Paid And Incurred Triangle Uncertainty Comparisons

Why Compare Paid and Incurred Triangles?

Paid and incurred triangles describe related but different views of claim development. A paid triangle records payments made to date. An incurred triangle usually combines those payments with case estimates of amounts still outstanding. The incurred view can contain information about expected ultimate losses earlier than the paid view, but it also reflects revisions in case estimates. Those revisions can add volatility, even when the underlying payment experience is steady.

A useful comparison therefore asks more than which triangle has the larger development factor or variance estimate. It asks whether the two triangles have comparable development patterns, whether their residuals show different structure, and whether case-reserve changes help explain that structure. A lower estimate of uncertainty in one triangle is not automatically evidence that its predictions are more reliable.

Definition: Let $P_{i,j}$ be cumulative paid losses for origin year $i$ through development age $j$, and let $R_{i,j}$ be case outstanding reserves at that valuation. Define cumulative incurred losses by $I_{i,j}=P_{i,j}+R_{i,j}$. The corresponding incremental changes between ages $j$ and $j+1$ satisfy $\Delta I_{i,j+1}=\Delta P_{i,j+1}+\Delta R_{i,j+1}$, where $\Delta R_{i,j+1}=R_{i,j+1}-R_{i,j}$.

A case-reserve release has $\Delta R<0$, so the incurred increment can be smaller than the paid increment, or even negative. That does not mean that payments were negative: it means the reduction in the estimated unpaid amount exceeded the payment made during that interval. This distinction is central when interpreting incremental incurred residuals.

A Mack-Style Basis for Comparing Residuals

For either a paid or incurred cumulative triangle, a basic Mack-style development model assumes that, at development age $j$, the next cumulative amount has conditional mean proportional to the current cumulative amount and conditional variance proportional to it. For one triangle, write $C_{i,j}$ for its cumulative amount and $f_j$ and $\sigma_j^2$ for the development factor and variance parameter at age $j$. The model states

E[C_{i,j+1}\mid C_{i,j}]=f_jC_{i,j}, \qquad \operatorname{Var}(C_{i,j+1}\mid C_{i,j})=\sigma_j^2C_{i,j}.

The usual interpretation also requires appropriate independence across origin years, positive cumulative amounts for the variance scaling, and a sufficiently stable development pattern. The assumptions are applied separately to paid and incurred triangles; they do not imply that the two triangles are independent of one another. Case-reserve revisions can make incurred development violate the assumed pattern, so a fitted variance parameter should be treated as a model-based summary, not a universal measure of uncertainty.

Given observed development pairs, the volume-weighted chain-ladder factor is

\hat f_j=\frac{\sum_i C_{i,j+1}}{\sum_i C_{i,j}}.

A development residual is $e_{i,j}=C_{i,j+1}-\hat f_j C_{i,j}$. A common estimate of the development-period variance parameter, when there are $n_j>1$ observed pairs, is

\hat\sigma_j^2= \frac{\sum_i C_{i,j}(f_{i,j}-\hat f_j)^2}{n_j-1} = \frac{\sum_i e_{i,j}^2/C_{i,j}}{n_j-1}, \qquad f_{i,j}=\frac{C_{i,j+1}}{C_{i,j}}.

The corresponding standardized residual is

r_{i,j}=\frac{e_{i,j}}{\sqrt{\hat\sigma_j^2C_{i,j}}}.

This residual measures the deviation from the fitted development amount relative to the variability allowed by the model at that cumulative scale. It is more useful than comparing unscaled residuals across cells of very different sizes. It remains a diagnostic, however: a small standardized residual does not prove that the model assumptions hold.

Why the Residual Scaling Works

The residual denominator follows directly from the conditional variance assumption. If the true factor $f_j$ were known, the deviation $C_{i,j+1}-f_jC_{i,j}$ would have conditional mean zero and conditional variance $\sigma_j^2C_{i,j}$. Dividing that deviation by its conditional standard deviation therefore gives a quantity on a common scale:

\operatorname{Var}\left( \frac{C_{i,j+1}-f_jC_{i,j}}{\sigma_j\sqrt{C_{i,j}}} \,\middle|\,C_{i,j} \right) = \frac{\sigma_j^2C_{i,j}}{\sigma_j^2C_{i,j}} =1.

In practice, $f_j$ and $\sigma_j^2$ are estimated, which is why the fitted residual uses $\hat f_j$ and $\hat\sigma_j^2$. Estimation affects the residuals, particularly when few origin years are available. The scaling explains the comparison; it does not make the residuals independent or guarantee a standard normal distribution.

