PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey of Forex and Stock Price Prediction Using Deep Learning

Zexin Hu, Yiqi Zhao, Matloob Khushi

arXiv:2103.09750v1q-fin.ST

TL;DR

Stock and Forex forecasting is difficult because prices and exchange rates reflect many interacting economic, political, and market factors. This survey compares 88 deep-learning studies across models, datasets, variables, and evaluation metrics, finding strong reported performance for several approaches, including LSTM, DNN, and reinforcement learning.

  • Problem

    Stock and Forex forecasting must process complex, rapidly changing information influenced by numerous economic, political, and market factors.

  • Method

    The survey reviews 88 studies, comparing datasets, variables, deep-learning models, and evaluation metrics across stock and Forex forecasting.

  • Results

    Reported strong results included MSE smaller than 0.01 with LSTM, accuracy of 90%-100% with LSTM, and return rates of 1.4-1.8 with reinforcement learning.

  • Takeaways & Limitations

    The survey offers a comparative view of deep-learning approaches and their reported performance for stock and Forex forecasting.

Abstract

from arXiv · show

The prediction of stock and foreign exchange (Forex) had always been a hot and profitable area of study. Deep learning application had proven to yields better accuracy and return in the field of financial prediction and forecasting. In this survey we selected papers from the DBLP database for comparison and analysis. We classified papers according to different deep learning methods, which included: Convolutional neural network (CNN), Long Short-Term Memory (LSTM), Deep neural network (DNN), Recurrent Neural Network (RNN), Reinforcement Learning, and other deep learning methods such as HAN, NLP, and Wavenet. Furthermore, this paper reviewed the dataset, variable, model, and results of each article. The survey presented the results through the most used performance metrics: RMSE, MAPE, MAE, MSE, accuracy, Sharpe ratio, and return rate. We identified that recent models that combined LSTM with other methods, for example, DNN, are widely researched. Reinforcement learning and other deep learning method yielded great returns and performances. We conclude that in recent years the trend of using deep-learning based method for financial modeling is exponentially rising.

1. Introduction

Stock and Forex price prediction is challenging because prices reflect numerous economic, political, social, market, and informational influences. This survey examines how different deep learning methods affect forecasting accuracy by comparing published studies across their datasets, variables, models, and results.

  • Stock prices reflect future growth expectations and fluctuate with macroeconomic conditions, market anticipation, management confidence, and increasingly abundant timely information.
  • Forex forecasting is challenging because exchange rates are influenced by countries’ economies, politics, societies, and the international situation.Exchange-rate predictions may help investors increase returns and reduce risk.
  • Forex and stock markets share technical indicators, candle charts, and sensitivity to national market sentiment, motivating their joint analysis.
  • The survey investigates deep learning applications in both markets and their effects on price-trend prediction accuracy.
  • Each reviewed paper is analyzed by dataset, variables, adopted deep learning model, and prediction-model results.

2 Related deep learning methods and input introduction

The section introduces deep learning methods used for stock and Forex prediction, emphasizing their suitability for extracting market patterns, modeling temporal dependencies, fitting nonlinear relations, and learning profitable trading actions.

  • CNN: CNN can extract stock-market characteristics from feature graphs and support timing-selection strategy construction.Its weight-sharing architecture differs from fully connected networks such as DBN, SAE, and BP.
  • RNN: RNN models sequential data by incorporating previous hidden states, making it suitable for stock and Forex fluctuations linked to prior trends.Its training considers contextual information across time.
  • LSTM: LSTM, an RNN variant, uses self-loops to maintain gradient paths, addressing vanishing gradients while modeling nonlinear time-series correlations for stock prediction.LSTM extracts data characteristics from historical stock or Forex data for trend prediction.
  • DNN: DNNs model complicated nonlinear functions with high-level abstraction and can be trained by backpropagation for stock and Forex trend prediction.The section associates DNNs with sizable data and complicated nonlinear mapping relations.
  • Reinforcement learning: Reinforcement learning learns strategies through exploration and feedback, treating market context as states, trading decisions as actions, and gains or losses as rewards.It can learn timing actions involving price, trading duration, and order size according to market response.
  • Other deep learning methods: Other methods discussed include Hybrid Attention Networks, self-paced learning mechanism NLP, multi-filters neural networks, and Wavenet, which occur infrequently in the selected articles.These methods are reserved for discussion in section 4.7.

