RIFT: Pretraining and Applications for Representations of Interrelated Financial Time Series

Abstract

This article introduces a new financial time-series representation model called representations of interrelated financial time series (RIFT). RIFT combines a novel pretraining task and neural network architecture to create generalized representations of multiple financial time-series inputs. The network uses a Siamese architecture to predict pairwise future correlations of securities; the encoder can then be used to create representations of individual securities for downstream tasks. Similar to successful applications of transfer learning in other domains, the authors test the representations on several downstream tasks common in quantitative finance, including dimensionality reduction, portfolio optimization, and portfolio reconstruction. In particular, the article introduces neural hierarchical risk parity (HRP), an improvement on the HRP algorithm, the current state of the art for portfolio optimization, and shows promising results across a variety of assessment criteria, including a 6.0% relative improvement in annualized returns and a 5.6% relative improvement in the Sharpe ratio.