Worked Example: Reconciling Paid and Incurred Increments

A hypothetical insurer has the following cumulative amounts, in thousands of dollars, for one origin year. At development ages 1, 2, and 3, cumulative paid losses are 100, 180, and 220. Case outstanding reserves are 60, 5, and 0. Find the incurred cumulative amounts and the paid and incurred increments from age 1 to age 2 and from age 2 to age 3. Explain the age-2 incurred increment.

Method. Add case outstanding reserves to cumulative paid amounts to obtain incurred amounts:

I_1=100+60=160,\qquad I_2=180+5=185,\qquad I_3=220+0=220.

The paid increment from age 1 to age 2 is $180-100=80$. The case-reserve change is $5-60=-55$, so the incurred increment is $80-55=25$. Directly, $185-160=25$, which checks the reconciliation.

From age 2 to age 3, the paid increment is $220-180=40$, and the case-reserve change is $0-5=-5$. The incurred increment is $40-5=35$. Directly, $220-185=35$, again matching the identity.

The age-2 incurred increment is positive but much smaller than the paid increment because a substantial part of the outstanding case reserve was released. Common mistake: treating the incurred increment as the paid increment plus the new case reserve. The correct adjustment is the change in case reserve, here $5-60$, not the ending reserve of 5.

Worked Example: Comparing One-Period Paid and Incurred Residuals

For a hypothetical development period, paid cumulative amounts are 100, 200, and 300, followed by 140, 300, and 480. Incurred cumulative amounts are 120, 220, and 320, followed by 180, 308, and 520. For each triangle, calculate the volume-weighted factor, the estimate $\hat\sigma_j^2$, and the standardized residuals. Use the variance estimate with denominator $n_j-1$.

Paid triangle. The aggregate factor is $(140+300+480)/(100+200+300)=920/600=1.5333$. The fitted values are $153.333$, $306.667$, and $460$; the residuals are $-13.333$, $-6.667$, and $20$.

The weighted squared residual terms are $13.333^2/100=1.7778$, $6.667^2/200=0.2222$, and $20^2/300=1.3333$. Thus $\hat\sigma_P^2=(1.7778+0.2222+1.3333)/2=1.6667$. The standardization denominators are $\sqrt{1.6667(100)}=12.9099$, $\sqrt{1.6667(200)}=18.2574$, and $\sqrt{1.6667(300)}=22.3607$. The paid standardized residuals are therefore $-1.033$, $-0.365$, and $0.894$, rounded to three decimals.

Incurred triangle. The aggregate factor is $(180+308+520)/(120+220+320)=1008/660=1.5273$. The fitted values are $183.2727$, $336$, and $488.7273$; the residuals are $-3.2727$, $-28$, and $31.2727$.

The weighted squared residual terms are $3.2727^2/120=0.0893$, $28^2/220=3.5636$, and $31.2727^2/320=3.0562$. Thus $\hat\sigma_I^2=(0.0893+3.5636+3.0562)/2=3.3545$. The standardization denominators are $\sqrt{3.3545(120)}=20.064$, $\sqrt{3.3545(220)}=27.166$, and $\sqrt{3.3545(320)}=32.764$. The incurred standardized residuals are $-0.163$, $-1.031$, and $0.955$, rounded to three decimals.

The incurred variance estimate is numerically larger, but it is estimated from incurred amounts and should not be interpreted as a direct proof that the incurred ultimate is less reliable. The standardized patterns also differ: the largest absolute paid residual is in the first origin year, whereas the incurred residuals are more pronounced in the second and third. Common mistake: comparing only raw residuals, or treating a variance estimate from one triangle as interchangeable with the other, without considering scale and residual pattern.

What the Two Views Say About Uncertainty

Paid development is driven by the timing and amount of payments. Incurred development reflects those payments plus changes in case estimates. Incurred data can provide a more developed view of expected ultimate losses at an earlier valuation, but its apparent information advantage depends on the quality and consistency of case reserving. A change in settlement practice, reserving guidelines, or adjuster estimates can alter the incurred pattern without a corresponding change in the underlying payment process.

Under the Mack-style model, the one-period conditional process variance is $\sigma_j^2C_{i,j}$. This is only the variation in the next cumulative amount if the development parameters are treated as known. A reserve prediction that also estimates development factors has additional uncertainty from estimating those factors. A full comparison should use the same valuation date, development ages, tail treatment, and uncertainty convention for both triangles. It should also distinguish process variation from parameter-estimation variation rather than comparing unlike totals.

