# NoxTrader: LSTM-Based Stock Return Momentum Prediction for Quantitative Trading

**Authors:** Hsiang-Hui Liu, Han-Jay Shu, Wei-Ning Chiu  
**arXiv:** [2310.00747](https://arxiv.org/abs/2310.00747)  
**Date:** 2023-10-01  
**Summary page:** https://www.summarizepaper.com/en/paper/2310.00747v2/

## Key points

- Financial markets are undergoing a significant transformation due to rapid advancements in algorithmic and machine learning-based trading.
- There is a lack of literature addressing the practical application and potential failures of artificial intelligence models in real-world financial markets.
- The paper aims to develop highly accurate models for successful implementation in actual market scenarios.
- NoxTrader is introduced as a sophisticated system designed for portfolio construction and trading execution with the goal of generating moderate to long-term profits in the stock market.
- NoxTrader leverages LSTM networks and supervised learning techniques for comprehensive analysis of historical data, capturing subtle fluctuations in market prices effectively.
- The system's unique label generation through "return momentum" enhances its predictive capabilities compared to other LSTM-based forecasting models.
- NoxTrader utilizes price and volume data from the US stock market for feature engineering, including Return Momentum, Week Price Momentum, and Month Price Momentum.
- Dynamic model updates during trading execution enable the system to continuously adapt to current market trends.
- NoxTrader features a comprehensive trading backtesting system that enables portfolio management based on predictive scores and custom evaluation metrics for performance assessment.
- Through rigorous feature engineering and precise selection of prediction targets, NoxTrader achieves impressive correlation ranges between 0.65 and 0.75.
- By monitoring dispersion of prediction data and conducting comparative analyses against actual market data with filtering techniques, NoxTrader significantly improves initial investment returns from -60% to 325%, positioning it as a cutting-edge tool for real-world applications in financial markets.

## Summary

In an era marked by rapid advancements in algorithmic and machine learning-based trading, financial markets are undergoing a significant transformation. While many studies focus on improving market prediction accuracy using artificial intelligence models, there is a lack of literature addressing their practical application and potential failures in real-world financial markets. This paper aims to develop highly accurate models and achieve successful implementation in actual market scenarios. Drawing from existing research such as "Machine Learning Approaches in Stock Price Prediction: A Systematic Review" by P. Soni, Y. Tewari, and D. Krishnan (2022) and "101 Formulaic Alphas" by Z. Kakushadze (2015), which provide insights into the use of artificial intelligence models in market prediction and the importance of feature engineering, we introduce NoxTrader as a sophisticated system designed for portfolio construction and trading execution with the primary goal of generating moderate to long-term profits in the stock market. Unlike traditional methodologies, NoxTrader leverages LSTM networks and supervised learning techniques to analyze historical data comprehensively, capturing subtle fluctuations in market prices effectively. The system's unique approach to label generation through "return momentum" enhances its predictive capabilities compared to other LSTM-based forecasting models. NoxTrader utilizes price and volume data from the US stock market for feature engineering, including Return Momentum, Week Price Momentum, and Month Price Momentum. By employing dynamic model updates during trading execution, the system continuously adapts to current market trends. Additionally, NoxTrader features a comprehensive trading backtesting system that enables portfolio management based on predictive scores and custom evaluation metrics for thorough performance assessment. Through rigorous feature engineering and precise selection of prediction targets, NoxTrader achieves impressive correlation ranges between 0.65 and 0.75. By monitoring the dispersion of prediction data and conducting comparative analyses against actual market data with filtering techniques, NoxTrader has significantly improved initial investment returns from -60% to 325%. This refined approach positions NoxTrader as a cutting-edge tool for successful real-world applications in financial markets.