Preserving Valuable Data: Backfilling Returns Using Deep Learning

Abstract

The investment professional industry relies heavily on historical financial data to develop asset allocation, risk, and prediction models. Investment professionals align all data points with variables with the shortest return series. Unfortunately, valuable information is lost during this process. Our findings indicate a robust method to backfill data, using a novel deep learning approach, which inherently considers the non-normality of financial time series. This method could preserve the amount of valuable information in the dataset while producing more robust estimates for the backfilled data series, to assist practitioners in developing a more robust asset allocation, risk, and expected returns models.

Keywords: Backfilling returns, asset allocation, Machine learning, deep neural networks, portfolio construction, risk modelling, expected returns