PaperlayerView exact revision on arXiv

Source-linked AI summary

Financial Time Series Forecasting with Deep Learning : A Systematic Literature Review: 2005-2019

Omer Berat Sezer, Mehmet Ugur Gudelek, Ahmet Murat Ozbayoglu

arXiv:1911.13288v1cs.LGq-fin.CPstat.ML

TL;DR

Financial forecasting attracts extensive ML research, but DL-focused reviews remain lacking. This paper comprehensively surveys DL forecasting studies, categorizes them by asset class and model, and finds rising interest with substantial opportunities for further research.

  • Problem

    Despite numerous traditional ML surveys, comprehensive reviews focused specifically on DL implementations for financial time series forecasting were lacking.

  • Method

    The paper reviews DL studies, grouping them by intended asset class and preferred DL model while examining their characteristics and applications.

  • Results

    The survey finds that overall interest in DL implementations for financial forecasting is rising, with new models generating many research opportunities.

  • Takeaways & Limitations

    Researchers and practitioners can use the survey’s model and application comparisons to assess DL choices for particular forecasting needs and domains.

  • Takeaways & Limitations

    RNN training has difficulty learning long-term dependencies because backward temporal dependence makes learning complex, motivating development of LSTMs.

Abstract

from arXiv · show

Financial time series forecasting is, without a doubt, the top choice of computational intelligence for finance researchers from both academia and financial industry due to its broad implementation areas and substantial impact. Machine Learning (ML) researchers came up with various models and a vast number of studies have been published accordingly. As such, a significant amount of surveys exist covering ML for financial time series forecasting studies. Lately, Deep Learning (DL) models started appearing within the field, with results that significantly outperform traditional ML counterparts. Even though there is a growing interest in developing models for financial time series forecasting research, there is a lack of review papers that were solely focused on DL for finance. Hence, our motivation in this paper is to provide a comprehensive literature review on DL studies for financial time series forecasting implementations. We not only categorized the studies according to their intended forecasting implementation areas, such as index, forex, commodity forecasting, but also grouped them based on their DL model choices, such as Convolutional Neural Networks (CNNs), Deep Belief Networks (DBNs), Long-Short Term Memory (LSTM). We also tried to envision the future for the field by highlighting the possible setbacks and opportunities, so the interested researchers can benefit.

1. Introduction

Financial time-series forecasting has attracted extensive ML research, while DL has recently emerged and remained insufficiently surveyed. This paper addresses that gap by reviewing DL models, applications, comparisons with traditional ML, and future directions.

  • Financial forecasting has extensive ML literature because automated trading and demand for higher yields sustain the search for better models.
  • DL has recently become a prominent predictor class, motivating increasing research on DL-based financial forecasting models.
  • The survey addresses a literature gap by comprehensively reviewing DL implementations for financial time-series forecasting.
  • The review asks which DL models are used, how they compare with traditional ML, and what future research directions exist.
  • The inclusion scope covers English-language journals, conferences, theses, book chapters, arXiv papers, technical publications, and older deep-model studies predating common DL terminology.
  • Studies are organized by DL models and forecasting applications including stocks, indices, trends, commodities, volatility, foreign exchange, and cryptocurrencies.

2. Financial Time Series Forecasting with ML

Financial time-series forecasting has been surveyed extensively through ML and soft-computing perspectives, including broad, model-specific, application-specific, and text-mining reviews. The paper positions DL forecasting as a comparatively under-surveyed area requiring dedicated coverage.

  • Existing surveys cover financial forecasting through traditional soft-computing and ML methods, alongside broader financial applications.
  • Some reviews combining financial time-series studies with other financial applications were included to preserve comprehensive coverage.
  • Model-specific reviews examined techniques such as evolutionary computation and artificial neural networks across financial applications.
  • Text-mining surveys studied market responses to news, abnormal returns, and stock or foreign-exchange prediction.
  • Application-specific surveys focused especially on stock-market forecasting, while also addressing other financial forecasting implementations.
  • DL implementations had not been comprehensively surveyed despite their emergence in financial time-series forecasting, motivating this review.

3. Deep Learning

Deep learning models extend neural-network forecasting through layered representations, sequential memory, generative architectures, and reinforcement-learning formulations. The section describes their structures, training procedures, optimization challenges, and limitations for financial time series.