Worked Example: Comparing One-Step Process Uncertainty

Suppose a hypothetical portfolio has one remaining development period. Treat the fitted factors and variance parameters as known, so this calculation measures process uncertainty only. For paid losses, the current cumulative amount is 400, the factor is 1.25, and $\sigma_P^2=16$. For incurred losses, the current cumulative amount is 460, the factor is 1.10, and $\sigma_I^2=25$. Calculate each predicted terminal amount, process standard deviation, and coefficient of variation.

Paid. The predicted terminal amount is $1.25(400)=500$. The conditional process variance is $16(400)=6{,}400$, so the standard deviation is $\sqrt{6{,}400}=80$. The coefficient of variation is $80/500=0.1600$, or 16.00%.

Incurred. The predicted terminal amount is $1.10(460)=506$. The conditional process variance is $25(460)=11{,}500$, so the standard deviation is $\sqrt{11{,}500}=107.24$. The coefficient of variation is $107.24/506=0.2120$, or 21.20%, rounded to four decimal places for the ratio.

On these stated assumptions, the incurred one-step process uncertainty is larger in both dollars and relative terms. The implied additional incurred development is $506-460=46$, whereas the paid development is $500-400=100$. These are not the same reserve measure: incurred already includes case estimates. Common mistake: interpreting this process-only comparison as a complete comparison of reserve prediction error. It omits uncertainty in estimated factors and does not establish that the incurred method is generally less accurate.

Case-Reserve Volatility and Diagnostic Traps

An incurred residual is affected by the change in case reserves, not just by payments. If case estimates are revised inconsistently across origin years, incurred link ratios can scatter even when paid development is comparatively smooth. A large case-reserve release may also create a negative incurred increment. Neither observation alone proves a data error, but both warrant investigation before relying on a standard Mack variance calculation.

Compare the residuals by development age and origin year, then reconcile unusual incurred cells to paid movements and case-reserve changes. Ask whether a pattern is concentrated in a particular origin year, appears across a development age, or follows a known change in case estimation practice. Also check whether cumulative incurred amounts remain positive and whether negative increments make the selected variance method unsuitable. The Mack variance scaling is not appropriate for a zero or negative cumulative base.

A common exam trap is to conclude that incurred is always more stable because it incorporates case estimates, or that paid is always more reliable because it avoids those estimates. Neither conclusion follows from the definitions. The appropriate conclusion is conditional: inspect the two development patterns, use comparable standardized diagnostics, and identify what source of variation each triangle contains.

Key takeaway: Paid and incurred triangles can support different views of the same ultimate losses, but their uncertainty diagnostics are not interchangeable. Reconcile incurred increments to paid increments and case-reserve changes, compare standardized residual patterns, and state clearly which uncertainty components are included.

Check Your Understanding

For each problem, give the requested result and interpret it in the stated context.

A hypothetical origin year has cumulative paid losses of 250 at age 2 and 310 at age 3. Case reserves are 40 and 15, respectively. Find the age-3 incurred increment and the paid increment. (Answer: incurred increment 35; paid increment 60, in the stated monetary units.)
For one Mack-style development step, $C_{i,j}=240$, $\hat f_j=1.20$, and $\hat\sigma_j^2=9$. Find the predicted next cumulative amount and conditional process standard deviation, treating parameters as known. (Answer: predicted amount 288; standard deviation $\sqrt{2{,}160}=\$46.48$, rounded to cents.)
Three paid cumulative amounts are 80, 120, and 200, and their next-age amounts are 104, 168, and 300. Using the volume-weighted factor and the variance estimator with denominator $n_j-1$, calculate $\hat f_j$ and $\hat\sigma_j^2$. (Answer: $\hat f_j=1.4300$; $\hat\sigma_j^2=1.2200$.)
A one-step paid triangle has current cumulative amount 500 and $\hat\sigma_P^2=4$. An incurred triangle has current cumulative amount 560 and $\hat\sigma_I^2=9$. Calculate their conditional process standard deviations, treating parameters as known, and identify which is larger. (Answer: paid $\sqrt{2{,}000}=\$44.72$; incurred $\sqrt{5{,}040}=\$70.99$; incurred is larger.)
During a development period, paid losses increase by 75 while case outstanding reserves decrease by 90. Find the incurred increment and explain what its sign means. (Answer: $-15$; the incurred amount decreases because the case-reserve release exceeds the paid increase.)

What You'll Learn