Loss Development Triangles: Construction and Age-to-Age Factors

The loss development triangle is the fundamental data structure of claims reserving: it organizes claims data by two time dimensions — when the claim originated and how long it has been developing — to reveal and project the pattern by which losses grow from immature reported values to their ultimate settled value.

Triangle construction

A triangle stacks origin periods (accident year, policy year, or report year) as rows against development periods (age since origin, e.g., 12 months, 24 months, ...) as columns, with the cell $C_{i,k}$ holding cumulative losses (paid or incurred) for origin year $i$ evaluated at age $k$. Because older origin years have had more time to develop, and each successive evaluation adds one more diagonal of data, the data naturally forms a triangle: origin year 1 has $n$ evaluations, origin year $n$ has only 1. The most recent diagonal — one value per origin year, all evaluated as of the same calendar date — is the latest diagonal, and it is these values that must be projected to ultimate.

Triangles can be built on paid losses (cash actually disbursed), incurred losses (paid + case reserves), reported claim counts, or closed claim counts; each answers a different diagnostic question and is used in different methods (see Reserving Diagnostics and Berquist–Sherman).

Age-to-age (ATA) factors

The age-to-age factor (or "link ratio" or "development factor") for development period $k$ to $k+1$, for a single origin year $i$, is

$f_{i,k} = \frac{C_{i,k+1}}{C_{i,k}}.$

To select a single factor for use across all origin years at that maturity, actuaries average the individual $f_{i,k}$ values across the available origin years using one of several averaging methods:

• Volume-weighted average: $\bar f_k = \dfrac{\sum_i C_{i,k+1}}{\sum_i C_{i,k}}$ — weights each origin year's link ratio by its own losses at $k$, so larger, more credible years dominate. This is the default/most common choice and is equivalent to the maximum-likelihood chain-ladder estimator under Mack's model assumptions.
• Simple (arithmetic) average: $\bar f_k = \frac1n\sum_i f_{i,k}$ — treats every origin year equally regardless of size, useful when years vary greatly in maturity-consistent reporting practice but volume shouldn't dominate.
• Medial (excl. high/low) average: drops the highest and/or lowest ratios before averaging, reducing the influence of a single anomalous year.
• Geometric average: multiplies ratios and takes the $n$th root; rarely used but sometimes appropriate when growth is genuinely multiplicative and skewed.

Practitioners often examine a 3-year or 5-year volume-weighted average alongside the all-year average, favoring shorter windows when there is a known trend (e.g., faster claim settlement in recent years) and longer windows for stability when data is thin.

Judgmental selection

Mechanical averages are a starting point, not the answer. The actuary must apply judgment for: outlier link ratios driven by one large claim; changes in claims handling (e.g., a new case-reserving philosophy creating a level shift mid-triangle — see Berquist–Sherman); known trend in the ratios across origin years (monotonic decline as claims close faster); and credibility of thin, immature years where a single link ratio can be wildly volatile. Selected factors should be reviewed against diagnostic ratios (loss ratios, average paid/reported severities) for consistency with what is otherwise known about the business.

Tail factors

Beyond the observed triangle's last column, a tail factor projects remaining development beyond the latest available age to ultimate, since many long-tailed lines (WC, GL, medical malpractice) have material development beyond 10 or even 20 years. Common tail methods: fitting a curve (inverse power, exponential decay) to the observed tail of the ATA factors and extrapolating; using industry benchmark tail factors (e.g., from Schedule P aggregate data or reinsurance industry tables); and the Bondy method, which assumes the last observed factor's ratio of decline continues (tail $= f_{n-1}^2 / f_{n-2}$, roughly).

Worked 4×4 example

Cumulative incurred losses (USD 000s):

Individual ATA factors, 12–24: $1500/1000=1.500$, $1700/1100=1.545$, $1850/1200=1.542$.

Volume-weighted 12–24: $(1500+1700+1850)/(1000+1100+1200) = 5050/3300 = 1.530$.

Individual ATA factors, 24–36: $1650/1500=1.100$, $1870/1700=1.100$.

Volume-weighted 24–36: $(1650+1870)/(1500+1700)=3520/3200=1.100$.

36–48: only one pair, $1700/1650=1.0303$.

Selected factors (volume-weighted, with a small selected tail of 1.02 for development beyond 48 months based on industry benchmarks): $f_{12-24}=1.530$, $f_{24-36}=1.100$, $f_{36-48}=1.0303$, Tail $=1.02$.

Cumulative development factors (CDFs) built by multiplying tail-to-date:

• $CDF_{48} = 1.02$
• $CDF_{36} = 1.0303\times1.02 = 1.0509$
• $CDF_{24} = 1.100\times1.0509 = 1.1560$
• $CDF_{12} = 1.530\times1.1560 = 1.7687$

Projected ultimates (applying the appropriate CDF to each origin year's latest diagonal value): AY2024 at 12mo, $1300\times CDF_{12}=1300\times1.7687=2{,}299$; AY2023 at 24mo, $1850\times CDF_{24}=1850\times1.1560=2{,}139$; AY2022 at 36mo, $1870\times CDF_{36}=1870\times1.0509=1{,}965$; AY2021 at 48mo, $1700\times CDF_{48}=1700\times1.02=1{,}734$.

Pitfalls

• Mixing paid and incurred data within one triangle inadvertently (e.g., a case-reserve reclassification injected mid-history).
• Using volume-weighted averages on a triangle with a level shift (see Berquist–Sherman) without adjusting the older diagonals first — the shift will bias every subsequent factor.
• Applying a single set of ATA factors uniformly across origin years with materially different (known) claims-handling eras.
• Ignoring the diminishing credibility of a tail factor built from very few, very old data points.

Exam relevance

Triangle construction and ATA factor selection are foundational to CAS Exam 6 and Exam 8, and to the SOA's general insurance reserving topics.

Further reading

• Friedland, Estimating Unpaid Claims Using Basic Techniques, CAS Study Note.
• Mack, Distribution-free calculation of the standard error of chain ladder reserve estimates, ASTIN Bulletin.
• Brosius, Loss Development Using Credibility, CAS Study Note.