Stock Prediction using Sequence-to-Sequence Models

Abstract

Aim: Long-Short-Term-Memory (LSTM) deep learning Sequence-to-Sequence(Seq2Seq) models have recently and successfully been used for analyzing text in Natural Language Processing (NLP) applications. As stock data has sequential price patterns over time, this paper aims to explore the application of Long-Short-Term-Memory (LSTM) deep learning Sequence-to-Sequence models to identify the stock price patterns in the stock market. Particularly, we focus on utilizing the Sequence-to-Sequence model and attention mechanism to give multi-day-ahead stock price predictions. The models are trained and tested using daily data on the Australia Stock Market. Then, its performance is analyzed and compared to other deep learning baseline models. To evaluate the profitability, we also propose a trading strategy which makes buy/sell decisions appropriately based on predictions from each model. Findings: The Sequence-to-Sequence prediction model achieved a high average daily accuracy of 97.07%, and outperformed other deep learning baseline models (Multi-Layer Perceptron (MLP), Gated Recurrent Unit (GRU), Attention model) with average daily accuracies of 96.51%, 96.82%, and 97.06% respectively. The Sequenceto- Sequence model showed promising profitability with a 23% return on investment on our trading strategy portfolio compared to the return on investment of the Multi-Layer Perceptron (MLP) (13.6%), Gated Recurrent Unit (GRU) (16.2%) and the Attention model (12.4%). We have demonstrated that our innovative Long-Short-Term-Memory Sequence-to-Sequence Model canto identify the sequential stock price patterns with high accuracy and high profitability.