Data Scarcity and the Decumulation Problem: Two Challenges in Finance

Abstract

This thesis investigates two challenges in quantitative finance through an end-to-end examination of the data-driven pipeline: the scarcity of usable financial data and the retirement decumulation problem. On the data side, the estimation of stochastic model parameters is found to depend more critically on the length of the observed series than on sampling frequency. The Stationary Block Bootstrap consistently outperforms TimeGAN in long-horizon distributional fidelity, with TimeGAN's failures attributed to an architectural mismatch with financial returns. On the decumulation side, the specific realized path of the underlying process emerges as the dominant determinant of neural network policy quality, with tail distributional fidelity serving as a reliable and computationally inexpensive screening criterion for expected policy performance. Together, the findings raise a question that no clean methodology can fully resolve: how much of what we learn from a single observed history reflects the underlying process, and how much is simply a portrait of one particular past.