PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Learning for Financial Applications : A Survey

Ahmet Murat Ozbayoglu, Mehmet Ugur Gudelek, Omer Berat Sezer

arXiv:2002.05786v1q-fin.STcs.LGstat.ML

TL;DR

Deep learning applications in finance are expanding, but comprehensive synthesis of their implementations remains limited. This survey categorizes financial deep-learning studies by application area and model, finding LSTM dominance alongside growing CNN and text-mining use, and identifies opportunities for future research.

  • Problem

    Despite rising financial deep-learning research, no comprehensive survey had focused on its implementations across finance application areas.

  • Method

    The paper surveys and categorizes financial deep-learning studies by finance subfield, analyzes their model structures, and identifies future research directions.

  • Results

    LSTM is the dominant model for financial time-series forecasting, while DMLP commonly serves classification and CNN and text-mining applications are gaining traction.

  • Takeaways & Limitations

    The survey presents finance deep learning as an early, expanding field with substantial opportunities for new models and applications.

  • Takeaways & Limitations

    The surveyed literature contains few noteworthy financial applications of GANs, capsule networks, and deep Gaussian processes, with DRL only recently attracting attention.

Abstract

from arXiv · show

Computational intelligence in finance has been a very popular topic for both academia and financial industry in the last few decades. Numerous studies have been published resulting in various models. Meanwhile, within the Machine Learning (ML) field, Deep Learning (DL) started getting a lot of attention recently, mostly due to its outperformance over the classical models. Lots of different implementations of DL exist today, and the broad interest is continuing. Finance is one particular area where DL models started getting traction, however, the playfield is wide open, a lot of research opportunities still exist. In this paper, we tried to provide a state-of-the-art snapshot of the developed DL models for financial applications, as of today. We not only categorized the works according to their intended subfield in finance but also analyzed them based on their DL models. In addition, we also aimed at identifying possible future implementations and highlighted the pathway for the ongoing research within the field.

1. Introduction

This survey maps the growing body of deep-learning research across financial applications, aiming to guide researchers and practitioners in selecting models and research paths. It assesses application areas, research maturity, model success relative to traditional techniques, and future opportunities.

  • Financial ML research addresses stock forecasting, algorithmic trading, credit risk, portfolio allocation, asset pricing, and derivatives markets.
  • Deep learning for finance is an emerging, increasingly published field whose implementations the survey organizes for researchers and practitioners.
  • The survey evaluates which financial application areas interest the DL community, how mature they are, and which offer promising academic or industrial research potential.
  • It examines preferred and successful DL models across applications and compares them with traditional soft computing and ML techniques.
  • The paper focuses on DL implementations for financial applications, excludes its separately surveyed financial time-series forecasting literature, and includes DL-based algorithmic trading studies.
  • The survey claims to be the first comprehensive study concentrated on DL implementations in finance and intended to highlight advancements and provide a research roadmap.

2. Machine Learning in Finance

Machine learning has been studied in finance for about 40 years, producing thousands of papers across financial fields. This survey focuses on deep learning while situating it within prior surveys of broader ML applications and techniques.

  • Prior ML research: Finance has been a major ML application area for about 40 years, with thousands of published studies and continuing research interest.The survey cites related surveys from the past 20 years while concentrating specifically on deep learning implementations.
  • General ML surveys: General ML surveys compare or review diverse approaches, including Artificial Neural Networks, Expert Systems, Hybrid models, data mining, Genetic Algorithms, rule-based systems, and Neural Networks.These surveys take a broad perspective rather than concentrating on one financial implementation area.
  • Technique-focused surveys: Technique-focused surveys cover Evolutionary Algorithms such as Genetic Algorithms and Particle Swarm Optimization, particularly for financial optimization and portfolio selection.Additional prior work includes books on Genetic Algorithms and Genetic Programming in computational finance.
  • Method- and application-focused surveys: Other reviews focus on specific methods or domains, including Artificial Neural Networks, agent-based computational finance, and kernel methods involving Principal Component Analysis and Support Vector Machines.Application-specific surveys are deferred to Section 4, followed by an overview of deep learning model structures in the next section.

3. Deep Learning … 3.3. Recurrent Neural Network (RNN)

The paper surveys Deep Learning as a multilayer form of machine learning that provides high-level data abstractions, covering DMLP, CNN, RNN, and other architectures. It describes DMLP learning, CNN components, and RNN sequence processing, memory, structures, and training.

