FORECASTING RUNOFF TRIANGLES

Piet de Jong*

ABSTRACT
This paper deals with the methodology of liability forecasting using the runoff triangle data. Techniques are based on time series models and methods that facilitate the calculation of forecast distributions and the assessment of model t. The models deal with correlation within triangles. Correlations are critical to proper reserving. The output of the methodology is the complete shape of the liability distribution. Methods are applied to a well-known runoff triangle and results compared to those from previous studies.

1. INTRODUCTION
Claims or loss reserving in casualty insurance is often based on runoff triangles. An example runoff triangle is displayed in Table 1. Outstanding liabilities correspond to the lower unlled portion of the rectangle that have to be forecast. This article deals with models and methods for performing the prediction. The models proposed in this article are extensions of Hertig (1985), and the methods are based on modern time series forecasting. There is an extensive literature on runoff triangle analysis: see, for example, the bibliography in England and Verrall (2002) or the book by Taylor (2000). Articles of particular interest include Zehnwirth (1985), Wright (1990), Mack (1993, 1994b), Verrall (1990), Goovaerts and Redant (1999), and Barnett and Zehnwirth (2000). References that take a time series approach include De Jong and Zehnwirth (1983), Kremer (1984), and Verrall (1989a, b). The approach in this article departs from the previous literature in that it implements difference equation time series models as discussed in, for example, Harvey (1989). These models are stated in terms of levels, trends, and correlations. This compares with the regression methods (such as generalized linear modeling; Taylor 2000) that forecast by extrapolating tted surfaces in terms of explanatory variables. Extrapolations from regression ts (as opposed to difference equation ts) often deliver inferior time series forecasts and forecast intervals. There are many advantages to a difference equation approach. First, there is a rich and widely studied range of practical models. Second, there is no need to rethink or reinvent optimal forecasting formulas and algorithms. Third, diagnostics are readily available. Fourth, the actuary is free to tackle actuarially relevant tasks including the selection of appropriate models and uncovering and dealing with features such as correlations between accident and calendar years or between different triangles. Thus the actuary need not spend time on estimation, forecasting, and diagnostic issues that have already been resolved. The layout of this article is as follows. The next section introduces notation and summarizes the proposed approach. Section 3 sets out Hertigs model and generalizations that are the modeling basis of the present approach. Section 4 discusses the tting and assessment of the proposed models and forecasting implementation. Section 5 deals with a case study involving the development correlation model. An appendix discusses state space forms.

* Piet de Jong, BEc, PhD, is a Professor in the Department of Actuarial Studies, Macquarie University, New South Wales 2109, Australia, [email protected].

28

FORECASTING RUNOFF TRIANGLES

29

2. LOG-LINK RATIOS
A runoff triangle contains cumulative liabilities with respect to accident years i and development years j. Entries are denoted cij. An example triangle displayed in Table 1 relates to Automatic Facultative General Liability (excluding asbestos and environmental) from the Historical Loss Development study. This triangle, called the AFG data, was considered by both Mack (1994a) and England and Verrall (2002) and is used here to illustrate methods and compare the present papers methods with those in previous studies. Entries in each row generally increase with j, indicating that as time progresses, incurred liabilities with respect to an accident year increase. Each calendar year leads to an oblique diagonal of observations. In Table 1 there are n 10 calendar years. Following Hertig (1985), the models of this article are stated in terms of the log-link ratios
ij

ln

cij ci, j
1

1, . . . , n, j

1, . . . , n

1,

(2.1)

with i0 ln(ci0). Thus ij, j 0, is the continuously compounded growth in accident years i cumulative in development year j. In terms of the log-link ratios the future rate of growth in cumulatives along any row of the runoff triangle through to development year n 1 is gi ln ci,n ci,n
1 i,n i 1 i i,n 1

, i

2, . . . , n.

