9.9 Diagnostics
Use diagnostics to judge the quality of the model:
Test model assumptions
Gauge quality of model fit
Guide the adjustments of model parameters
5 diagnostics
9.9.1 Residual Graphs
Plot residuals versus
CY, AY, Age, forecast loss (on x-axis)
Want to see random variability around zero
Bare in mind that we don’t have the same number of residuals at each point (helpful to plot the line for average as well)
Test assumption of iid residuals across the entire triangle
This helps with grouping for hetero adjustment
Plot the relative \(\sigma(r_{wd})\) for each group to further help with the groupings
Do all the plots again after adjustment
9.9.2 Normality Test
Normality is not required, only need this if we’re doing parametric bootstrap with normal distribution
Plot residuals against the normal best fit based on the percentiles
- QQ-plot
Statistical tests:
Check if p-value > 5%
\(R^2\)
AIC
\[2p + n \left [ 1 + \ln(2\pi\dfrac{RSS}{n})\right]\]
- BIC
\[n \ln\left( \dfrac{RSS}{n}\right) + p \ln(n)\]
- \(RSS\) = actual residual - expected residual from normal then squared and summed
9.9.3 Outlier
Remove true outliers but do not remove points that are realistic extreme scenarios
Use box & whisker plot
Box hows 25%-tile to 75%-tile
Whiskers are 3 times the inter quartile range (both side total)
Residuals outside the range are graphed
9.9.4 Parameter Adjustments
Test model with different sets of parameters using GLM bootstrap
- Check parameter significance based on t-statistics (>2)
Parameter selection process:
Start with all the AY and Age parameters (\(\alpha_w\) and \(\beta_d\)) and remove the insignificant ones until only significant parameters are left
Add CY parameter (\(\gamma\)) and check for significance
After selecting parameters:
9.9.5 Review Model Results
Review outputs once we have decided on a model and run the bootstrap
Mean, s.e., CoV, Min/Max, and percentiles by AYs
Incremental fitted mean, s.e. and CoV for each cell in the triangle
- Check for reasonability and consistency
AY | Mean Unpaid | Standard Error | CoV | Min | Max | 50%-tile | 75%-tile | 95%-tile | 99%-tile |
---|---|---|---|---|---|---|---|---|---|
(1) | (2) | (3) = (2) / (1) | (4) | (5) | (6) | (7) | (8) | (9) | |
1 | - | - | - | - | - | - | - | - | - |
\(\vdots\) | - | - | - | - | - | - | - | - | - |
\(w\) | - | - | - | - | - | - | - | - | - |
Total | \(\sum\) | - | - | - | - | - | - | - | - |
Remark. Standard Error: Col (2)
Total s.e. should be greater than any individual year but less than the straight sum of each AY’s s.e.
Expect s.e. to increase going down the column
- This is different from Mack? Where we expect the total s.e. to be greater than the simple sum due to correlation between AYs? (questionable statement here)
Remark. Coefficient of Variation: Col (3)
Total CoV should be less than any individual year (due to diversification of results across AYs)
Except for the most recent AYs, CoV should decrease going down the column (due to larger based of unpaid losses for the more recent AYs)
Higher CoV for the most recent AYs due to:
More parameters used to forcase for the most recent AYs \(\therefore\) parameter uncertainty \(\gg\) process variance
- Model maybe overestimating the uncertainty \(\Rightarrow\) Use BF of Cape Cod
Remark. Min & Max Col (4) - (5)
Check for reasonability (e.g. extreme outcomes from negative values)
- Implausible results can affect the mean