As part of the assumption of risks in primary insurance and reinsurance, risk-related losses arise which lead to cash flows between the contractual partners. Losses are assigned to a certain underwriting year after the loss date. However, depending on the insurance industry (for example motor insurance, workers' compensation), losses are incurred over a period of several years, which are then grouped into development years after their respective payout date.
As development progresses over time, the claim already incurred will result in further claims payments (so-called IBNER claims) or claims that have already been incurred but have not yet been reported to the risk carrier (so-called IBNR claims).
The task of the risk taker (primary and / or reinsurance undertaking) is now to estimate the future outstanding claims payments (IBNR / IBNER) to complete the development triangle to a rectangle.
There are various procedures for this, all of which are subject to a certain estimation error, since future claims development is subject to chance.
The IBNR prototype provides functionalities for calculating outstanding reserves at a single loss (Micro level) or an aggregated level (Macro and Meso level) through a Shiny app.
The main goal of the app is to implement all three levels of processes in one kernel:
- using the already known R package ChainLadder for Macro level,
- using the already known R package DCL for Meso level
- and implementing a powerful modern deep learning model for Micro level using Keras, and alternatively a classical micro level approach with k-nearest neighbors.
Note: all the data shown in this app are cumulative.
With such a solution it would be possible to:
- in the case of reserving activities, test all three levels of the procedure (if available) and select the most efficient one (based on the respective estimation error),
- on the micro-level through machine learning, obtain higher accuracies than the ones obtained with the previous approaches,
- avoid redistribution at individual risk level by using the micro model and thereby,
- evaluate retrocession effects more accurately
- reduce solvency capital requirements.
By using a suitable software architecture, this calculation kernel could be used in an insurance company in the various fields of activities (group reserving, pricing of reinsurance contracts, etc.) and, if necessary, also be integrated into existing solutions.
The available data sets are of three different kinds/sources: Macro - claims aggregated per year
Meso - claims aggregated per year; additionally the number of claims per year is given
Micro - Single claims and claim development over the years; the number of claims per year is fix