Deep Learning for Financial Applications : A Survey

1. Introduction

• The survey asks which financial areas use DL, how mature or promising they are, which models are preferred or successful, and how they compare with traditional methods.It also asks about future directions for DL research in finance.Source: §1 · pp. 1–2
• The survey focuses on DL implementations for finance, excluding financial time-series forecasting studies covered in a separate survey while including DL-based algorithmic trading with or without forecasting.Source: §1 · p. 2
• The review spans journals, conferences, theses, book chapters, arXiv papers, and noteworthy technical papers, includes English-language works, and also covers older models such as RNNs and Jordan-Elman networks.The older models were included even when they were not called DL at publication.Source: §1 · p. 2
• To the authors’ knowledge, this is the first comprehensive survey focused on DL implementations in financial applications, intended to highlight advances and guide researchers developing models across application areas.Source: §1 · p. 2
• The paper reviews prior finance ML surveys and basic DL models, then covers financial implementations, analyzes publication and field statistics, and discusses achievements, open areas, and recommended research topics.Source: §1 · pp. 2–3

• Finance has been a major ML application area for about 40 years, with thousands of studies; this survey focuses on DL and reviews earlier ML surveys for context.Source: §2 · p. 3
• Broad-perspective surveys cover financial methods and applications, including ANNs, expert systems, data mining, soft computing, banking, econometrics, and Fintech –; books cover soft computing.Source: §2 · p. 3
• Some surveys are not application-area-specific but focus on particular ML techniques; evolutionary algorithms are commonly used in financial optimization, including Portfolio Selection, and later surveys cover MOEAs in portfolio optimization and other financial applications,,.Chen et al. wrote a book covering GAs and GP in Computational Finance.Source: §2 · p. 3
• ANN-focused surveys cover early financial implementations, stock-price forecasting and other applications, and economics and management research.Source: §2 · p. 3
• Other reviews cover agent-based computational finance and kernel methods, including PCA and SVM; application-specific surveys are discussed in Section 4, before the next section introduces DL models.Source: §2 · pp. 3–4

3. Deep Learning

• Deep learning comprises multiple ANN layers that provide high-level abstraction for data modelling, with DMLP, CNN, RNN, LSTM, RBM, DBN, and autoencoders among its model types.Source: §3 · p. 4

3.1. Deep Multi Layer Perceptron (DMLP)

• DMLP extends an ordinary MLP with more layers, where neurons combine weighted inputs and bias before applying a nonlinear activation function.Common activations include sigmoid, hyperbolic tangent, ReLU, leaky ReLU, swish [26], and softmax [27].Source: §3.1 · p. 4
• DMLP networks contain input, hidden, and output layers, with weights connecting neurons across layers, as shown in Figure 1.Source: §3.1 · p. 4, Figure 1 · p. 5
• Multi-layer deep ANNs achieve more efficient classification and regression performance than shallow nets, and DMLP learning uses backpropagation [21].Output-layer error is reflected back to preceding layers; SGD is mostly used to update connection weights, as illustrated in Figure 1.Source: §3.1 · p. 4, Figure 1 · p. 5

3.2. Convolutional Neural Networks (CNNs)

• CNNs are deep neural networks mostly used for image classification and recognition, scanning images with filters that commonly measure 1x1, 3x3, or 5x5.Source: §3.2 · p. 5
• Common CNN architectures use convolutional, pooling, and fully connected layers; pooling may use average or maximum operations.Figure 2 presents a generalized architecture with convolutional, subsampling (pooling), and fully connected layers.Source: §3.2 · p. 5, Figure 2 · p. 5

3.3. Recurrent Neural Network (RNN)

• RNNs are used mainly for sequential data, including time series, audio, speech, and language, and use internal memory to process incoming inputs.The passage also notes their use in handwriting and speech recognition.Source: §3.3 · p. 6
• RNNs process inputs sequentially, with hidden-layer state vectors retaining input history, and support one-to-many, many-to-one, and many-to-many structures.Source: §3.3 · p. 6
• RNNs can be trained with Backpropagation Through Time, which makes the loss gradient at time t reflect weights from earlier times; training is harder and longer than for FFNNs.Source: §3.3 · p. 6
• The RNN illustration shows hidden-layer information across discrete time steps, with state, input, and output values indexed by time and parameters U, W, and V reused at each step [21].Source: §3.3 · p. 6, Figure 3 · p. 6

3.4. Long Short Term Memory (LSTM)

• LSTM cells use input, output, and forget gates to regulate information flow and retain selected values over arbitrary time intervals.Cells combine into neural-network layers; Figure 4 illustrates a basic LSTM unit.Source: §3.4 · p. 6, Figure 4 · p. 7
• LSTMs are intended for sequential-data analysis and are used for speech recognition, language translation, handwritten-character recognition, and time-series forecasting.Source: §3.4 · p. 6

