**Authors:** Simão Moraes Sarmento, Nuno Horta
**Published:** _Expert Systems with Applications_, 2020
**Link:** [DOI](https://doi.org/10.1016/j.eswa.2020.113490)
---
### Summary
This study enhances the traditional **Pairs Trading** strategy by integrating **Machine Learning** techniques to:
1. **Find Profitable Pairs Efficiently**
2. **Avoid Long Decline Periods** caused by divergent price movements
---
### Key Points
- **Pairs Trading:** A market-neutral strategy where an investor buys an undervalued security and sells an overvalued one, profiting when prices converge.
- **Problem 1 - Finding Pairs Efficiently:**
- Traditional methods are exhaustive or sector-limited, which can miss opportunities.
- Solution: Apply **PCA (Principal Component Analysis)** for dimensionality reduction followed by **OPTICS Clustering** to identify high-potential pairs with reduced noise.
- Results: This approach achieved an average **Sharpe Ratio of 3.79**, outperforming traditional methods.
- **Problem 2 - Reducing Decline Periods:**
- Divergent price movements can lead to long drawdowns.
- Solution: Use **time-series forecasting** models like **ARMA**, **LSTM**, and **LSTM Encoder-Decoder** to predict spread movements and adjust trades dynamically.
- Results: Reduced portfolio decline periods by **75%**, although profitability decreased slightly.
- **ETFs as the Asset Class:** The study focuses on 208 commodity-linked ETFs, trading at 5-minute intervals, which showed strong mean-reversion characteristics suitable for Pairs Trading.
---
### How It Works
1. **PCA + OPTICS Clustering:** Identifies robust, co-integrated pairs with minimal noise.
2. **Forecasting Models:** Predicts spread changes to optimize entry and exit points.
3. **Trading Simulation:** Uses real 5-minute data from 2009 to 2018 to validate models and strategies.
---
### Strengths
- Enhanced pair selection with clustering...less noise, better matches.
- Machine learning forecasting reduces bad trades and decline periods.
- Market-neutral approach suitable for volatile environments.
---
### Limitations
- Slightly lower profitability when optimizing for reduced drawdowns.
- Performance drops with fewer available pairs during longer formation periods.
- Needs more exploration with different asset classes like stocks or futures.