3. Review methodology and criteria · 4. Results

The survey selected and reviewed 88 post-2015 stock and Forex forecasting papers from DBLP and Microsoft Academic, then compared six deep learning method categories across datasets, variables, and model types. The results section also summarized CNN-based studies and their reported paper information.

  • 3.1 Paper selection methods: The reviewed articles were sourced from the DBLP computer science bibliography and Microsoft Academic.The survey focused on publications after 2015 to provide timeliness.
  • 3.1 Paper selection methods: Searches used CNN, LSTM, deep learning, RNN, and reinforcement learning stock/Forex keywords, while Microsoft Academic results were filtered to 2015–2021 and deep learning.DBLP selection also excluded informally published journals and conferences.
  • 3.1 Paper selection methods: 88 papers published from 2015–2020 were selected for review, with the largest yearly groups appearing in 2018 and 2019.The distribution was 4 papers in 2015, 1 in 2016, 15 in 2017, 30 in 2018, 28 in 2019, and 10 in 2020.
  • 3.2 Selected paper statistics: The study compared CNN, LSTM, DNN, RNN, reinforcement learning, and other deep learning methods across datasets, input variables, and model types.Results were individually analyzed and then compared for discussion and conclusions.
  • 3.2 Selected paper statistics: The survey presented annual paper counts, method distributions, systematic-review information, and article-type distributions through figures and Table 1.These materials documented the composition of the selected-paper set.
  • 4.1.1 CNN: 17 articles using CNN technology for stock and Forex prediction were briefly described in the CNN results subsection.Table 2 records the authors, variables, datasets, and models for the CNN papers.

4.1.2 RNN · 4.1.3 LSTM

The survey separately reviews RNN- and LSTM-based approaches for stock and Forex prediction. It describes five RNN articles and 27 LSTM articles, with corresponding tables summarizing authors, variables, datasets, and models.

  • 4.1.2 RNN: Five articles using RNN technology for stock and Forex prediction are briefly described.The review presents their overall information in Table 3.
  • 4.1.2 RNN: The RNN section focuses on technology used for stock and Forex prediction.It introduces the reviewed articles before referring readers to Table 3.
  • 4.1.2 RNN: Table 3 organizes the RNN papers by author, variables, dataset, and model.The table provides the overall information for papers using an RNN model.
  • 4.1.3 LSTM: Twenty-seven articles using LSTM technology for stock and Forex prediction are briefly described.The review presents their overall information in Table 4.
  • 4.1.3 LSTM: The LSTM section focuses on technology used for stock and Forex prediction.It introduces the reviewed articles before referring readers to Table 4.
  • 4.1.3 LSTM: Table 4 organizes the LSTM papers by author, variables, dataset, and model.The table provides the overall information for papers using an LSTM model.

4.1.4 DNN

This section briefly describes 13 articles that applied deep neural network (DNN) technology to stock and Forex prediction. It also presents overall information about the papers using DNN models.

  • 4.1.4 DNN: 13 articles used DNN technology for stock and Forex prediction.The articles are briefly described in the section.
  • 4.1.4 DNN: Table 6 provides overall information on papers using DNN models.The table is identified as covering papers that used DNN models.

4.1.5 Reinforcement learning

This section briefly describes eight articles that applied reinforcement learning to stock and Forex prediction, with Table 6 summarizing their authors, variables, datasets, and models.

  • 4.1.5 Reinforcement learning: Eight articles using reinforcement learning for stock and Forex prediction are briefly described.Table 6 presents the overall information for these papers.
  • 4.1.5 Reinforcement learning: Table 6 summarizes the authors, variables, datasets, and models of the reviewed papers.The table is presented as the overall information for papers using reinforcement learning models.

4.1.6 Other Deep Learning Methods

This section briefly describes eight articles that applied other deep learning technologies to stock and Forex prediction. Table 7 summarizes each paper’s authors, variables, dataset, and model.

  • Scope: Eight articles used other deep learning technologies for stock and Forex prediction.The articles are briefly described in this section.
  • Table 7: Table 7 presents the authors of the papers using other deep learning methods.
  • Table 7: Table 7 reports the variables, dataset, and model for the included papers.