  • 3. Deep Learning: Deep Learning uses multiple ANN layers to provide high-level abstraction for data modelling.The surveyed models include DMLP, CNN, RNN, LSTM, RBMs, DBNs, and AEs.
  • 3.1. Deep Multi Layer Perceptron (DMLP): DMLP extends ordinary MLPs by using more layers while retaining input, output, and hidden layers.Each neuron has input, weight, and bias terms and applies a nonlinear activation function.
  • 3.1. Deep Multi Layer Perceptron (DMLP): DMLPs achieve more efficient classification and regression than shallow networks through multilayer learning.Their learning uses backpropagation, with output errors reflected to preceding layers.
  • 3.1. Deep Multi Layer Perceptron (DMLP): DMLP optimization mostly uses Stochastic Gradient Descent to update connection weights.Backpropagation supplies the output-error signal used during learning.
  • 3.2. Convolutional Neural Networks (CNNs): CNNs are Deep Neural Networks mainly used for image classification and recognition by scanning images with filters.Common filter sizes are 1x1, 3x3, and 5x5.
  • 3.2. Convolutional Neural Networks (CNNs): CNN architectures commonly contain convolutional, pooling, and fully connected layers.The convolutional layers are based on the convolution operation.
  • 3.3. Recurrent Neural Network (RNN): RNNs process sequential data such as time series, audio, speech, and language using consecutively structured units and internal memory.Unlike feed-forward networks, they use memory to process incoming inputs.
  • 3.3. Recurrent Neural Network (RNN): RNNs process input sequences one element at a time, retain history in hidden-layer state vectors, and can be trained with Backpropagation Through Time.RNN structures include one-to-many, many-to-one, and many-to-many; BPTT reflects loss differentiation at time t through weights from previous times.

3.4. Long Short Term Memory (LSTM) · 3.5. Restricted Boltzmann Machines (RBMs) · 3.6. Deep Belief Networks (DBNs)

Sections 3.4–3.6 describe LSTMs for sequential data, RBMs for probabilistic representation learning, and DBNs as stacked RBMs for unsupervised feature discovery followed by supervised learning.

  • 3.4. Long Short Term Memory (LSTM): LSTMs retain both short-term and long-term values, making them suitable for sequential data analysis and time-series forecasting.They are composed of LSTM units.
  • 3.4. Long Short Term Memory (LSTM): Figure 4 presents the basic structure of an LSTM unit.
  • 3.5. Restricted Boltzmann Machines (RBMs): RBMs learn the probability distribution of an input set and are used for dimensionality reduction, classification, and feature learning.An RBM is a bipartite, undirected graphical model with visible and hidden layers.
  • 3.5. Restricted Boltzmann Machines (RBMs): RBM units within each layer are unconnected and make stochastic decisions about transmitting input data.Each unit functions as a computational point that processes the input.
  • 3.5. Restricted Boltzmann Machines (RBMs): Figure 5 illustrates the visible and hidden layers of an RBM.
  • 3.6. Deep Belief Networks (DBNs): DBNs consist of stacked RBM layers and use latent variables in a probabilistic generative model.They learn to reconstruct the input set probabilistically during training.
  • 3.6. Deep Belief Networks (DBNs): DBNs identify independent and discriminative input features through unsupervised learning, followed by supervised learning after the representation-learning step.The network layers begin detecting discriminative features after learning.
  • 3.6. Deep Belief Networks (DBNs): Figure 6 depicts a Deep Belief Network.

3.7. Autoencoders (AEs) · 3.8. Other Deep Structures

Autoencoders support unsupervised feature learning, feature extraction, and dimensionality reduction in financial deep-learning models, but may lose significant data relationships. Other deep structures are recognized in the literature, although DRL is the only one noted as having begun to attract financial applications.

  • 3.7. Autoencoders (AEs): Autoencoders remap input features into more representative representations through unsupervised feature learning.They learn a dataset representation by reducing dimensionality.
  • 3.7. Autoencoders (AEs): AEs have been used for feature extraction and dimensionality reduction.
  • 3.7. Autoencoders (AEs): Their symmetrical architecture uses equal-sized input and output layers with encoder and decoder components.
  • 3.7. Autoencoders (AEs): Dimensionality reduction and feature extraction can lose significant data relationships when the AE code minimizes relationship-loss.
  • 3.8. Other Deep Structures: Other well-known deep structures include Deep Reinforcement Learning, Generative Adversarial Networks, Capsule Networks, and Deep Gaussian Processes.
  • 3.8. Other Deep Structures: The authors found no noteworthy academic or industrial financial-application publications for these structures so far, except that DRL had started gaining attention.
  • 3.8. Other Deep Structures: Ongoing deep-learning model development continues to introduce new structures, while the aforementioned models cover almost all published work.

