Speaker(s): Pietro Parodi, Peter Watson (SCOR)
Benchmark loss curves (examples of which are the SR/Lloyd's curve for property rating, the ISO increased limit factor curves, tail severity distributions for casualty business) often arise out of resource-intensive portfolio studies that are usually done on a one-off basis as it would be too expensive to re-calibrate periodically.
As a result, these benchmark curves remain in circulation for much longer than the lifetime of the assumptions they are based on. In the most extreme cases (e.g. SR curves), practitioners are still using them unchanged after about 50 years.
Also, data is often not used optimally as the number of different curves is not chosen scientifically but based on the underwriter's judgement. In this paper, we show that most of the work necessary to produce the loss curves relevant to a portfolio of risks can be automated and performed by bots (or "intelligent agents"), i.e. autonomous pieces of software that run in the background without need for a user to launch them and that perform a specific task.
Apart from the creation of the curves from scratch, the bots can be used
1. to continually update the parameters of the curves (or the underlying models) in the background (although their actual use in production may need to be approved explicitly);
2. to produce a suitable audit trail of the changes in the curves over time and the reason behind these changes;
3. to split the current curves into further categories whenever sufficient data becomes available, or determine the functional dependency of the parameters of the curves on the significant factors (e.g. type of industry, type of property);
4. to find new categories based on unplanned and possibly unexpected features of the business, using machine learning techniques such as unsupervised learning