(2.2)

Further, ci,n 1 ci,n i ci,n i (e gi 1) is the further increase in the liability with respect to development year i. The total future liability with respect to all accident years and through to development year n 1 is
n i 2 n

(ci,n

ci,n i)

i 2

ci,n

(e gi

1).

(2.3)

Forecasting the liability (2.3) requires forecasts of g (g2, . . . , gn) or, from expression (2.2), the future log-link ratios ij, i j n. The approach to forecasting the liability (2.3) advocated in this article consists of the following steps: 1. Modeling and tting. The log-link ratios ij are modeled, and the chosen model is tted to the runoff triangle data. Models described in this article are time series models, while tting is on the basis of maximum likelihood with the Kalman lter used to evaluate the likelihood. 2. Model assessment. The model is assessed using diagnostics. Assessment and diagnostics may suggest areas of model inadequacy and appropriate extensions. Assessment is important for minimizing the chance of model error. Diagnostics are generated using the Kalman lter (De Jong and Penzer 1998). Barnett and Zehnwirth (2000) have stressed the importance of model assessment generally.
Table 1 AFG DataCumulative Incurred Claim Amounts
Accident Year i 1 2 3 4 5 6 7 8 9 10 Development Year j 0 5,012 106 3,410 5,655 1,092 1,513 557 1,351 3,133 2,063 1 8,269 4,285 8,992 11,555 9,565 6,445 4,020 6,947 5,395 2 10,907 5,396 13,873 15,766 15,836 11,702 10,946 13,112 3 11,805 10,666 16,141 21,266 22,169 12,935 12,314 4 13,539 13,782 18,735 23,425 25,955 15,852 5 16,181 15,599 22,214 26,083 26,180 6 18,009 15,496 22,863 27,067 7 18,608 16,169 23,466 8 18,662 16,704 9 18,834

30

NORTH AMERICAN ACTUARIAL JOURNAL, VOLUME 10, NUMBER 2

3. Forecast future log-link ratios. Future log-link ratios ij, i j n are forecast using the Kalman lter based on the tted model. Forecast error variances and covariances also are derived. i,n i 1 i,n 1 are derived, i 4. Derivation of future accident year growth rates. The i g 2, . . . , n as well as the associated covariance matrix . 5. Simulation of the liability distribution. Repeated draws are made from the multivariate distribution with means i and covariance matrix . Each draw is combined as in equation (2.3) to derive the g simulated forecast liability distribution. The simulated liability distribution incorporates both process and estimation error and provides estimated percentiles of the liability distribution. Details of each step are described in subsequent sections. All the steps have been implemented in an Excel environment, superimposed on the detailed computer algorithms. Figure 1 displays the estimated liability distribution derived from the AFG data using the above steps. Subsequent sections discuss these steps in detail. The estimated liability distribution forms the basis for inferences about the mean, standard deviation, and percentiles. For example, recent Australian legislation has mandated the 75th percentile as an appropriate level of reserves.

3. HERTIGS MODEL

AND

EXTENSIONS

Hertig (1985) introduced a simple yet useful model for the growth rates ij. The model states (see also Murphy 1993 and Taylor 2000) that the growth rates are uncorrelated with means and variances depending only on the development year:
ij j

hjij,

1, . . . n, j

0, . . . , n

1,

(3.1)

where ij (0, 2) and h0 1. Thus the mean and standard deviation of ij are j and hj , respectively, where is the standard deviation of i0. A seemingly minor addition to Hertig (1985) is the inclusion in the model of i0 ln(ci0). This addition turns out to be practically important. Despite this addition we will call equation (4) Hertigs model. Trends in claims over accident years are allowed for with the Hertigs model since each accident years development starts off from from the relevant ci0. Hence a high or low value in ci0 automatically shifts up or down the subsequent development prole for that accident year. Thus the assumption that the i0 all have the same mean 0 is of no import from the forecasting point of view since the forecast liability for each accident year takes off from ci,n i, the latest observed cumulative for that accident year.
Figure 1 Estimated Histogram of Forecast Incurred Liabilities for AFG Data

$0

$50

$100

$150

$200
Thousands

FORECASTING RUNOFF TRIANGLES

31

A key feature of Hertigs model is that the growth rates ij are uncorrelated across both accident years i and development years j. The next three subsections relax this assumption.