4. Financial Applications

This section categorizes deep-learning applications across financial subfields and summarizes representative studies by implementation area and model. It highlights algorithmic trading, risk assessment, and financial fraud as major application areas while noting overlaps between categories.

  • Section organization: The survey organizes financial DL implementations into separate application-area subsections and tabulates representative study features.Some papers are assigned to multiple areas because they address separate problems or fit multiple categories within the survey structure.
  • Algorithmic trading: Algorithmic trading studies commonly combine price or trend forecasting with buy-sell decisions, with LSTM the most preferred DL model and stock or index prediction the dominant focus.The survey separates forecasting-embedded, classification-based, and stand-alone algorithmic-trading studies across three tables.
  • Algorithmic trading: Algorithmic-trading research also covers high-frequency limit-order-book data, forex and cryptocurrency trading, reinforcement learning, and dynamic cryptocurrency portfolio construction.Examples include midprice prediction, GBP/USD and GBP/PLN trading, and portfolios spanning Bitcoin, Dash, Ripple, Monero, Litecoin, Dogecoin, Nxt, and Namecoin.
  • Risk assessment: Risk assessment applies DL to problems including bankruptcy prediction, credit scoring and evaluation, underwriting, bond rating, loan applications, and financial distress prediction.The survey presents snapshot information about risk-assessment studies using various DL models, and reports good DL performance for detecting stock-market crises.
  • Financial fraud: Financial fraud is another extensively studied finance application, encompassing credit-card fraud, money laundering, consumer-credit fraud, tax evasion, bank fraud, and insurance-claim fraud.Governments and authorities are described as seeking a permanent solution to these fraud cases.

Art. Data Set · 4.4. Portfolio Management · 4.5. Asset Pricing and Derivatives Market (options, futures, forward contracts)

The survey identifies portfolio management and asset pricing as financial application areas for deep learning, while noting that asset pricing had received only limited DL research. Derivative-product models, including options, futures, and forward contracts, were emerging targets for DL methods.

  • 4.5. Asset Pricing and Derivatives Market (options, futures, forward contracts): Asset pricing or valuation is a fundamental finance research area, but DL had not yet been applied extensively to it.The authors found only a handful of studies within the DL and finance community.
  • 4.5. Asset Pricing and Derivatives Market (options, futures, forward contracts): Options pricing, hedging strategy development, and financial engineering with options, futures, and forward contracts were identified as potential DL application areas.Recent studies indicated growing interest in DL solutions for this complex field.
  • 4.5. Asset Pricing and Derivatives Market (options, futures, forward contracts): Table 8 summarizes studies applying DL models to asset pricing and derivatives-market problems.The table covers the intended purposes of the identified studies.
  • 4.4. Portfolio Management: Iwasaki et al. used a DFNN and analyst reports for sentiment analysis to predict stock prices, followed by different portfolio selection approaches.The workflow combined text-derived sentiment information with portfolio construction after price prediction.
  • 4.5. Asset Pricing and Derivatives Market (options, futures, forward contracts): Culkin et al. proposed a feedforward DNN to predict option prices and compared its results with the Black & Scholes option pricing formula.The study is presented as a novel DL method for option pricing.
  • 4.5. Asset Pricing and Derivatives Market (options, futures, forward contracts): Hsu et al. proposed a method for predicting TAIEX option prices using bid-ask spreads.The supplied passage introduces this study but does not provide its complete inputs or results.

5. Current Snaphot of DL research for Financial Applications

The survey reviewed 144 papers to summarize the current state of deep learning research across financial applications. Financial text mining and algorithmic trading were the most studied topics, while RNN, DMLP, and CNN dominated model use, with CNN gaining recent attention.

  • 5. Current Snaphot of DL research for Financial Applications: The survey reviewed 144 papers across financial application areas, analyzing topics, publication types, problem types, methods, datasets, feature sets, and performance criteria.The authors provide general summary statistics because of space limitations.
  • 5. Current Snaphot of DL research for Financial Applications: Financial text mining and algorithmic trading were the two most researched fields, followed by risk assessment, sentiment analysis, portfolio management, and fraud detection.Most papers were published within the last 3 years, indicating active research interest.
  • 5. Current Snaphot of DL research for Financial Applications: RNN, DMLP, and CNN dominated the surveyed model types, with LSTM comprising more than 70% of RNN papers.LSTM was especially popular for time series forecasting, regression, and algorithmic trading.
  • 5. Current Snaphot of DL research for Financial Applications: DMLP was commonly used for classification problems and across most financial application areas, partly because it has a longer history than other deep learning models.Its longer history reflects its development as an extension of the shallow MLP.
  • 5. Current Snaphot of DL research for Financial Applications: CNN research increased within the last 3 years, alongside a growing trend of representing financial data as 2-D images for CNN use.CNN passed DMLP during the last 3 years and might surpass other models in the future.

