# Slow Momentum with Fast Reversion: A Trading Strategy Using Deep Learning and Changepoint Detection

**Authors:** Kieran Wood, Stephen Roberts, Stefan Zohren  
**arXiv:** [2105.13727](https://arxiv.org/abs/2105.13727)  
**Date:** 2021-05-28  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2105.13727v3/

## Key points

- Momentum strategies are important in alternative investments, especially in commodity trading advisors (CTAs).
- These strategies struggle to adapt to rapid changes in market conditions, as seen during the 2020 market crash.
- Time-series momentum (TSMOM) strategies make poor bets after momentum turning points.
- The authors propose a novel approach that combines an online changepoint detection (CPD) module with a Deep Momentum Network (DMN) pipeline.
- The DMN uses Long Short-Term Memory (LSTM) architecture for trend estimation and position sizing.
- The CPD module helps balance a slow momentum strategy and a fast mean-reversion strategy regime.
- It generates outputs indicating the location and severity of changepoints, allowing the model to respond appropriately.
- Backtesting from 1995 to 2020 shows that incorporating the CPD module improves the Sharpe ratio by one-third compared to traditional models.
- Performance boosts of approximately two-thirds were observed during recent years tested (2015–2020).
- This approach addresses challenges faced by momentum strategies when adapting to changing market conditions.

## Summary

Momentum strategies play a crucial role in alternative investments, particularly in commodity trading advisors (CTAs). However, these strategies often struggle to adapt to rapid changes in market conditions, as evidenced by the challenges faced during the 2020 market crash. One specific issue arises immediately after momentum turning points, where trends reverse from an uptrend to a downtrend or vice versa. Time-series momentum (TSMOM) strategies are prone to making poor bets in such situations. To address this problem and enhance the response to regime change, the authors propose a novel approach that incorporates an online changepoint detection (CPD) module into a Deep Momentum Network (DMN) pipeline. The DMN utilizes a deep-learning architecture called Long Short-Term Memory (LSTM), which enables simultaneous learning of trend estimation and position sizing. By integrating the CPD module, the model can effectively balance two key components: 

1) A slow momentum strategy that capitalizes on persistent trends without overreacting to localized price movements; and 2) A fast mean-reversion strategy regime that promptly flips positions and then reverts them back to exploit localized price moves. The CPD module generates outputs indicating the location and severity score of changepoints, allowing the model to learn how to respond appropriately based on data-driven insights. To evaluate their model's performance, the authors conducted backtesting over the period from 1995 to 2020. The inclusion of the CPD module resulted in a one-third improvement in Sharpe ratio—a measure of risk-adjusted returns—compared to traditional models. The benefits of this module were particularly pronounced during periods of significant nonstationarity, with performance boosts reaching approximately two-thirds during recent years tested (2015–2020). Notably, this is intriguing because traditional momentum strategies have underperformed during this period. Overall, this study introduces a promising approach for addressing challenges faced by momentum strategies when adapting to changing market conditions. By incorporating an online CPD module into a DMN pipeline, it achieves improved performance and better navigates regime changes while enhancing risk-adjusted returns for investors.