3.1 Development Correlation Model

A useful extension of Hertigs model (3.1) is to allow the ij to be correlated across development years j for given accident years i. This is achieved with the following model:
ij j

hj(ij

j i, j 1 ij

), j

1, . . . , n
i, j 1

1, i

1, . . . , n.
i0

(3.2) and
i1

Thus is

models the correlation between

and

. For example, the correlation between

2

cov( i0, i1) var( i0)var( i1)

2 2 1

h1 h (1

1 2 1

2 1

Similar expressions apply for higher-order correlations. Case studies, such as in Section 5, suggest that development correlation between i0 and i1 is often important, but higher-order development year correlations, such as between i1 and i2, often can be ignored.

3.2 Accident Correlation Model

The accident correlation model formalizes the notion that for forecasting, more weight should be given to the more recent data (Taylor 2000, p. 50). It does so by imposing random walk behavior, and hence correlation, across accident years:
ij ij

hjij,

i 1, j

ij

j ij

,i

1, . . . , n, j

0, . . . , n

1.

(3.3)

This extends Hertigs model (3.1) by allowing the mean ij within any development year to evolve slowly over accident years. Hertigs model (3.1) is the special case where j 0 for all j, implying each development years mean is constant over the accident years: i 1, j ij. The accident year correlation model implies that, when forecasting future development ratios, more weight is given to more recent observed ratios as opposed to those in the more remote past. For example, the estimate of the log-link ratios in the upcoming calendar year is a geometrically declining average of the previous ratios falling in the same development year. The rate of decline in each development year is controlled by the signal-to-noise ratio j/hj. A ratio near zero implies a very low rate of decline. Weighting loss ratios has two other consequences, both impacting on the variability associated with the forecasts. First, basing the ratio estimates on the more recent evidence implies there is less certainty about them because, in effect, fewer observations are used to make the forecasts. Second, since the ratios evolve over time, evolution is likely to continue into the future, also implying increased uncertainty in the estimates.

3.3 Calendar Correlation Model

A model yielding correlation across calendar years is
ij j

hj(

i j

ij),

i j 1

i j

i j

, j

1, . . . , n

1, i

1, . . . , n,

(3.4)

where ij and i j are uncorrelated mean zero, variance 2 noise terms. The calendar year effects k thus are assumed to evolve as a random walk in calendar time, and each k serves to increase or decrease all the log-link ratios falling in calendar year k. The effect of k on a particular log-link ratio is scaled by hj, and hence the effect is assumed proportional to the standard deviation associated with the loglink ratio. Hertigs model (3.1) results when 0.

32

NORTH AMERICAN ACTUARIAL JOURNAL, VOLUME 10, NUMBER 2

3.4 State Space Form

The models discussed in the above subsections can be cast into the state space form. State space forms permit the immediate implementation of the forecasting, estimation, and diagnostic checking technology centered around the Kalman lter. The generic state space form is yt Xt Zt
t

Gtt,

t 1

Wt

Tt

Htt, t

1, . . . , n,

(3.5)

where yt is the vector of observations at time t. In this article yt is the vector of log-link ratios observed at t, yt ( 1,t 1, 2,t 2, . . . , 0t) , and is diagonal t of the runoff triangle. The Appendix shows how each of the above models can be cast into the state space form and the interpretation accorded to the parameters, appearing in the right-hand side of equation (3.5).

4. FITTING, ASSESSMENT,

AND

FORECASTING

This section discusses the detailed steps in the tting, assessment, and forecasting associated with the models described in the previous section.

4.1 Model Fitting

Hertigs model (3.1) is tted to a runoff triangle using the formulas j 1 n j
n j i 1 ij, n j

1 n j

(
i 1

ij

j)2, j

0, . . . , n

1.