3.5. Restricted Boltzmann Machines (RBMs)

• RBMs learn input distributions and are used for dimensionality reduction, classification, and feature learning.Their units make stochastic decisions about transmitting input data.Source: §3.5 · p. 7
• An RBM has visible and hidden layers in a bipartite, undirected graph, with no connections among units within either layer.Reconstruction outputs re-enter the network, and repeated comparisons with the previous input aim to reduce their difference.Source: §3.5 · p. 7, Figure 5 · p. 8
• RBM training is tricky because no known cheap method estimates the log-likelihood itself, despite good estimators of its gradient.Source: §3.5 · p. 7

3.6. Deep Belief Networks (DBNs)

• A DBN is a probabilistic generative model with latent variables, formed by stacking RBM layers.Figure 6 illustrates its structure.Source: §3.6 · p. 8, Figure 6 · p. 8
• DBNs use an unsupervised approach to find independent and discriminative features in the input set [31].Source: §3.6 · p. 8
• During training, a DBN probabilistically reconstructs the input set; after learning, supervised learning performs classification.Source: §3.6 · p. 8

3.7. Autoencoders (AEs)

• AEs learn more representative input features through unsupervised feature learning and are used for feature extraction and dimensionality reduction.Source: §3.7 · p. 8
• AEs have symmetric input and output layers, with one or more hidden layers connecting them and encoder and decoder components.Figure 7 shows the basic AE structure.Source: §3.7 · pp. 8–9, Figure 7 · p. 9
• AE dimensionality reduction and feature extraction can lose significant data relationships when minimizing loss in the code [33].Source: §3.7 · p. 9

3.8. Other Deep Structures

• Other structures discussed include DRL, GANs, Capsule Networks, and DGPs.Source: §3.8 · p. 9
• The survey found no noteworthy financial applications using these models other than DRL, which has started gaining attention.Source: §3.8 · p. 9
• Despite limited published financial applications, these structures may suit finance, and ongoing DL research continues to produce new structures.The authors state that the models discussed cover almost all published work to date.Source: §3.8 · p. 9

4.1. Algorithmic Trading

• Algorithmic-trading DL spans forecast-linked, buy/sell or trend-classification, and standalone strategies; most studies target stock or index prices, with LSTM the most preferred model.Studies also address HFT and forex or cryptocurrency trading, including dynamic cryptocurrency portfolio construction.Source: §4.1 · pp. 10–12

4.2. Risk Assessment

• Risk-assessment DL studies chiefly address credit scoring and bank distress, with additional work on bankruptcy, mortgage and transaction risk, and crisis detection.One bankruptcy study found DBN outperformed SVM, while DL models showed good performance in classifying whether crash-causing events occurred.Source: §4.2 · pp. 15, 17, 18

4.3. Fraud Detection

• Fraud-detection DL studies are generally framed as anomaly detection or classification and cover card and payment fraud, money laundering, insurance fraud, and tax-related cases.Examples include LSTM applied to credit-card transaction sequences and deep autoencoders used to detect anomalies in export tax claims and parliamentary spending.Source: §4.3 · p. 18

4.5. Asset Pricing and Derivatives Market (options, futures, forward contracts)

• DL had not been applied broadly to asset pricing, leaving substantial opportunities for future research in pricing and valuation.The survey reports only a handful of studies in this area.Source: §4.5 · p. 21
• Options, hedging, financial engineering, futures, and forward contracts were identified as potential DL applications, with recent studies signaling growing interest.The survey characterizes the field as complex and challenging.Source: §4.5 · p. 21
• Iwasaki et al. used analyst-report sentiment and a DFNN for stock-price prediction, then applied different portfolio-selection approaches.Source: §4.5 · p. 22
• Culkin et al. predicted option prices with a feedforward DNN against Black & Scholes, while Hsu et al. used a 3-layer DMLP for TAIEX options.Hsu et al. used bid-ask spreads and Black & Scholes pricing-model parameters.Source: §4.5 · p. 22
• A Fama-French n-factor DL model used characteristic features to predict US equity returns across NASDAQ, AMEX, and NYSE indices [144].Source: §4.5 · p. 22

