Decision methods in pension finance: Large-scale optimization
Project Description
In a former project funded by the FFF (Fin-19-2) we have developed a software package that models the financial decisions an individual insured in Liechtenstein has to take regarding his/her pensions. This includes optimal consumption and saving decisions in the context of the specific financial situation of the insured. It also includes the allocation of savings to the three pillars as well as the direct investment in various financial assets and decisions on whether to receive the entire savings as a lump sum when entering retirement or the option to transform the savings into a (life-long) annuity.
As those optimizations require a large amount of time (10 minutes) per parameter setting we have run a limited number of such optimizations (approx. 900’000) for a pre-specified parame-ter grid (amounting to 3.2 mn combinations) and then used the optimizations to train a Machine Learning model to accurately and quickly (2-5 seconds) approximate these optimal decisions. Even after the end of the former project, we continued running these optimizations to improve the accuracy of the Machine Learning (ML) model, but due to the Ransomware attack at the University of Liechtenstein, many of our results as well as the computing resources were lost and are not recoverable. In this project, we therefore apply for additional funding to finish the computations on a large-scale optimization cluster, retrain our ML model and then provide the insured with accurate and quickly available optimal pension decisions on the dedicated website (app currently unavailable due to the server loss): https://apps.resqfin.com/pfli.
As those optimizations require a large amount of time (10 minutes) per parameter setting we have run a limited number of such optimizations (approx. 900’000) for a pre-specified parame-ter grid (amounting to 3.2 mn combinations) and then used the optimizations to train a Machine Learning model to accurately and quickly (2-5 seconds) approximate these optimal decisions. Even after the end of the former project, we continued running these optimizations to improve the accuracy of the Machine Learning (ML) model, but due to the Ransomware attack at the University of Liechtenstein, many of our results as well as the computing resources were lost and are not recoverable. In this project, we therefore apply for additional funding to finish the computations on a large-scale optimization cluster, retrain our ML model and then provide the insured with accurate and quickly available optimal pension decisions on the dedicated website (app currently unavailable due to the server loss): https://apps.resqfin.com/pfli.