(4.1)

In turn 0 and hj j/ 0, j 0, . . . , n 1. These estimates are discussed in Hertig (1985) and Taylor (2000, Section 7.3) and illustrated with respect to the AFG data in Section 5. The correlation models of Sections 3.1, 3.2, and 3.3 can be tted using likelihood maximization with the Kalman lter employed to evaluate the likelihood. There is no need to tailor estimation formulas or software to the specic extension if the model is cast in the general state space form amenable to the general form of the Kalman lter. The appendix shows how the correlation models can be written in the state space form. The Kalman lter equations are displayed in Anderson and Moore (1979) or Harvey (1989), while associated smoothing and diagnostic algorithms are discussed in De Jong (1989) and De Jong and Penzer (1998). These smoothing and diagnostic algorithms facilitate signal extraction and model assessment.

4.2 Model Assessment

Hertigs model (3.1) is assessed using the standardized log-link ratios zij
ij

j ,

1, . . . , n, j

0, . . . , n

i.

(4.2)

If Hertigs model (3.1) holds, then the z-scores are approximately normally distributed with mean zero and unit standard deviation. Large zij suggests the model does not t well at the given accident year i and development year j. The z-scores can be plotted against development j, accident i, or calendar i j to reveal any structure. If Hertigs model is appropriate, there should be no structure. The standardized residuals (4.2) form a basis for assessing correlation in runoff triangles. For example, development correlation between two development years is computed from the zij corresponding to the two years, with the accident years i forming the cases. Similarly, correlation between accident years is computed by taking the zij for the two accident years and letting the development years j be the cases. Calendar year correlation can be dealt with similarly. In the development correlation case, the correlation between the zij is the same as the correlation between the corresponding ij. However, this property does not hold for the accident or calendar year correlation computation.

FORECASTING RUNOFF TRIANGLES

33

Residuals also are available for the correlation models. These are calculated with the Kalman lter as discussed in De Jong and Penzer (1998).

4.3 Forecasting with the Fitted Model

The forecasting formulas for Hertigs model (3.1) are well known (Taylor 2000). If the then the linear predictor of gi and associated prediction error variance are
n i 1 n 1 j

are known, (4.3)

2 i

2 n i 1

2 n 1

, i

2, . . . , n.

In practice the j and j in these expressions are replaced by the estimates j and j given in equations (4.1). Using the j ij rather than the j in equations (4.3) introduces estimation correlation between the i g i,n
i 1

i,n

i 1

n 1, i

2, . . . , n,

(4.4)

since the same estimates j are used in different i. Expressions for the elements of the covariance g matrix of g (2, . . . , n) are given in Taylor (2000, Section 7.3). g g For the more general correlation models, the mean vector and covariance matrix of g, and indeed all the ij, can be computed readily with the Kalman lter, again provided the model is in the state space form. In this setting is interpreted as the minimum mean square error linear predictor of g, g and is the associated error covariance matrix. Given and the covariance matrix , and assuming normality, repeated draws can be made from the g (, ) distribution, with each draw combined as in equation (2.3) to yield an estimate of the outstandg ing liability. The convenient distribution to work with is the multivariate normal, although its applicability would have to be assessed. If the normal assumption is appropriate, then it follows (Aitchison and Brown 1957) that the conditional expected value of ci,n 1 and associated coefcient of variation are
g ci,n i e i
2 i/2

2 i

1,

2, . . . , n,

respectively. These expressions in turn lead to the conditional expected value of each accident years liability ci,n 1 ci,n i and associated coefcient of variation. In these expressions the 2 dened in i equations (4.3) are assumed known. In practice they are replaced with estimates.

4.4 Using the State Space Form

Given the state space form, the Kalman lter and associated algorithms simplify estimation, forecasting, and assessment into estimates of future yt (and hence ij and gi) and associated variances and covariances (for example, ). A user of the technology need not concern him- or herself with the details of these algorithms.

