Applications of deep learning in stock market prediction: recent progress

Scope and motivation

• Deep learning models have shown better performance than linear and machine learning models on stock market prediction tasks by learning nonlinear relationships from big data.The survey focuses on deep neural network structures and their application to stock market prediction.Source: §1 · p. 2

Survey coverage

• The survey summarizes recent deep learning progress, especially work from the past three years, and presents prediction-workflow trends to help newcomers avoid obsolete technologies.The workflow provides a basis for classifying previous studies and guiding future work at each step.Source: §1 · pp. 3–4

Scope and limitations

• Although focused on stock market applications rather than all financial problems, the findings may also inform exchange-rate and cryptocurrency time-series prediction.The broader applicability is presented as insight for other finance-related prediction problems, not as a coverage objective.Source: §1 · p. 3

Implementation and reproducibility

• It gives special attention to implementation and reproducibility, including open data and code that enable validity checks, baseline implementation, and fair comparisons on shared datasets.This emphasis is identified as often neglected in similar surveys.Source: §1 · pp. 3–4

Future directions

• The survey identifies future research directions to help readers follow research frontiers and choose subsequent research movements.Some highlighted directions are described as ongoing.Source: §1 · pp. 3–4

Earlier reviews

• Prior surveys review neural and neuro-fuzzy forecasting, artificial neural networks, text mining, evolutionary computation, financial-data processing, and neural-network applications in business.These works address stock prediction alongside options, exchange rates, banking, financial crises, and other financial problems.Source: §2 · pp. 4–5

Recent reviews

• Recent reviews broaden coverage to NLP-based financial forecasting, quantitative finance, fundamental and technical analysis, and financial instruments, while identifying challenges in live testing, long-term prediction, and sentiment analysis.Nti et al. (2019) finds support vector machine and artificial neural network are the most used techniques for stock market prediction.Source: §2 · p. 5

Limitations

• Existing literature has problems including unsuitable benchmarks, short evaluation periods, and nonoperational trading strategies.Reschenhofer et al. (2019) emphasizes economic significance while identifying these limitations.Source: §2 · p. 6

Motivation

• The review treats stock-market forecasting as distinct from general time-series prediction by considering profitability alongside prediction accuracy.The motivation is to track deep-learning techniques as historical prices and diverse inputs such as financial news and Twitter accumulate.Source: §2 · pp. 6–7

Review scope

• This review focuses on deep-learning research from 2017–2019 for stock price and market-index prediction, rather than Sezer et al. (2019)’s broader financial time-series scope from 2005 to 2019.It also emphasizes implementation workflow, dataset and code availability, and reproducibility of previous studies.Source: §2 · p. 6

Corpus and sources

• The review covers 124 papers collected from Google Scholar, including 56 journal papers, 58 conference papers, and 10 arXiv preprints.Most covered papers, 115 of 124, were published during 2017–2019; arXiv papers were included to keep the survey updated.Source: §3 · p. 7

Prediction scope

• The survey focuses on predicting individual-stock and market-index close prices, also covering ETFs, equity-index futures, and intraday limit-order-book mid-prices.Bond and cryptocurrency prices are excluded, while exact-price targets are treated as regression and directional targets as classification.Source: §3 · pp. 7–8

Prediction frequency

• 105 of 124 papers study daily prediction, 18 of 124 study intraday prediction, and only one paper considers both frequencies (Liu & Wang,Intraday studies include five-minute or hourly prediction settings.Source: §3 · p. 9

Problem types

• 52 of 124 papers address daily classification, 54 of 124 daily regression, 8 of 124 intraday classification, and 11 of 124 intraday regression.The paper count for these four problem types is summarized in Figure 1.Source: §3 · p. 9, Figure 1 · p. 9

Surveyed markets

• The survey spans mature and emerging markets, with most studies evaluating one market and some evaluating multiple markets.Daily prices and news titles are easier to collect than intraday data, which is limited in academic research.Source: §3 · pp. 9–10

