After publication, the authors discovered a minor bug in the code used to produce the results published in “Applications of Gaussian-Inverse Wishart Process Regression Models in Claims Reserving.”

This bug resulted in a numerical error in the data presented in Table 4 and slight alterations in the graphs shown in Figures 8 to 12. While these corrections affect some reported values, they do not impact the methodology or the overall conclusions of the paper.

The revised table, along with updated commentary and code, has been uploaded to the GitHub repository (https://github.com/marcopark90/GaussianProcesses). It is also presented here.

Table 4.Revised GPR and ODP results
Triangle Obs GPR ODP RMSE GPR RMSE ODP Obs. Percentile GPR Obs. Percentile ODP
EMC PA 68,330 67,698 70,711 632 2,381 0.45 0.33
EMC OL 158,514 163,419 134,554 4,905 23,960 0.35 0.95
CC WC 130,095 123,931 127,603 6,164 2,492 0.56 0.65
PRI MM 407,525 425,707 751,244 18,182 343,719 0.59 0.00
HNI CA 28,527 27,055 30,101 1,472 1,574 0.54 0.34

Commentary

It is worth noting that even in the most divergent case, the GPR absolute percentage error on the observed values is 4.7%, compared to 1.9% for the ODP. The two methodologies yield reasonably similar results.

EMC PA
Figure 8.EMC PA—total reserve distribution.
EMC OL
Figure 9.EMC OL—total reserve distribution.
CC WC
Figure 10.CC WC—total reserve distribution.
PRI MM
Figure 11.PRI MM—total reserve distribution.
HNI CA
Figure 12.HNI CA—total reserve distribution.