8 - 15 Jun 2022
Neural Networks Meet Least Squares Monte Carlo at Intern. Model Data
Solvency II aims at implementing a set of robust solvency rules for insurance companies, which takes the most material risks into account in an adequate way. In principle, the Solvency II framework requires the derivation of the full loss distribution of the available Own Funds, with the goal of deriving its correct Value-at-Risk. This particularly does not only involve a market consistent calculation of the economic balance sheet items at the valuation date but also its re-evaluation for each possible scenario at the risk horizon (one year within Solvency II).
Most insurance companies avoid this enormous effort by applying the standard formula approach to calculate the Solvency Capital Requirement (SCR). But the largest life insurers usually stick to the original Solvency II requirement and develop a full-scale internal model which allows them to calculate the economic balance sheet for thousands of one-year scenarios. The focus of this web session is on presenting a regression-based Monte Carlo approach in order to estimate the SCR. By doing so, we challenge the state-of-the-art Least Squares Monte Carlo approach based on polynomials by the most promising machine learning technique, namely ensemble of neural networks.Organised by the EAA – European Actuarial Academy GmbH.
The web session is open to all interested persons. A certain level of programming knowledge will be beneficial, since we will work with an R notebook. Further, a basic understanding of the risk modelling and SCR calculation will be advantageous for the comprehension of the topic.
Technical RequirementsPlease check with your IT department if your firewall and computer settings support web session participation (the programme Zoom is used for this online training). Please also make sure that you are joining the web session with a stable internet connection.
In case you want to run and adapt the provided R code, please install an R environment; a notebook will be sent out for testing purposes in advance. The published use case at https://github.com/DeutscheAktuarvereinigung/insurance_scr_data is based on Python.
Purpose and Nature
In August 2020 we published the "Comprehensive Internal Model Data for Three Portfolios", see https://github.com/DeutscheAktuarvereinigung/insurance_scr_data, as an outcome of our work for the working group "Statistical Methods - Actuarial Data Science" of the German Actuarial Association. In the first part of this web session, we introduce the framework for calculating the SCR with an internal model and explain the data set for our use case. Our internal model approach, which is of high practical interest for insurance companies, is based on deriving proxy functions via regression in order to derive and explain the stress behaviour of balance sheet items. By doing so, we focus our interest on neural networks and introduce this important machine learning method. We also concentrate on a practical implementation in R with the aim of having a notebook ready to train neural networks before the end of the first day.
In the period between the first and the second part of this web session the participants should perform their own experiments. A hyperparameter search will be included in the code, but it will also be possible to amend the code and challenge the best results obtained using neural networks. Afterwards, we discuss the results and want to find out whether neural networks provide higher accuracy to estimate the SCR. Acknowledging that the volume of our published data exceeds the computational resources of most insurance companies, we then analyze whether the high accuracy can be preserved by training the neural networks model using significantly less scenarios and a smart way of deriving the hyperparameters. Hereby, the focus is on reducing the computational time and memory consumption without compromising the approximation quality. Finally, we showcase an efficient approach which requires a feasible number of scenarios and can be fully implemented using the same open-source software and libraries used in the first day.
Dr Christian JonenChristian is leading the internal model validation unit at Generali Deutschland AG. He holds a PhD in mathematical finance and is a member of the German Association of Actuaries (Aktuar DAV). Before his time at Generali, he worked as an IT project manager in the department Change Delivery at HSBC.
Dr Zoran Nikolić Zoran works at B&W Deloitte in the area of actuarial data science, risk modelling and actuarial analytics. Between 2010 and 2019 he has worked for Generali in the modelling team. He is a Certified Actuarial Data Scientist (DAV) and member of the DAV Committee Actuarial Data Science.
Dr Tamino Meyhöfer Tamino works as risk modeler at Generali Deutschland AG. He holds a PhD in financial mathematics and is a member of the German Association of Actuaries (Aktuar DAV).