  • Deep learning uses multiple processing layers to extract higher-level features automatically from input data.
  • Financial forecasting studies use models including DMLP, RNN, LSTM, CNN, RBM, DBN, autoencoder, and deep reinforcement learning.Most financial time series forecasting studies frame prediction as regression, while many trend-prediction studies use classification.
  • Deep Multi Layer Perceptron (DMLP): DMLPs use layered input, hidden, and output units for regression or classification, but fully connected architectures can increase parameters, computation, and storage requirements as input features grow.
  • Deep Multi Layer Perceptron (DMLP): Backpropagation propagates output-layer error backward while optimization algorithms update network weights, including SGD, RMSProp, and ADAM.SGD updates parameters using a randomly selected subset of samples at each iteration; RMSProp adapts learning rates across parameters.
  • Deep Multi Layer Perceptron (DMLP): Deep networks can suffer vanishing gradients when updates weaken in early layers, especially as depth and complexity increase.
  • Recurrent Neural Network (RNN): RNNs process sequential inputs with internal state vectors that retain information about prior observations, making them suitable for longer time periods than traditional recurrent models.
  • Recurrent Neural Network (RNN): RNNs use backpropagation through time, but backward temporal dependence makes training complex and long-term information difficult to learn, motivating LSTM development.
  • Recurrent Neural Network (RNN): RNN performance depends on architectural and training hyperparameters such as sequence length, recurrent-unit type, learning rate, regularization, and optimization algorithm.The reviewed literature includes grid, random, and Bayesian methods for hyperparameter selection.

4. Financial Time Series Forecasting

Financial time-series forecasting studies are organized by forecasting target, feature set, and deep-learning model. Stock-price forecasting is the most studied area, with LSTM-based and hybrid models prominent, while other studies use raw, technical, fundamental, macroeconomic, and text-derived inputs.

  • Forecasting Areas: More than half of deep-learning implementations focus on forecasting the next movement of an underlying financial asset.
  • Forecasting Targets: Financial forecasting studies commonly distinguish price prediction from directional trend prediction, with trend prediction treated as more crucial in many applications.
  • Model Choices: LSTM variants and hybrid models dominate when temporal characteristics are retained, whereas CNN and DFNN are commonly chosen after feature extraction or temporal transformation.
  • Survey Organization: The survey groups papers first by subject, then feature set, and finally deep-learning model or method.
  • Stock-Price Inputs: Stock-price studies use raw OCHLV series, technical or fundamental data, macroeconomic information, financial statements, news, investor sentiment, and social-media feeds.
  • Stock-Price Models: Raw-series stock studies compare DNN, RNN, SRNN, LSTM, GRU, CNN, MLP, ARIMA, and hybrid approaches including CNN+LSTM, DBN+MLP, and RNN+CNN.
  • Multi-Source Studies: Studies using varied data include DNN models with fundamental or macroeconomic features, alongside LSTM and RNN models using broader financial inputs.
  • Text-Mining Studies: Text-mining studies are organized into CNN/LSTM, GRU/LSTM/RNN, and novel-method subgroups for stock-price forecasting.

4.2. Index Forecasting

Index forecasting studies use diverse index datasets, feature sets, and deep-learning or hybrid model families. The survey organizes this literature by input data and model groupings.

  • Index Forecasting: Indices are often forecast instead of individual stocks because they are generally less volatile and better indicate overall economic momentum.
  • Index Forecasting: Common datasets include S&P500, CSI300, NIFTY, NIKKEI225, DJIA, SSE180, HSI, SZSE, and FTSE100.
  • Index Forecasting: The survey groups index studies first by feature set, including raw price or index data, OCHLV, technical indicators, social-media feeds, news, and statistical features.
  • Index Forecasting: Examples include reinforcement-learning/deep-learning combinations for SZSE and S&P500 index prediction, alongside LSTM and RNN applications to several major indices.
  • Index Forecasting: DNN, RNN, LSTM, and CNN are identified as the most commonly used models for index forecasting with varied auxiliary inputs.

4.3. Commodity Price Forecasting

Commodity forecasting research covers widely traded assets such as gold, silver, oil, and copper, using several deep-learning and hybrid model families. Studies evaluate both directional movement and price prediction with varied criteria.

  • Commodity Price Forecasting: Commodity forecasting studies address assets including gold, silver, oil, and copper, with interest expected to grow as online trading availability increases.
  • Commodity Price Forecasting: DNN, RNN, FDDR, and CNN are the most frequently used models for commodity price prediction.
  • Commodity Price Forecasting: Studies use CNN for next-week and next-month directional movement, while DNN, RNN, and LSTM are also applied to commodity, forex, and index time series.
  • Commodity Price Forecasting: Hybrid studies compare FNN and stacked denoising autoencoders with SVR, random walk, and Markov regime-switching models for WTI oil forecasting.
  • Commodity Price Forecasting: Evaluation criteria include accuracy, MAPE, RMSE, MSE, profit, return, Sharpe ratio, and profit-loss curves across forecasting and trading applications.

