Overview

In this project we developed an R-package (available on GitHub at sstoeckl/pensionfinanceLi) to optimize decisions individuals in Liechtenstein’s pension system have to take. The package contains several optimizers as well as full documentation (available via vignette("model") once installed). We determined the most relevant drivers of optimal pension decisions using large-scale optimization, then trained three machine learning models — a hyperparameter-tuned random forest performs best — to allow individuals to receive fast, near-optimal decisions. Results are publicly available at apps.resqfin.com/pfli.

Step-by-Step Project Description

1. Built the R-package pensionfinanceLi (devtools::install_github("sstoeckl/pensionfinanceLi")), carefully validated to avoid counter-intuitive results.

2. Provided detailed documentation of the package and Liechtenstein pension system (also available as vignette("model")).

3. Determined plausibility checks for all input parameters, drew preliminary conclusions on parameter dependencies, and reduced undesired outcomes (e.g. moral hazard).

4. Compiled a final grid of 3,110,400 feasible parameter combinations for large-scale optimization.

5. Due to the time-consuming full-parameter optimization (~25 min/run), optimized 3 key decision variables: consumption share c, retained wealth share alpha, and optimal asset allocation w. Parallelized across Amazon AWS clusters (3 × 96 CPU cores) and continued on an in-house 40-CPU cluster.

6. Ran one- and two-dimensional comparative static analyses to identify the most relevant drivers of optimal pension decisions.

7. Trained three machine learning models in Python (linear regression, k-nearest neighbors, tuned random forest). The random forest achieves surprisingly good out-of-sample performance.

8. Deployed a public Shiny app where anyone can input their individual situation and retrieve near-optimal pension decisions from the trained heuristics.