# Two-Stage Sector Rotation Methodology Using Machine Learning and Deep Learning Techniques

**Authors:** Tugce Karatas, Ali Hirsa  
**arXiv:** [2108.02838](https://arxiv.org/abs/2108.02838)  
**Date:** 2021-08-05  
**Summary page:** https://www.summarizepaper.com/en/paper/2108.02838v1/

## Key points

- Two-stage sector rotation methodology using machine learning and deep learning techniques
- First stage involves selecting sector-specific macroeconomic indicators and identifying important features with Recursive Feature Elimination algorithm
- Second stage uses Recurrent Neural Networks models to forecast ETF prices for each sector and rank sectors based on predicted rate of returns
- Evaluation of model performance based on annualized return, Sharpe ratio, and Calmar ratio of portfolios containing top four ranked sectors
- Robustness tests conducted with varying lookback windows and look ahead windows show outperformance of equally weighted portfolios in the long run
- Echo State Networks exhibit exceptional performance compared to other RNN models while being faster to implement
- Paper structure includes literature review, data set description, model background, sector ranking methodology, performance evaluation measures, backtesting results, analysis of performance evolution with increasing lookback window sizes, initial findings summary, and plans for future extensions.

## Summary

This paper presents a two-stage sector rotation methodology that utilizes machine learning and deep learning techniques to predict ETF prices for each sector and rank sectors based on their predicted rate of returns. The first stage involves selecting sector-specific macroeconomic indicators and using the Recursive Feature Elimination algorithm to identify the most important features for each sector. In the second stage, various Recurrent Neural Networks models are employed to forecast future ETF prices for each sector, allowing for the ranking of sectors based on their predicted rate of returns. The study evaluates the performance of different models by considering annualized return, annualized Sharpe ratio, and Calmar ratio of portfolios containing the top four ranked sectors chosen by the model. Robustness tests are also conducted to assess model performance with varying lookback windows and look ahead windows. Empirical results demonstrate that this methodology outperforms equally weighted portfolios even in the long run. Furthermore, a comparison of different RNN models reveals that Echo State Networks exhibit exceptional performance while being faster to implement compared to other models. The paper is structured as follows: Section 2 reviews existing literature on market indicators; Section 3 describes the data set and preprocessing techniques used in this study; Section 4 provides a technical background on implemented models; and Section 5 outlines the sector ranking methodology. Performance measures utilized in evaluating model outcomes are introduced in Section 6 along with backtesting results and an analysis of performance evolution with increasing lookback window sizes. Finally, initial findings are summarized in Section 7 with plans outlined for future extensions. This research contributes to two main areas of literature: using macroeconomic variables in asset return forecasting and developing sector rotation methodologies based on tactical asset allocation strategies.