4.4. Volatility Forecasting

Volatility forecasting studies target price variation for risk assessment and asset pricing. The surveyed methods include CNN, RNN, LSTM, GARCH, Markov, and hybrid approaches using financial, search, and high-frequency data.

  • Volatility Forecasting: Volatility forecasting is motivated by its relationship to price variation and its use in risk assessment and asset pricing.
  • Volatility Forecasting: The survey groups volatility studies into CNN, RNN/LSTM, and hybrid or novel model categories.
  • Volatility Forecasting: A CNN study uses high-frequency trading data, while other studies apply Markov models, Elman RNN, and LSTM to volatility changes.
  • Volatility Forecasting: LSTM-based volatility studies combine price data with Google Domestic trends, while another uses CSI300 and Baidu search-volume data.
  • Volatility Forecasting: Hybrid approaches include RMDN-GARCH and HAR-GASVR, with traditional forecasting and deep-learning models also compared for volatility estimation.

4.5. Bond Price Forecasting

Bond prices are presented as an economically informative forecasting target because they are used to assess economic conditions and yield-curve behavior. The passage introduces the usefulness of bond-price prediction but does not detail surveyed deep-learning implementations.

  • Bond Price Forecasting: Some experts use bond-price changes to analyze economic conditions, viewing bond prices as more representative of economic health than the stock market.
  • Bond Price Forecasting: The passage describes inverted yield curves before recessions and links accurate bond-price prediction to economic analysis.

4.6. Forex Price Forecasting

Forex forecasting studies span multiple deep-learning architectures and hybrid approaches, with models commonly organized into four groups. Research covers major currency pairs and regional variants, using methods including DBNs, DNNs, RNNs, LSTMs, CNNs, autoencoders, and hybrids.

  • Forex research covers major currency pairs, with USD commonly included and regional differences influencing model development.
  • Studies group forex models into CDBN, DBN, DBN+RBM, autoencoder, DNN, RNN, PSN, LSTM, CNN, and hybrid categories.
  • CDBN and DBN studies predicted pairs including EUR/USD, GBP/USD, INR/USD, and CNY/USD, sometimes comparing DBNs with FFNNs.
  • DNN, RNN, PSN, and LSTM studies addressed futures and EUR/USD forecasting, alongside regression, Kalman-filter, and LASSO-based combinations.
  • CNN-based forex studies included multi-model EUR/USD forecasting and an AE+CNN trading agent that outperformed the other agents in a GBP/PLN environment.
  • Hybrid studies compared traditional and deep-learning models for currency forecasting, including TAR-VEC-RHE and LRNFIS variants.

4.7. Cryptocurrency Price Forecasting

Cryptocurrency forecasting research emerged alongside heightened interest in digital assets and focused mainly on price prediction. The literature includes deep and classical models, as well as combinations of market sentiment, text mining, and price data.

  • Cryptocurrency studies emerged in recent years and focused primarily on price forecasting.
  • Bitcoin’s rise from 1000 USD in January 2017 to 20,000 USD in January 2018 drew attention to cryptocurrency prediction and trading research.
  • Cryptocurrency price forecasting used DNN, LSTM, GRU, RNN, ARMA, ARIMA, ARCH, and GARCH models.
  • Some studies combined cryptocurrency price prediction with opinion or text mining, using CNN and LSTM models on Bitcoin, Litecoin, and StockTwits data.

4.8. Trend Forecasting

Trend forecasting is treated as classification rather than regression when predicting asset direction, with studies organized by input data and model family. The literature uses raw prices, technical and fundamental variables, text, order books, and image-like representations.

  • Trend forecasting changes the task from regression to classification when models predict price direction instead of the actual price.
  • The survey groups trend studies by raw time series, technical indicators with price and fundamental data, text mining, and other data variations.
  • Raw time-series studies used ANN, DNN, FFNN, LSTM, RNN, PNN, CNN, DBN, and novel image-based or decomposition-based methods.
  • Studies using technical, price, and fundamental data included ANN, MLP, DBN, RBM, LSTM, GRU, SVM, and XGBoost models.
  • CNN studies transformed financial time series or technical-indicator outputs into two-dimensional images for directional classification.
  • Text-based forecasting combined financial news, social media, word embeddings, and neural models including DNN, CNN, GRU, LSTM, and hybrid architectures.
  • Stock2Vec and Two-stream GRU models used news and stock-price inputs for classification, with results reported as better than state-of-the-art models.
  • Other studies represented limit-order-book data as sequences, flows, histories, or images and applied LSTM, RNN, GRU, and CNN models.

5. Current Snaphot of The Field

The survey’s 140 papers show strong concentration in stock, trend, and index forecasting, with RNNs—especially LSTM—dominant among model choices. Publication activity accelerated recently, CNN interest was expected to grow, and Python-related tools led development environments.