5. CASE STUDY: DEVELOPMENT CORRELATION

This section implements the steps discussed in previous sections to the AFG data of Table 1 to arrive at the estimated liability distribution displayed in Figure 1. The bottom two rows of Table 2 display the j and standard deviations j dened in formulas (4.1). As expected, average growth rates tend to be high in the early development years and gradually fall off. The z-scores (4.2) are displayed in the body of Table 2. Boldface numbers are those more than 1.645 standard deviations away from zero. Five out of the 55 numbers are outside the 10% normal bounds, indicating reasonable conformance to normality. The two left panels of Figure 2 display ln( j ) and ln( j ) as functions of development year j as well as least squares straight line ts. Both j and j decline geometrically with development year j. Table 2 and Figure 2 contain compelling evidence to suggest that Hertigs model (3.1) is not appro-

34

NORTH AMERICAN ACTUARIAL JOURNAL, VOLUME 10, NUMBER 2

Table 2 z-Scores and Hertigs Model Estimates for the AFG Data
Accident Year i 1 2 3 4 5 6 7 8 9 10 j j Development Year j 0 1.04 2.39 0.70 1.15 0.31 0.02 0.91 0.12 0.62 0.25 7.35 1.12 1 1.06 2.27 0.57 0.83 0.68 0.07 0.48 0.12 1.01 1.52 0.96 2 0.94 1.14 0.28 0.80 0.02 0.42 2.14 0.58 3 0.87 2.17 0.51 0.24 0.42 0.77 0.68 4 0.58 1.76 0.35 1.38 0.18 0.72 5 1.00 0.10 0.87 0.17 1.79 6 1.59 1.17 0.31 0.11 7 0.15 1.29 1.14 8 1.00 1.00 9 0.00

0.50 0.24

0.25 0.20

0.17 0.05

0.12 0.06

0.04 0.04

0.03 0.01

0.02 0.01

0.01

Note: Boldface numbers indicate signicance at the 10% level.

priate. In particular, the top right panel of Figure 2 displays a scatter plot of the z-scores corresponding to development years 0 and 1. The z-scores are almost perfectly negatively correlated: a high value in development year 0 is almost invariably followed by a low value the following development year. Ignoring the correlation between i0 and i1 can lead to sizable forecast errors. Since the cumulative ci0 for accident year i 10 is above the average, it is highly likely that the increment in development year 1 will be below the average. This has major implications for the forecast since a large part of the total liability is with respect to the nal accident year n 10. It may be argued that negative correlation always will be present since i0 ln(ci0) and i1 ln(ci1/ ci1). However, in our experience not all runoff triangles display the correlation. Further, the existing

Figure 2 Statistical Features of the AFG Runofff Triangle

mean
2
log(mu)

z-scores

0 -2 -4 0 1 2 3 4 5 6 7 8 9

year 1 value

1 0 -1 -2 -2 -1 0 1 2

std dev
0
year 1 value log(sigma)

cumulatives
12 9 6 3 0

-2 -4 0 1 2 3 4 5 6 7 8 9

development year

development year 0 value

FORECASTING RUNOFF TRIANGLES

35

literature and methods ignore the correlation. For example, the often used chain-ladder approach forecasts cn1 by cn0( i ci0/ i ci1). This formula does not adjust the forecast according to the relative size of cn0. Hertig (1985) does not explicitly model ci0, and hence the possibility of correlation is ignored. Taylor (2000), illustrating the use of Hertigs model, also ignores the correlation. A notable exception is Barnett and Zehnwirth (2000), who argue that the chain-ladder method is misleading because it does not allow for an intercept term in the simple regression corresponding to the bottom right panel in Figure 2, which plots ci0 versus ci1. However, in this plot the correlation is obscured and may be rationalized on the misleading basis that a large value of ci0 will lead, other things equal, to a large value for ci1.

