Financial Time Series Forecasting with Deep Learning : A Systematic Literature Review: 2005-2019 (1911.13288v1)

Abstract: 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.

A Comprehensive Review of Deep Learning Applications in Financial Time Series Forecasting

The paper "Financial Time Series Forecasting with Deep Learning: A Systematic Literature Review: 2005-2019" by Sezer, Gudelek, and Ozbayoglu provides an essential and comprehensive summary of deep learning (DL) applications in the domain of financial time series forecasting. The extensive review compiles various DL methodologies, their practical implementations, and implications, offering a detailed analysis that will benefit researchers and practitioners in both academia and the financial industry.

The paper meticulously categorizes the plethora of existing studies into specific forecasting domains such as stock price, index, commodity, volatility, bond, forex, cryptocurrency price forecasting, and trend forecasting. This structured approach aids in comprehending the scope of DL's applicability in different financial contexts.

Deep Learning Models in Financial Time Series Forecasting

The review highlights several DL models, each with its unique characteristics and suitability for different types of forecasting problems:

1. Recurrent Neural Networks (RNNs): Predominantly featuring Long Short-Term Memory (LSTM) networks, RNNs are recognized for their ability to handle sequential data, making them ideal for financial time series forecasting. The survey shows a clear inclination towards RNN-based models for both price and trend forecasting, emphasizing their robustness in capturing temporal dependencies.
2. Deep Multilayer Perceptrons (DMLPs): These models are frequently chosen for classification tasks such as trend prediction. The review underscores the importance of data preprocessing to achieve stationarity in time series data before applying DMLPs, aligning with neural network characteristics to improve model performance.
3. Convolutional Neural Networks (CNNs): While traditionally used in computer vision, CNNs find innovative applications in financial forecasting by transforming 1-D time series data into 2-D images. This transformation enables CNNs to leverage their filtering capabilities for better feature extraction and trend classification.
4. Autoencoders (AEs), Restricted Boltzmann Machines (RBMs), and Deep Belief Networks (DBNs): These models are explored for unsupervised learning and feature extraction, providing foundational layers for more complex architectures like hybrid models.
5. Deep Reinforcement Learning (DRL): Although fewer in number, DRL applications are emerging, particularly in algorithmic trading. The paper points out the potential of these models in developing intelligent trading systems through agent-based interactions and decision-making processes.

Practical Implementations and Implications

The extensive categorization of studies reveals how different DL models are implemented across various financial instruments. For example, stock price forecasting, with the highest number of studies, showcases a preference for RNNs and hybrid models integrating multiple DL techniques. The use of technical indicators and fundamental analysis in conjunction with DL models is a common theme, indicating a holistic approach to capturing market dynamics.

Key Observations and Trends

• Stock and Index Forecasting: These areas dominate the DL applications with a pronounced use of LSTMs and CNNs. The integration of technical and fundamental analysis is prevalent, highlighting a comprehensive feature set for improving prediction accuracy.
• Volatility Forecasting: LSTM and hybrid models incorporating GARCH techniques show promising results in capturing market volatility, emphasizing the need for models that can handle complex volatility structures.
• Cryptocurrency Forecasting: Emerging as a new area of interest, cryptocurrency forecasting leverages LSTM and CNN models to handle the high volatility and unconventional trading patterns in the crypto markets.
• Text Mining and Sentiment Analysis: Increasingly, text mining and sentiment analysis techniques are being integrated with DL models to capture market sentiment. These hybrid approaches illustrate the evolving nature of DL applications, aiming for more comprehensive market analysis.

Future Directions

The survey suggests several future directions:

• Advanced Text Mining: Integrating NLP and sentiment analysis to enhance DL models further.
• Agent-based DRL Models: Exploring DRL for developing autonomous trading systems that can interact and learn from market conditions.
• New DL Architectures: Investigating the applicability of newer architectures like GANs and Capsule Networks in financial forecasting.
• Hardware Optimization: Leveraging GPUs and FPGAs for constructing high-frequency trading models and optimizing financial forecasting algorithms.

Conclusion

The paper provides a crucial reference point for researchers entering the field of DL applications in financial time series forecasting. By systematically presenting the categorized studies, methods, and future research directions, it serves as an essential guide for developing more accurate and efficient financial forecasting models. The detailed review underscores the evolving landscape of DL in finance, driven by increasing computational power and sophisticated data integration techniques, paving the way for innovative solutions in financial forecasting.