reserve evaluation (method selection, diagnostics, backtest, stochastic distribution). The first only produces a number; the second produces actuarial judgement defensible to the regulator.
The most-skipped point in the field: chain-ladder assumes development patterns are stable. After product changes, claim-handling policy changes, or external shocks (pandemic, high inflation, regulatory change), the historical pattern no longer predicts the future. The chief actuary must judge when to override CL with BF or selected development factors. A software that just runs CL without diagnostic plots is dangerous.
The second most-skipped point: inflation in long-tail lines. Liability claims paid 5-10 years after accident accumulate medical and wage inflation. Chain-ladder implicitly assumes inflation continues at the historical rate; in high-inflation environments this breaks. Inflation-adjusted CL (paid-in-real-terms then re-inflated to nominal at payment year) is mandatory.
Step-by-step path — for the SMB insurer
Stage 1 — Clean the data first. At least 8-10 years of claim records from the claims management system: accident date, report date, payment dates and amounts, case reserve changes, closure status. Line of business, claim type, channel, geography fields valid. Re-opened claims flagged.
Stage 2 — Build the triangles. Separate triangles by line: cumulative paid, cumulative incurred, claim count. Attritional vs large claim split (e.g. top 5%). Tail estimation by Sherman or curve fitting. Homogeneity test.
Stage 3 — Method selection and pilot. Mature years (5+ development): Chain-Ladder + Mack standard error. Green years (1-2 development): Bornhuetter-Ferguson. Volatile line: GLM or bootstrap. Pilot 1-2 lines; backtest over 4 quarters.
Stage 4 — Rollout and regulator integration. All lines in 6-9 months. Quarterly actuary committee; annual regulator actuary report. If Solvency II or equivalent applies, the SCR Non-Life Premium and Reserve module is fed by the reserve distribution.
Risks — what can go wrong
- Data quality. Wrong payment dates, missing re-opened flags or late case reserve updates make the triangle misleading. Data quality control is the first hurdle.
- Method blindness. Running only CL — ignoring diagnostics. Residual plots showing calendar-year effects mean CL is broken; BF or GLM is needed.
- Inflation fallacy. Nominal CL on a long-tail line — future payments lose real value; reserves come out too low. Real-nominal conversion is mandatory.
- Single point estimate. When regulator asks for 75th percentile or risk margin you are caught unprepared. Stochastic methods must be in place beforehand.
- Missing audit trail. Which development factor was selected and why, which a-priori was used — without records, the audit cannot be defended.
Technical view of the solution method
| Approach | Typical use | Stochastic? | Data need |
|---|---|---|---|
| Chain-Ladder (deterministic) | Mature years, stable pattern | No | 5-10 year triangle |
| Mack 1993 distribution-free | CL + standard error | Yes (analytic) | 5-10 year triangle |
| Bornhuetter-Ferguson | Green years, sparse data | No (classical) | A-priori expected loss ratio |
| Cape Cod | BF variant, observation-driven a-priori | No | Triangle + exposure |
| GLM (over-dispersed Poisson) | Volatile line, multiple covariates | Yes | Incremental triangle |
| Bootstrap (England-Verrall 2002) | Full distribution, 75th percentile, risk margin | Yes (simulation) | Triangle + resampling |
| Munich Chain-Ladder | Paid-incurred inconsistency | Limited | Two triangles jointly |
| Bayesian/credibility (Wüthrich-Merz) | Very sparse data, strong prior | Yes | Prior + triangle |
Objective choice (reserving philosophy rather than optimization):
- Objective 1 — Best-estimate central reserve: Classical actuarial point estimate.
- Objective 2 — Minimum mean-squared error: Via Mack or bootstrap.
- Objective 3 — Conservative reserve at 75th percentile: Typical Solvency II use.
- Objective 4 — Risk margin (Cost-of-Capital): Regulator risk margin standard.
Multi-objective: best-estimate + risk margin = technical provision. Local regulators use similar decomposition.
Academic sources
Listed in the sources field of the frontmatter.
Sources
- Mack, T. (1993). Distribution-free calculation of the standard error of chain ladder reserve estimates. ASTIN Bulletin, 23(2), 213–225. Foundational stochastic chain-ladder reference.
- Bornhuetter, R. L. and Ferguson, R. E. (1972). The actuary and IBNR. Proceedings of the Casualty Actuarial Society, 59, 181–195. Original BF source.
- England, P. D. and Verrall, R. J. (2002). Stochastic claims reserving in general insurance. British Actuarial Journal, 8(3), 443–518. Bootstrap reserving reference.
- Wüthrich, M. V. and Merz, M. (2008). Stochastic Claims Reserving Methods in Insurance. Wiley. Standard textbook on stochastic reserving.
- Local insurance regulators (e.g. SEDDK in TR, EIOPA in EU) — actuarial regulations and technical provisions directives. Reference for compliance frameworks.
Glossary
- Claims Reserving
- The actuarial process of estimating future claim payments for events that have already occurred.
- Chain-Ladder Method
- Classical actuarial loss-reserving method that derives development factors from cumulative ratios in a loss development triangle.