Reserves from a growth curve fitted to the incremental claims by maximum likelihood, after Clark (2003): the LDF method, with an ultimate for each origin, or the Cape Cod method, with one expected loss ratio on the premiums. Its standard errors have a process part and a parameter part.
Arguments
- triangle
A triangle of cumulative claims, as for
chain_ladder.- curve
"loglogistic"(the default) or"weibull".The premium of each origin: needed by the Cape Cod method, and used for the loss ratios when given to the LDF method.
- method
"ldf"(the default) or"cape cod".- truncation
The age, in development periods, at which the curve is cut off and the claims are taken to be fully developed:
Inf(the default) for none.- offset
The average date of the claims in an origin period, as a share of a period from its start: 0.5 (the default) for claims spread evenly over it, 0 to measure ages from the start.
Value
An object of class netsimr_reserve: a list with by_origin,
a data frame with each origin's latest claims, share developed, ultimate, IBNR,
standard error, coefficient of variation and its process and parameter parts
(and premium and ultimate loss ratio when there are premiums); total;
parameters, the curve's \(\omega\) and \(\theta\) (and the loss
ratio of the Cape Cod method), \(\sigma^2\) and the degrees of freedom;
growth, the fitted curve as a function of the age; fitted and
residuals, the fitted increments and the Pearson residuals
\((c - \mu) / \sqrt{\sigma^2 \mu}\), as triangles; and covariance.
Details
The share of an origin's ultimate developed by age \(x\) is a growth
curve: the log-logistic \(G(x) = x^\omega / (x^\omega + \theta^\omega)\) or
the Weibull \(G(x) = 1 - e^{-(x / \theta)^\omega}\). The age is measured from
the average date of the claims in an origin period, offset of a period
after its start (a half for claims spread evenly over it), so the increment of
development period \(t\) is expected to be
\(\mu_{it} = U_i (G(t - o) - G(t - 1 - o))\), with \(G = 0\) at ages up to 0.
\(U_i\) is the ultimate of each origin (LDF method) or its premium times one
expected loss ratio (Cape Cod method). The increments are taken to be
over-dispersed Poisson: the parameters maximise
\(\sum c_{it} \log \mu_{it} - \mu_{it}\), so that each \(U_i\) (or the loss
ratio) has a closed form given \(\omega\) and \(\theta\), which are found
numerically. The scale \(\sigma^2\) is the sum of the squared Pearson
residuals over the degrees of freedom (the cells less the parameters).
The reserve of an origin is \(U_i (G(T - o) - G(a_i - o))\), from its latest
age \(a_i\) to the truncation age \(T\) (infinite by default; the growth
curves can have long tails, so a finite truncation is often used with the LDF
method). Its process variance is \(\sigma^2\) times the reserve; its parameter
variance comes from the covariance of the parameters, \(\sigma^2\) times the
inverse of minus the second derivatives of the log-likelihood, by the delta
method. The results agree with the ChainLadder package's ClarkLDF() and
ClarkCapeCod().
References
Clark, D. R. (2003). LDF curve-fitting and stochastic reserving: a maximum likelihood approach. Casualty Actuarial Society Forum, Fall 2003, 41-92.
See also
chain_ladder and cape_cod.
Examples
clark <- clark_growth(raa_triangle)
clark
#> Clark LDF method, log-logistic curve
#>
#> origin latest developed ultimate ibnr se cv process_se parameter_se
#> 1981 18,834 0.822 22,906 4,072 2,792 0.686 1,950 1,998
#> 1982 16,704 0.799 20,899 4,195 2,834 0.676 1,980 2,028
#> 1983 23,466 0.771 30,440 6,974 4,049 0.581 2,553 3,143
#> 1984 27,067 0.735 36,821 9,754 5,143 0.527 3,019 4,164
#> 1985 26,180 0.689 37,994 11,814 5,855 0.496 3,322 4,821
#> 1986 15,852 0.628 25,225 9,373 4,879 0.521 2,959 3,879
#> 1987 12,314 0.547 22,527 10,213 5,202 0.509 3,089 4,186
#> 1988 13,112 0.434 30,219 17,107 7,570 0.443 3,998 6,429
#> 1989 5,395 0.278 19,397 14,002 7,503 0.536 3,617 6,574
#> 1990 2,063 0.081 25,567 23,504 17,196 0.732 4,686 16,545
#> Total 160,987 0.592 271,994 111,007 36,075 0.325 10,184 34,608
clark$parameters
#> omega theta sigma2 df
#> 1.346491 3.045890 934.249344 43.000000
# the Cape Cod method with a Weibull curve
premium <- c(50000, 52000, 54000, 56000, 58000, 60000, 62000, 64000, 66000, 68000)
clark_growth(raa_triangle, curve = "weibull", premium = premium, method = "cape cod")$total
#> latest developed ultimate ibnr
#> 1.609870e+05 6.742494e-01 2.387648e+05 7.777777e+04
#> se cv process_se parameter_se
#> 1.534206e+04 1.972550e-01 8.201597e+03 1.296582e+04
#> premium ultimate_loss_ratio
#> 5.900000e+05 4.046860e-01