Prediction Workflow

• Most studies follow a four-step prediction workflow comprising Raw Data, Data Processing, Prediction Model, and Model Evaluation.The workflow is presented as a general approach intended to support easy reproduction in future work.Source: §4 · p. 12

4.1.1. Data Types

• The survey categorizes stock-prediction inputs into seven raw-data types, spanning market, text, macroeconomic, knowledge-graph, image, fundamental, and analytics data.These sources differ in availability, processing difficulty, and suitability for deep learning.Source: §4.1.1 · pp. 13–14
• Market data is used most because it provides the largest sample volume, while text data is second-most common and analytics data is absent from surveyed studies because of sparsity and access costs.Figure 3 summarizes usage across raw-data types.Source: §4.1.1 · pp. 14–15
• Studies used more diverse data types in 2018 and 2019 than in 2017, indicating that market data alone made better prediction harder and that new tools, such as graph neural networks, enabled additional data use.Source: §4.1.1 · p. 15

4.1.2. Data Length

• Data length requires balancing short-period overfitting against long-period exposure to changing market styles and outdated results, while also considering availability and cost.Intraday studies generally use less than one year because high-quality intraday data is more expensive; Figure 4 shows the surveyed distribution.Source: §4.1.2 · p. 16
• Most studies target short horizons such as one day or five minutes, with few predicting five or ten days; technical-indicator lags range from 2 to 252 periods.A lag defines the historical input window, whereas horizon defines the future period being predicted.Source: §4.1.2 · pp. 16–17

4.2.1. Missing Data Imputation

• Lower-frequency data are propagated forward from the last valid observation to align sampling frequencies while avoiding future-information leakage.Market data missingness is described as less severe than in domains such as sensor data because market data are relatively reliable and well maintained.Source: §4.2.1 · p. 17

4.2.2. Denoising

• Wavelet transforms and kNN classifiers are used to reduce market-data noise that can misrepresent price trends and misguide prediction.Sun et al. (2017) applies kNN classifiers using two training sets with different labels in a data-preparation layer, while Bao et al. (2017) and Liang et al. (2019) use wavelet transforms.Source: §4.2.2 · p. 17

4.2.3. Feature Extraction

• Historical prices and technical indicators are the most commonly used input features, followed by text and macroeconomic data, as shown in Figure 5 and Table 4.Feature engineering combines domain-informed handcrafted features with raw data, while text processing uses embeddings, event and sentiment representations, and other neural or linguistic methods; extracted features can substantially boost model performance.Source: §4.2.3 · pp. 19–20, Figure 5 · p. 19, Table 4 · p. 20
• Feature extraction combines historical prices with technical indicators, fundamentals, knowledge sources, graphs, text, macroeconomic data, or images across surveyed studies.The listed combinations include historical prices with technical indicators, fundamentals, knowledge graphs, text, macroeconomic variables, graph data, and images.Source: §4.2.3 · p. 22

4.2.4. Dimensionality Reduction

• PCA-ANN gives slightly higher next-day SPY directional accuracy than FRPCA and KPCA, while other studies use diverse reduction and feature-selection methods.Dimensionality reduction is used to mitigate correlated technical indicators and deep-learning overfitting; alternatives include ICA, autoencoders, RBMs, EMD, SMC, chi-square, and MRMR.Source: §4.2.4 · pp. 22–23

4.2.5. Feature Normalization & Standardization

• Feature normalization maps values between 0 and 1 or between -1 and 1, while standardization subtracts a location measure and divides by a scale measure, helping training speed and performance.The z-score method subtracts the mean and divides by a measure of scale.Source: §4.2.5 · p. 24

4.2.6. Data Split

• Stock prediction commonly uses train/validation/test or in-sample/out-of-sample splits, with rolling or successive windows adapting these schemes to time-series evaluation.Training fits the model, validation optionally tunes hyperparameters, and testing evaluates final performance; k-fold cross-validation is another option.Source: §4.2.6 · pp. 24–25, Figure 6 · p. 25