4.2 Papers Results grouped by the method used

The survey grouped reported results by deep learning method and analyzed them using the differing performance metrics across studies. Results represented the average performance of the best-performing models reported in the papers, with tables covering CNN, RNN, LSTM, and DNN models.

  • CNN: CNN results were analyzed by performance metric and summarized in Table 8 using the average performance of the best-performing models.The passage states that the survey analyzed differing metrics and reported average results for the best-performing models.
  • RNN: RNN results were analyzed by performance metric and presented in Table 9 as the average performance of the best-performing models.The table was specifically reported for papers using the RNN model.
  • LSTM: LSTM results were analyzed by performance metric and presented in Table 10 as the average performance of the best-performing models.The table was specifically reported for papers using the LSTM model.
  • DNN: DNN results were analyzed by performance metric and presented in Table 11 as the average performance of the best-performing models.The table was specifically reported for papers using the DNN model.

4.2.5 Reinforcement learning

The reinforcement-learning studies were analyzed using the differing performance metrics reported across articles. Results were represented by the average performance of the best-performing models, as presented in Table 12.

  • 4.2.5 Reinforcement learning: The survey analyzed reinforcement-learning papers according to the different performance metrics used across articles.The reported results were based on the average performance of the best-performing models in the cited papers.
  • 4.2.5 Reinforcement learning: Table 12 presents the metrics and results for papers using reinforcement-learning models.
  • Other deep learning methods: The survey applied the same metric-based analysis approach to papers using other deep-learning methods, with results averaged from their best-performing models.These findings were presented in Table 13.
  • Other deep learning methods: Table 13 presents the metrics and results for papers using other deep-learning methods.

5. Discussion

The discussion identifies distinct dataset, input-variable, and model-modification patterns across CNN, RNN, LSTM, DNN, reinforcement-learning, and other deep-learning approaches. It also compares studies using RMSE, MAPE, MAE, MSE, accuracy, F-measure, Sharpe ratio, and return rate.

  • CNN models: CNN studies mainly used technical analysis, with 6 papers combining it with sentiment and news analysis; close price was used by all models.Five papers used close price only, while 12 modified traditional CNN models, most commonly combining CNN and LSTM.
  • RNN-based models: RNN studies primarily used technical analysis and multivariable inputs containing open, close, highest, and lowest prices.All papers modified the traditional RNN model, and two selected C-RNN models.
  • LSTM papers: LSTM studies mostly used technical analysis, with close price included in 23 models; 15 papers modified traditional LSTM models.Attention-based LSTM and LSTM with RNN were most frequent, while two papers used LSTM with GRU.
  • DNN and reinforcement learning: DNN and reinforcement-learning studies relied on technical analysis and predominantly multivariable inputs, while 3 of 6 reinforcement-learning papers combined model improvements with LSTM.Three DNN papers modified traditional DNN models, and all improved DNN models were different.
  • Other deep learning methods: Other deep-learning methods included sentiment and news analysis in 4 papers, used multivariable inputs in five models, and featured HAN three times.The remaining models were not duplicated; paper had the highest accuracy, while paper [70] had the highest return rate.

6. Conclusions

The paper reviewed 88 studies on deep-learning approaches for stock and Forex price prediction, comparing datasets, variables, models, and evaluation metrics. It identified limited research combining multiple deep-learning methods and highlighted hybrid networks as promising for future work.

  • Scope and contribution: The review covered 88 papers from 2015 to the present on stock and Forex price-movement prediction using deep-learning methods.It analyzed datasets, variables, model use, and evaluation metrics.
  • Scope and contribution: The surveyed techniques included CNN, LSTM, DNN, RNN, reinforcement learning, HAN, NLP, and Wavenet.The review compared a broad range of deep-learning methods for stock and Forex prediction.
  • Future research: The review identified a lack of studies combining multiple deep-learning methods, especially with other deep-learning methods.This gap motivates further investigation of combined approaches.
  • Future research: Hybrid networks were described as showing promising signs for future stock and Forex prediction research.The authors stated that they would design a specific hybrid model based on the review’s analysis.
Loading 2103.09750v1…