5.1 Fitting and Forecasting

We now t the development correlation model (3.2) to the AFG runoff data in Table 1 and forecast the liabilities. Both tting and forecasting employ the Kalman lter. To t the model it is assumed that only the correlation between development year 0 and 1 is signicant, and hence equation (3.2) applies with j 0 except for j 1. The bottom left panel of Figure 2 suggests ln( j) a bj and hence 1, hj ea
1

j 1
b( j 2) 2

0, 1, 2, . . . , n 1.

, j j

Thus the parameters to be tted are , h1 (equivalently 1), , a, b, and 0, . . . , n 1. When n 10, as in the case of the AFG data, this gives 15 unknown parameters compared to 55 data points. This is possibly an excessive number of unknown parameters compared to data points. The n 10 mean parameters 0, . . . , 1 can possibly also be structured as linear in the logs, as suggested in the top left panel of Figure 2. Table 3 reports moment- and normal-based maximum likelihood estimates of development correlation model parameters using the AGF data in Table 1. The moment estimates are where the estimate of is derived from the empirical correlation between i0 and i1, and a and b are derived from a least squares regression j on j 2 for j 2, . . . , n 2. Maximum likelihood values are derived using the Kalman lter to evaluate the likelihood. In the table is the value of the negative of the log-likelihood. Table 1 indicates there are no material differences in the two sets of estimates. Given the parameter estimates displayed in Table 3, the Kalman lter was used to derive generalized least squares estimates of the j and to compute the forecasts i as well as the associated error covarg iance matrix. Results are displayed in Table 4 with estimated standard errors in parentheses. The standard errors factor in both process and estimation uncertainty of the j, but not the uncertainty of the estimates displayed in Table 3. Simulations from the distribution, including covariances, yield the forecast liability distribution displayed in Figure 1. The development correlation model was assessed by computing the linear predictors of the disturbances ij in equation (3.2) given the runoff triangle data, and where unknown parameters are replaced by the Table 3 estimates. Computations are performed with the smoothing lter companion to the Kalman lter (De Jong 1989). These residuals indicate that the outcome in accident year 2, develop-

Table 3 Parameter Estimates for Correlation Model and Table 1 Data

Parameter Moment estimates Maximum likelihood estimates h1 0.21 0.21 a 1.52 1.43 b 0.54 0.53 3.93 3.74 1.12 1.12 152.07 152.72

36

NORTH AMERICAN ACTUARIAL JOURNAL, VOLUME 10, NUMBER 2

Table 4 Liability Forecasts for the AFG Data

Development Year 2 3 4 5 6 7 8 9 10 Total Forecast Growth 0.01 0.03 0.06 0.10 0.22 0.39 0.64 1.14 2.40 (0.01) (0.01) (0.02) (0.04) (0.07) (0.12) (0.20) (0.35) (0.45) Forecast Liability 154 642 1,701 2,843 3,948 5,941 12,243 12,475 22,957 62,982 (146) (375) (753) (1,271) (1,462) (2,290) (5,463) (6,747) (11,551) (16,260) Chain Ladder 154 617 1,636 2,747 3,649 5,435 10,907 10,650 16,339 52,135 (206) (623) (753) (1,456) (2,007) (2,228) (5,344) (6,284) (24,509) (26,909) Poisson Model 243 885 2,033 3,582 3,849 5,393 11,091 10,568 17,654 55,297 (486) (984) (1,589) (2,216) (2,301) (2,873) (4,686) (5,563) (12,801) (17,357)

ment year 1, previously agged as extreme, is in fact not extreme in the context of the development correlation model. Also the correlation between the development year 0 and 1 residuals is now an insignicant 0.19. The model has thus both explained the correlation and exploited it to arrive at a defensible loss forecast distribution.

