Reproduce “Few-Shot Learning Patterns in Financial Time-Series for Trend-Following Strategies” on OpenResearch

About this paper

Kieran Wood, Samuel Kessler, Stephen J. Roberts, Stefan Zohren

The X-Trend deep learning model enhances systematic trend-following strategies by rapidly adapting to changing financial market conditions and novel assets using few-shot learning and a cross-attention mechanism over strategically segmented historical regimes. This approach yields an 18.9% increase in Sharpe ratio over a neural baseline during turbulent periods and achieves profitable signals for unseen assets in zero-shot settings.

Key insights

• Intelligently segmenting historical data into distinct regimes using change-point detection for context set construction significantly improves model performance and adaptability.
• Modeling the full (non-Gaussian) distribution of financial returns via quantile regression in the few-shot setting leads to superior risk-adjusted returns compared to simpler Gaussian assumptions.
• Cross-attention provides an interpretable mechanism, allowing the model to dynamically identify and transfer relevant 'universal patterns' across diverse assets and time periods for forecasting.

Summary by alphaXiv.