  • The survey included 140 papers and analyzed them by forecasted asset, deep-learning model, and development framework.
  • 46 stock-price, 38 trend, and 33 index studies together represented more than 70% of the surveyed literature.
  • Forex accounted for 19 papers, volatility for 7, and cryptocurrency for 3, although cryptocurrency research was expected to increase.
  • Stock, trend, and index forecasting dominated the field, while publication activity accelerated during the last three years.
  • RNN-based models led with 65 papers, followed by DMLP with 23 and CNN with 20.
  • Within RNNs, LSTM had 58 papers, compared with 27 for Vanilla RNN and 10 for GRU, making LSTM the most popular model.
  • DMLP and CNN were generally preferred for classification problems, including stock, index, and trend forecasting.
  • CNN used increasingly common image-like transformations of financial series, and the survey projected that CNN interest might grow further.

6. Discussion and Open Issues

The survey finds that RNN-based models, especially LSTM, dominate financial time series forecasting, while feature engineering, hybrid models, and emerging architectures shape ongoing research. It also identifies scope boundaries and open opportunities involving profitability, new model classes, automated trading, and specialized hardware.

  • Discussion and Open Issues: Financial forecasting requires attention to profitability because higher prediction accuracy does not necessarily reflect a profitable model.The survey states that the risk-and-reward structure must also be considered when evaluating financial forecasting systems.
  • DL Models for Financial Time Series Forecasting: RNN models are generally preferred for time-series regression, whereas DMLP models are generally preferred for trend classification or buy-sell point identification.The survey reports success for both approaches and does not claim that one model type consistently outperforms the other.
  • Selected Features: Raw time-series data is commonly incorporated into feature vectors, while fundamental, technical, and financial sentiment features broaden inputs for stock and index forecasting.Financial text mining combines sentiment extracted from news, tweets, statements, or blogs with numerical analysis features.
  • Forecasted Asset Types: Stock and index forecasting attract the most interest among asset groups, although price or trend prediction and algorithmic trading are commonly embedded in forecasting studies.Forex forecasting remains popular, while cryptocurrency forecasting is identified as an increasingly active area.
  • Future Model Choices: Enhanced RNN or LSTM variants, hybrid systems, and two-dimensional CNNs that convert time series into image-like data are identified as promising model directions.The survey reports increasing interest in image-based 2-D CNN implementations and expects further studies of this approach.
  • Future Directions: Deep reinforcement learning and agent-based models may fit automated trading environments such as high-frequency trading and robo-advisory systems.The survey connects these opportunities with systems that decide what and when to buy without human intervention.
  • DL Models for Financial Time Series Forecasting: RNN-based models, particularly LSTM, are the most commonly used models, while CNN and DMLP are used extensively for classification after appropriate raw-data processing.The survey attributes LSTM’s popularity to its handling of time-varying data and embedded feedback representations.
  • Responses to Initial Research Questions: Most surveyed studies report better DL performance than ML, but many find comparable results and two studies report better ML performance.The survey also identifies NLP, semantics, and text-mining hybrids combined with time-series data as a likely future direction.

7. Conclusions

The survey reviews DL implementations for financial time series forecasting by organizing studies by asset class and preferred DL model. It finds rising interest in new DL models and identifies substantial opportunities for further research.

  • The survey groups studies by intended asset class and associated deep-learning model.This organization provides a snapshot of the current research status of DL implementations for financial time series forecasting.
  • Overall interest in DL-based financial forecasting is rising through the use of new DL models.
  • The field offers many opportunities for researchers despite financial forecasting’s long research history.

Glossary

The glossary expands abbreviations used throughout the review, covering datasets, forecasting methods, neural architectures, optimization procedures, and evaluation measures.

  • CDBN denotes Continuous-valued Deep Belief Networks, and CDBN-FG denotes Fuzzy Granulation with Continuous-valued Deep Belief Networks.
  • CNN denotes Convolutional Neural Network, while CRBM denotes Continuous Restricted Boltzmann machine.
  • GARCH denotes Generalised Auto-Regressive Conditional Heteroscedasticity, and ARIMA denotes Autoregressive Integrated Moving Average.
  • RMSE denotes Root Mean Square Error, MAPE denotes Mean Absolute Percentage Error, and MSFE denotes Mean Squared Forecast Error.
  • RNN denotes Recurrent Neural Network, while SRNN denotes Stacked Recurrent Neural Network.
  • SVM denotes Support Vector Machine, SVR denotes Support Vector Regressor, and XGBoost denotes eXtreme Gradient Boosting.
Loading 1911.13288v1…