5.2 Comparison with Results from Previous Studies

Table 4 compares our results with those of previous studies by Mack (1994a, p. 130) and England and Verrall (2002, p. 45). The Mack results are in the columns labeled Chain Ladder and are based on a chain ladder analysis and using the Mack (1993) formula for the standard deviation. The England and Verrall results are based on an overdispersed Poisson model with a Hoerl curve. Table 4 displays substantial differences between forecasts from the different methods. The chain ladder method estimates the standard deviation of the prediction error for accident year 2 of 206, while the Poisson model gives 486. The data directly relevant to this estimate are the 0.92% or 172 growth in liabilities in accident year 1 between development years 8 and 9. Using the chain ladder estimates and the arguably conservative normal approximation yields the conclusion that accident year 2 liabilities may grow in excess of 154 1.65 206 500 with a probability of around 5%, while the Poisson approach would give an even more surprising 5% limit of 1045. These conclusions seem at odds with the data. Further, consider the standard deviation of total liabilities under the chain ladder method. The standard deviation 26,909 is only marginally higher than the standard deviation associated with the estimated accident year 10 liability. This has the counterintuitive implication that the individual accident year liability estimates are at most marginally positively correlated and most likely negatively correlated. The methods of this article focus on the whole distribution of forecast liabilities, not just the rst two moments. Some of the differences between the results displayed in Table 4 can be explained in terms of the distribution. For example, the mode of the forecast distribution displayed in Figure 1 is close to the means in Table 4 reported for the chain ladder and Poisson method. The distribution is slightly skewed to the right, forcing up the actual mean. The standard deviation seems similarly affected. Hence differences between the methods may arise from an inability of the chain ladder and Poisson methods rst two moments to address the skew in the distribution effectively. Having the whole distribution allows one to make considered statements about, for example, the upper percentiles of the distribution. To illustrate, the development correlation model gives $73,653, $83,860, and $92,609 as the 75th, 90th, and 95th percentiles of the liability distribution for the AFG data. Such upper percentiles recently have been suggested as an appropriate measure of risk.

FORECASTING RUNOFF TRIANGLES

37

APPENDIX STATE SPACE FORMS

This appendix indicates how the three discussed models can be cast into the state space form.

A.1. DEVELOPMENT CORRELATION MODEL

For t 1, . . . , n put yt ( 1,t 1, . . . , t0) and t (1,t 1, . . . , t0) . Thus yt is diagonal t of the runoff triangle of the ijs. Put Xt 0, Zt I and Gt diag(ht 1, . . . , h1, 1). These matrices are of dimension t n, t t, and t t respectively. In the second equation in equations (3.5), put Wt as the last t 1 rows of the row-permuted identity matrix of order n, Tt 0 of dimension (t 1) t, and Ht a (t 1) t matrix of zeros except in position (t, t), where it is h1 . With this parameterization ( 0, . . . , n 1) and t ( t 1, . . . , 2, 1 h1 t 1,0, 0) . Hertigs model is the special case 0.

A.2 ACCIDENT CORRELATION MODEL

Dene yt ( 1,t 1, . . . , t0) t ( 1,t 1, . . . , t0) , t (1,t 1, . . . , t0, 1,t 1, . . . , t0) , and ( 10, . . . , 1,n 1) . Dene Xt as a t n matrix of zeros, Zt as an identity matrix of order t, and put Gt as a t 2t matrix of zeros except on the rst complete diagonal, where it contains ht 1, . . . , h1, 1. Further, put Wt as a (t 1) n matrix of zeros except in position (1, t 1), where it is 1, Tt as an identity matrix of order t with a row of zeros on top, and Ht as a (t 1) 2t matrix of zeros except on the last complete diagonal, where it contains 0, t 1, . . . , 0. Hertigs model is the special case where each j 0. The log-link ratio random walk model in the accident year direction is attained by letting each hj/ j .