6. Discussion and Open Issues

LSTM remains the preferred model for financial time-series forecasting, while CNN, DRL, and hybrid approaches are gaining traction across applications. DL models generally outperform ML counterparts, and expanding research interest leaves substantial opportunities in emerging and underexplored financial areas.

  • Model trends: LSTM remains the dominant and preferred DL model for financial time-series forecasting, whereas CNN and DMLP are better suited to classification applications.LSTM fits time-varying financial data, while CNN is more suitable for static representations and classification.
  • Model trends: DRL implementations, particularly those coupled with agent-based modelling, are attracting rising interest, especially for algorithmic trading.The approach can also be developed for other problem types.
  • Model trends: Hybrid models are preferred over native models in most studies for better performance, although increasing complexity can make them difficult to build and interpret.Cascaded CNN-LSTM blocks may combine spatial and temporal information, while ensemble models integrating EC and DL offer another research direction.
  • Performance and adoption: DL models generally outperform ML counterparts on the same problems, handle larger amounts of data, and continue improving relatively each year.Improvements are most notable in trend-prediction algorithmic trading and text-mining studies, although their magnitude is difficult to quantify explicitly.
  • Application trends: Price/trend prediction, algorithmic trading, risk assessment, portfolio management, text mining, sentiment analysis, and fraud detection are among the most studied financial applications.Financial text mining is receiving particularly strong attention through integration of numerical and textual data.
  • Open issues: Cryptocurrencies, blockchain, behavioral finance, HFT, and derivatives markets offer promising research opportunities, while NLP, semantics, and new ensemble models may improve existing systems.Behavioral finance remains relatively untouched, and derivatives-market DL research is scarce despite their practical flexibility.

7. Conclusions

Deep learning is gaining traction across financial applications in academia and industry, with research increasing rapidly. The survey presents a state-of-the-art snapshot while emphasizing that the field remains in its early years and will continue expanding.

  • 7. Conclusions: Deep learning’s potential is increasingly recognized across financial application areas by both academia and industry.The passage describes growing realization of DL’s potential in the financial industry and academia.
  • 7. Conclusions: Research on deep learning for finance is increasing every year at an accelerated pace.The passage states that the number of research works keeps increasing annually in an accelerated fashion.
  • 7. Conclusions: The field remains in its early years, with further studies and new models expected to emerge.The passage characterizes this as a new era and anticipates additional studies and models.
  • 7. Conclusions: The survey aims to highlight state-of-the-art deep learning research in financial applications and provide a snapshot of the existing research status.These aims are stated directly in the conclusion passage.

Glossary

The glossary defines abbreviations spanning artificial intelligence, machine learning, deep learning, neural-network architectures, financial markets, indicators, datasets, and evaluation measures used throughout the survey.

  • Core concepts: AI, ML, and DL denote Artificial Intelligence, Machine Learning, and Deep Learning, respectively.These abbreviations recur across the survey’s cited sections.
  • Deep-learning models: Key deep-learning architectures include ANN, CNN, DBN, DFFN, GAN, GRU, LSTM, SAE, and TDNN.The glossary expands these as Artificial Neural Network, Convolutional Neural Network, Deep Belief Network, Deep Feed Forward Network, Generative Adversarial Network, Gated-Recurrent Unit, Long-Short Term Memory, Stacked Autoencoder, and Timedelay Neural Network.
  • Financial markets and data: Financial instruments, markets, and datasets include AMEX, BIST, DAX, DJI, DJIA, HFT, NASDAQ, NYSE, S&P500, SPY, and TAQ.The glossary identifies these as exchanges, indexes, trading activity, an exchange-traded fund, and Trade and Quote data.
  • Methods and indicators: Technical-analysis and statistical methods include ARIMA, Bollinger Band, PCA, RSI, SVM, SVR, VAR, and XGBoost.These abbreviations cover forecasting, indicators, dimensionality reduction, machine-learning models, vector autoregression, and gradient boosting.
  • Evaluation measures: Evaluation measures include AUC, AUROC, BA, G-mean, MCC, MDD, ROA, RSE, SR, STD, and WBA.The glossary expands these as classification, correlation, financial-performance, error, risk, volatility, and weighted-accuracy measures.
Loading 2002.05786v1…