• The survey reviewed 144 papers, analyzing each by topic, publication type, problem, method, dataset, features, and performance criteria.Because of space limitations, the section reports general summary statistics rather than analyses of every paper.Source: §5 · p. 31
• Financial text mining and algorithmic trading were the most-studied topics, followed by risk assessment, sentiment analysis, portfolio management, and fraud detection.Most papers appeared in the last three years; the few published before 2013 all used RNN-based models.Source: §5 · p. 31, Figure 8 · p. 31, Figure 9 · p. 32
• RNN, DMLP, and CNN dominated model choices, and more than 70% of RNN papers used LSTM models.LSTM is popular in time-series forecasting and regression, and is also used in algorithmic trading.Source: §5 · pp. 32–33, Figure 10 · p. 32
• DMLP commonly serves financial classification problems and is a common choice across application areas, while its longer history reflects its extension of shallow MLP.Source: §5 · p. 34
• CNN use has grown in the last three years, alongside a trend of representing financial data as 2-D images; CNN passed DMLP during that period and might surpass other models in the future.Source: §5 · pp. 34–35
• The journals with the most papers in the last three years included Expert Systems with Applications, Decision Support Systems, Applied Soft Computing, Neurocomputing, Knowledge-based Systems, and European Journal of Operational Research.Source: Figure 13 · p. 34, §5 · p. 35

6.2. Discussions on Implementation Areas

• Price and trend prediction, algo-trading, risk assessment, and portfolio management attract substantial interest, while financial text mining is gaining particular attention and cryptocurrency research is growing.Combining text mining with sentiment analysis and numerical data is presented as a promising direction for prediction and evaluation models.Source: §6.2 · p. 36

6.3.1. Model Development Perspective

• Further model-development opportunities include exploring 2-D CNNs, graph transformations, GANs, capsule networks, transfer learning, and one-shot learning for financial applications.The survey notes that graph-CNN applications are scarce and that no known financial implementations of GANs or capsule networks exist.Source: §6.3.1 · pp. 36–37
• Financial text models could be strengthened with Stock2Vec, NLP ensembles, and semantics, while hybrid architectures such as cascaded CNN-LSTM blocks could combine spatial and temporal information.The survey expects hybrid models to remain popular and suggests that evolutionary-computation and deep-learning ensembles could support portfolio decision systems.Source: §6.3.1 · p. 37

6.3.2. Implementation Perspective

• Implementation opportunities include cryptocurrencies and blockchain, robo-advisory and portfolio optimization, derivatives research, and low-latency HFT systems using GPU- or FPGA-based hardware.The survey identifies option strategies, futures trading, option pricing, and arbitrage as derivatives applications with little existing deep-learning research.Source: §6.3.2 · pp. 37–38

6.3.3. Suggestions for Future Research

• Behavioral finance remains largely untouched by deep-learning research, but text mining, NLP, semantics, and agent-based computational finance could help address the challenge of quantifying its inputs and outputs.The survey also anticipates increased attention to cryptocurrency and blockchain research.Source: §6.3.3 · p. 38

6.4. Responses to our Initial Research Questions

• The survey identifies financial text mining, algo-trading, risk assessment, sentiment analysis, portfolio management, and fraud detection as studied areas, while interest continues to rise across fields.It describes deep-learning results as better than traditional counterparts in almost all areas, while noting that overall research interest remains on the rise.Source: §6.4 · p. 39

6.1. Discussions on DL Models

• LSTM is preferred for financial time-series forecasting, CNN suits classification and static representations, and deep learning generally outperforms machine-learning counterparts, though hybrid models can be harder to build and interpret.The survey also notes rising interest in DRL, especially with agent-based modeling, and improving results in trend-prediction algo-trading and text mining.Source: §6.1 · p. 35, §6.4 · p. 39

6.4. Responses to our Initial Research Questions

• LSTM is preferred for time-series forecasting, while DMLP and CNN suit classification; the survey reports that deep learning generally outperforms machine-learning counterparts and expects greater importance for spatio-temporal, NLP, semantics, and text-mining hybrids.The survey lists cryptocurrencies, blockchain, behavioral finance, HFT, and derivatives as promising research areas.Source: §6.4 · p. 39

• The survey provides a snapshot of existing deep learning research in financial applications and identifies a future research roadmap.Source: §7 · p. 39
• Deep learning research in finance is growing rapidly, but the field remains in its early years, with more studies and models expected.Source: §7 · p. 39
• The survey finds substantial opportunities in the field and encourages interested researchers to explore them.Source: §7 · p. 39

• AI, ANN, ML, and DL are defined as Artificial Intelligence, Artificial Neural Network, Machine Learning, and Deep Learning.Source: Glossary · pp. 40–41
• Evaluation abbreviations include AUC, AUROC, Balanced Accuracy, Matthew Correlation Coefficient, Sharpe-ratio, and Standard Deviation.Source: Glossary · pp. 40–41
• Financial terms include CDS for Credit Default Swaps, ETF for Exchange-Traded Fund, and HFT for High Frequency Trading.Source: Glossary · pp. 40–41
• The glossary defines neural-network architectures including CNN, LSTM, GRU, and FFNN, alongside models such as GAN and DBN.Source: Glossary · pp. 40–41
• The glossary names market indices and exchanges such as DAX, NASDAQ, NYSE, and S&P500, with their full names.Source: Glossary · pp. 40–41