4.2.7. Data Augmentation

• Data augmentation is less used for time-series tasks, but some studies enlarge training data by clustering different stocks or using highly correlated companies to augment samples.Source: §4.2.7 · pp. 25–26
• Data augmentation combines 10 companies taken 5 at a time to help prevent overfitting.Source: §4.2.7 · p. 26
• An LSTM module processes the data samples before they are fed into a prediction LSTM module for stock market index prediction.Source: §4.2.7 · p. 26

Prediction Model

• Most prediction models use supervised learning with separate training and test sets, while only a few studies use semi-supervised learning when feature-extraction labels are unavailable.The survey classifies prediction models into standard models and variants, hybrid models, and other models.Source: §4.3 · p. 26
• Standard deep-learning models commonly use feedforward, convolutional, and recurrent neural networks, whereas generative adversarial, transfer, and reinforcement learning are categorized as other models.The latter approaches appeared only in recent years and remain at an early stage of application to stock-market prediction.Source: §4.3 · p. 26
• The survey defines ANN as a neural network with one or zero hidden layers and DNN as one with two or more hidden layers.BPNN, MLP, ELM, IDLNN, STEFNN, and RBFN are grouped with deep neural networks because they share similar structures.Source: §4.3 · p. 31
• FFNNs are the simplest artificial neural networks because their node connections do not form cycles, with neurons aggregating inputs and applying activation functions.Autoencoders are described as ANN subsets with equal-sized input and output layers.Source: §4.3 · p. 31
• The standard-model taxonomy also includes BPNN, MLP, ELM, IDLNN, STEFNN, and RBFN variants alongside FFNNs.ELM need not tune hidden-node parameters, IDLNN uses increasing-decreasing-linear processing units, and RBFN uses radial basis functions as activations.Source: §4.3 · pp. 31–32

Convolutional neural network (CNN)

• CNNs use shared-weight filters and pooling to reduce parameters and input size, while 1D CNNs adapt these operations to time-series prediction.Deng et al. (2019) uses a 1-D fully convolutional network with equal hidden-layer and input lengths maintained through zero padding.Source: §4.3 · p. 32

Recurrent neural network (RNN)

• RNNs model temporal dynamics but standard forms suffer vanishing gradients; LSTM and GRU address this limitation, with GRU using fewer parameters than LSTM.Bidirectional LSTM and bidirectional GRU connect recurrent representations in opposite temporal directions and have been used for stock market prediction.Source: §4.3 · p. 32

Model variants

• Deep prediction models extend standard architectures through stacked ensembles and attention mechanisms that bias alignment over inputs.Attention is described as a generalized pooling method introduced into recurrent neural network models.Source: §4.3 · p. 33

Other deep models

• RBMs and stacked DBNs learn input probability distributions, while seq2seq models use recurrent encoder-decoder architectures for sequence prediction; both families appear in stock-prediction studies.DBNs are reported in Li et al. (2017b) and Karathanasopoulos & Osman (2019), while seq2seq is reported in Liu & Wang (2019).Source: §4.3 · p. 33

Baselines and trading strategies

• Linear and traditional machine-learning models commonly serve as comparison baselines, alongside Buy&Hold and indicator-based strategies such as momentum, MACD, RSI, and SMA.Long-short trading goes long when predicted direction rises and short otherwise; Buy&Hold purchases at the start and holds through testing without further trades.Source: §4.3 · pp. 33, 35

Prediction Model

• Cited studies include Chen et al. (2017) with a DNN and Althelaya et al. (2018) with stacked LSTM architectures, while recurrent neural networks are represented as a type.Source: §4.3 · pp. 36–38

Hybrid Models

• Hybrid models are categorized into combinations of deep-learning models with traditional models and combinations of different deep-learning models.The survey presents the first category in Table 7 and distinguishes it from deep-learning-only hybrids.Source: §4.3 · pp. 40–41, Table 7 · p. 41
• Li et al. (2019b) combine sentiment analysis, ARMA, LSTM, VADER, and differential privacy to integrate information from multiple news sources in a hybrid DNN.The model uses news articles as hidden information and includes LSTM, VADER, and a differential privacy mechanism.Source: §4.3 · p. 41
• Under noisy conditions, Liu & Song (2017) combine weak ANNs through optimized bagging, while Wu & Gao (2018) use AdaBoost to ensemble multiple LSTM predictors.Liu & Song (2017) aim to avoid over-fitting, whereas Wu & Gao (2018) generate training samples and ensemble weights for each LSTM predictor.Source: §4.3 · p. 41