A.3 CALENDAR CORRELATION MODEL

Dene yt as above, ( 0, . . . , n 1) , t ( t 1, . . . , 0, t) , and t (1,t 1, . . . , t0, t) . Further put Xt as a t n matrix of zeros, Zt as an identity matrix of order t with an extra nal column containing (ht 1, . . . , h1, 1) , and Gt a t (t 1) matrix of zeros except on the main diagonal, where it contains (ht 1, . . . , h1, 1). Put Wt as the last t 1 rows of the row-permuted identity matrix of order n augmented with a nal column of zeros, Tt (t 1) (t 1) matrix of zeros except in position (t 1, t 1), where it is 1, and Ht a (t 1) (t 1) matrix of zeros except in position (t 1, t 1), where it is . REFERENCES
AITCHISON, J., AND J. A. C. BROWN. 1957. The Lognormal Distribution, with Special Reference to Its Uses in Economics. Cambridge: Cambridge University Press. ANDERSON, B. D. O., AND J. B. MOORE. 1979. Optimal Filtering. Englewood Cliffs, NJ: Prentice-Hall. BARNETT, G., AND B. ZEHNWIRTH. 2000. Best Estimates for Reserves. Proceedings of the Casualty Actuarial Society 87: 16667. DE JONG, P. 1989. Smoothing and Interpolation with the State-Space Model. Journal of the American Statistical Association 84: 1085 88. DE JONG, P., AND J. R. PENZER. 1998. Diagnosing Shocks in Time Series. Journal of the American Statistical Association 93: 796806. DE JONG, P., AND B. ZEHNWIRTH. 1983. Claims Reserving, State-Space Models and the Kalman Filter. Journal of the Institute of Actuaries 110: 15781. ENGLAND, P., AND R. VERRALL. 2002. Stochastic Claims Reserving in General Insurance. Journal of the Institute of Actuaries 129: 1 76. GOOVAERTS, M., AND R. REDANT. 1999. On the Distribution of IBNR-Reserves. Insurance: Mathematics and Economics 25: 110. HARVEY, A. C. 1989. Forecasting, Structural Time Series Models and the Kalman Filter. Cambridge: Cambridge University Press. HERTIG, J. 1985. A Statistical Approach to IBNR-Reserves in Marine Reinsurance. Astin Bulletin 15(2): 17183. KREMER, E. 1984. A Class of Autoregressive Models for Predicting the Final Claim Amount. Insurance: Mathematics and Economics 3: 11119.

38

NORTH AMERICAN ACTUARIAL JOURNAL, VOLUME 10, NUMBER 2

MACK, T. 1993. Distribution-Free Calculation of the Standard Error of Chain Ladder Reserve Estimates. ASTIN Bulletin 23(2): 213 25. . 1994a. Measuring the Variability of Chain Ladder Reserve Estimates. In Proceedings of the Casualty Actuarial Society Spring Forum, 10182. . 1994b. Which Stochastic Model Is Underlying the Chain Ladder Method? Insurance, Mathematics and Economics 15(2/3): 13338. MURPHY, D. 1993. Unbiased Loss Development Factors. In Proceedings of the Casualty Actuarial Society Meeting, May 1993, pp. 183 246. TAYLOR, G. 2000. Loss Reserving: An Actuarial Perspective. Boston: Kluwer. VERRALL, R. 1989a. Modelling Claims Run-Off Triangles with Two-Dimensional Time Series. Scandanavian Actuarial Journal: 129 38. . 1989b. A State Space Representation of the Chain Ladder Linear Model. Journal of the Institute of Actuaries 116: 589610. . 1990. Bayes and Empirical Bayes Estimation for the Chain Ladder Model. ASTIN Bulletin 20: 21743. WRIGHT, T. S. 1990. A Stochastic Method for Claims Reserving in General Insurance. Journal of the Institute of Actuaries 117: 677 731. ZEHNWIRTH, B. 1985. Interactive Claims Reserving Forecasting System. St. Kilda, Australia: Insureware P/L.

Discussions on this paper can be submitted until October 1, 2006. The author reserves the right to reply to any discussion. Please see the Submission Guidelines for Authors on the inside back cover for instructions on the submission of discussions.