Enhancing Cross-Sectional Currency Strategies by Context-Aware Learning to Rank with Self-Attention

Abstract

Globally learned learning-to-rank (LTR) algorithms at the core of cross-sectional strategies ignore differences between the distribution of asset features over portfolio rebalances. This flaw produces inaccurate asset rankings that can happen over risk-off episodes and cause unwanted drawdowns.n The authors tackle this shortcoming using the idea of the local ranking context from information retrieval: that a query's top retrieved documents provide vital information about the query's own characteristics, which can then be used to refine the initial ranked list.n Using cross-sectional currency momentum as a case study, the authors encode the ranking context of instruments with a transformer-based architecture and demonstrate the superiority of strategies using context-aware LTR on various performance measures over both normal and risk-off market states.

Sequential asset ranking in nonstationary time series

“…Classical approaches that rely on the ranking of assets include ranking annualised returns [16], or more recent regress-then-rank approaches [13]. Poh et al [28] use a context-aware learning-to-rank model based on the transformer architecture to encode top/bottom-ranked assets, learn the context and exploit this information to rerank the initial results.…”

Spatio-Temporal Momentum: Jointly Learning Time-Series and Cross-Sectional Strategies

“…On the other hand, cross-sectional momentum strategies [15,33,29,30,31], require a momentum score to first be quantified for each individual asset in the portfolio, before computing a relative ranking of these scores in order to formulate positions for a select group of assets. In the first step of calculating a momentum score, cross-sectional momentum strategies ultimately consider only an asset's own historical returns, independent of returns from other assets.…”