10.2 Stochastic Models for the Chainladder

Stochastic models that have the same estimate of Unpaid losses as the Chainladder

For each model, we can calculate the MSE of prediction and therefore a prediction interval

Proposition 10.1 Prediction variance = process variance + estimation variance

Proof. \[\begin{array} \text{MSEP} &= &\mathrm{E}[(y - \hat{y})^2] \\ &= &\mathrm{E}[((y-\mathrm{E}[y]) - (\hat{y} - \mathrm{E}[y]))^2] \\ &\approx &\mathrm{E}[((y-\mathrm{E}[y]) - (\hat{y} - \mathrm{E}[\hat{y}]))^2] \\ &= &\mathrm{E}[(y - \mathrm{E}[y])^2] \\ & &-2\mathrm{E}[(y-\mathrm{E}[y])(\hat{y} - \mathrm{E}[\hat{y}])] \\ & &+ \mathrm{E}[(\hat{y} - \mathrm{E}[\hat{y}])^2]\\ &= &\underbrace{\mathrm{E}[(y - \mathrm{E}[y])^2]}_{\text{Process Variance}}+ \underbrace{\mathrm{E}[(\hat{y} - \mathrm{E}[\hat{y}])^2]}_{\text{Estimation Variance}}\\ \end{array}\]

Remark.

  • Assume future observations are independent of past observations and the -2 term above goes to 0

  • S.e. = \(\sqrt{\text{Estimation Variance}}\)

  • Prediction error includes both the error in estimating our parameters and the process variance (inherent variability in the data being forecast)

  • Both bootstrap and MCMC allows us to calculate the prediction error

10.2.1 Mack-1993 (Non-parametric)

Only the first 2 moments of cumulative claims are specified

\[\begin{equation} \mathrm{E}[D_{ij}] = \lambda_j D_{ij-1} \tag{10.1} \end{equation}\] \[\begin{equation} \mathrm{Var}(D_{ij}) = \sigma^2_j D_{ij-1} \tag{10.2} \end{equation}\]

Remark.

  • Mean is the same as the Chainladder

  • Variance is \(\propto\) claims reported to date \(D_{ij-1}\)

    \(\sigma^2_j\) has to be estimated separately from the development factors

  • The simplicity of this allows the parameter estimate and prediction errors to be obtained from a spreadsheet

  • Downside is that without specifying a distribution we don’t get a predictive distribution

10.2.2 Over-dispersed Poisson and Negative Binomial

Separate stream of research that focused on the use of GLM

Incremental claims distribution OD Poisson:

\[\begin{equation} \begin{split} & c_{ij} \mid c, \alpha, \beta, \varphi \sim ODP(m_{ij}, \varphi m_{ij}) \\ & \mathrm{E}[c_{ij}] = m_{ij} \\ & \mathrm{Var}(c_{ij}) = \varphi m_{ij} \\ & \ln(m_{ij}) = c + \alpha_i + \beta_j \: ; \: \alpha_1 = \beta_1 = 0 \end{split} \tag{10.3} \end{equation}\]

Remark.

  • Model result is the same as Chainladder

  • \(m_{ij}\) mathematically the same as the one defined in Shapland, we can not include the \(c\) term and just use \(\alpha_i\)’s as the individual level parameters instead of having \(\alpha_1 = 0\) and use the \(alpha_i\)’s as adjustments to the level parameter constant \(c\)

  • This allows for calculating prediction error

Alternative way of writing ODP

\[\begin{equation} \begin{split} & c_{ij} \mid x,y, \varphi \sim ODP(x_i y_j, \varphi x_i y_j) \\ & \sum_{k=1}^n y_k = 1 \\ & x = \{x_1, x_2,...,x_n\} \:\: \text{row parameters}\\ &x_i = \mathrm{E}[D_{in}] \:\: \text{Expected ultimate cumulative losses up to }n\\ & y = \{y_1, y_2,...,y_n\} \:\: \text{col parameters}\\ & y_i \:\: \text{Proportions of ultimate losses that emerge in each development year} \end{split} \tag{10.4} \end{equation}\]

Recall:

\[\begin{equation} \begin{split} &X \sim \text{Poisson}(\mu) \\ &Y = \varphi X \sim ODP(\varphi \mu, \varphi^2 \mu)\\ \end{split} \tag{10.5} \end{equation}\]
  • Where \(\varphi\) is typically > 1 \(\therefore\) over-dispersed

  • This can be extend to other distribution beyond Poisson

    e.g. see over-dispersed negative binomial below

We can get the same predictive distribution (as ODP) with ODNB

  • ODNB also make the connection between ODP and Chainladder more apparent

Incremental claims distribution for OD Negative Binomial:

\[\begin{equation} \begin{split} & c_{ij} \mid D_{ij-1}, \lambda_j, \varphi \sim ODNB \\ & \mathrm{E}[c_{ij}] = (\lambda_j - 1)D_{ij-1} \\ & \mathrm{Var}(c_{ij}) = \varphi \lambda_j \mathrm{E}[c_{ij}] \\ \end{split} \tag{10.6} \end{equation}\]

Cumulative claims distribution for OD Negative Binomial:

\[\begin{equation} \begin{split} & D_{ij} \mid D_{ij-1}, \lambda_j, \varphi \sim ODNB \\ & \mathrm{E}[c_{ij}] = \lambda_j D_{ij-1} \\ & \mathrm{Var}(c_ij) = \varphi (\lambda_j - 1) \mathrm{E}[c_{ij}] \\ \end{split} \tag{10.7} \end{equation}\]

Remark. For both ODP and ODNB

  • Use a quasi-likelihood approach so that the loss data are not restricted to the positive integers

  • Reserve estimates are the same as Chainladder

  • Both are subject to the positivity constraints

    • It’s apparent in the ODNB formula that the column sums must be positive or else we’ll have development factor \(\lambda_j < 1\) \(\Rightarrow\) Varaiance to be negative
  • Advantages (over Kremer):

    • Does not necessarily break down if there are negative incremental loss values

      • In a strict sense, the model requires the incremental losses in a column to be positive otherwise it is more difficult to justify and interpret the inferences (but it doesn’t necessarily break the model)
    • Gives the same reserve estimate as Chainladder

    • More stable than the log-normal model of Kremer

  • Verrall suggest that model can be specified for either incremental or cumulative loss

  • Advantage of the NB:

    • Form of the mean is the same as chainladder
  • If we replace the NB and use normal instead we can deal with the problem of negative incremental claims (not discussed in paper)