Hybrid models

• Hybrid models combine CNNs with RNNs or different RNN architectures to capture multi-scale features, temporal dependencies, and complementary information for stock prediction.Examples include TreNet, convolution-Inception-LSTM models, convolutional recurrent models, stacked LSTM autoencoders, and attention-based recurrent hybrids.Source: §4.3 · p. 42

Other model types

• Other promising approaches include GANs, including Zhou et al. (2018)’s generic GAN employing LSTM and CNN for adversarial training to predict high-frequency stock market data, and GNNs for graph-structured data.Source: §4.3 · p. 45
• Capsule networks were introduced for stock-movement prediction from social media and were reported as effective for this task.Liu et al. (2019a) introduced the capsule-network model for stock-movement prediction.Source: §4.3 · p. 45
• Unlike supervised learning, reinforcement learning trains an agent to choose the optimal action given a current state, maximizing cumulative rewards while using information from the next and all subsequent time steps for stock prediction.Source: §4.3 · pp. 45–46

Model usage trends

• RNN models were used most frequently in the surveyed papers, but their proportion declined in 2019 as new model types emerged.The trend is shown in Figure 7.Source: §4.3 · p. 47

Prediction Model

• Adam is the most frequently used optimizer for stock prediction.It combines RMSprop with stochastic gradient descent and momentum, offering computational efficiency and low memory requirements.Source: §4.3 · p. 47
• Survey studies use linear, machine-learning, and deep-learning models as baselines.The baseline-model composition changes over time, as shown in the model-usage figure.Source: §4.3 · pp. 47–48, Figure 7 · p. 48
• The proportion of deep-learning models used as stock-prediction baselines has increased over the past three years.This trend accompanies further exploration of deep-learning models for stock prediction.Source: §4.3 · p. 48, Figure 7 · p. 48

Classification metrics

• Classification metrics evaluate movement prediction using accuracy, precision, recall, sensitivity, specificity, F1 score, AUC, Theil’s U, hit ratio, and related measures.Confusion matrices and boxplots for daily accuracy also support classification-performance analysis.Source: §4.4 · p. 48

Regression metrics

• Regression metrics assess stock or index price prediction with MAE, RMAE, MSE, NMSE, RMSE, relative RMSE, NRMSE, MAPE, RMSRE, mutual information, and R2.R2 is defined as the coefficient of determination.Source: §4.4 · p. 49

Profit Analysis

• Profit analysis evaluates prediction-based trading strategies through portfolio return and risk, including maximum drawdown, annualized volatility, and Sharpe Ratio.Maximum drawdown represents the largest peak-to-trough decline, while Sharpe Ratio combines excess return with volatility.Source: §4.4 · p. 50

Metric usage catalog

• Table 10 provides a detailed list of studies using each evaluation metric, including metric abbreviations.The catalog includes classification measures such as Accuracy, Precision, Recall, Sensitivity, Specificity, F1 score, Macro-average F-score, Matthews Correlation Coefficient, Average AUC Score, Theil’s U Coefficient, and Hit Ratio.Source: §4.4 · p. 51, Table 10 · p. 51

• Studies also evaluate predictions with normalized and relative errors, R2, returns, drawdown, volatility, Sharpe ratio, and statistical tests.The listed measures include NMSE, rRMSE, NRMSE, Mutual Information (MUL), Maximum Drawdown, Annualized Volatility, Sharpe Ratio, Kruskal-Wallis Test, and Diebold-Mariano Test.Source: §4.4 · pp. 55–57
• Accuracy and F1 score are the most often used classification metrics, followed by precision, recall, and MCC.The review presents their changing usage in Figure 10 and Figure 11.Source: §4.4 · p. 58, Figure 10 · p. 58
• RMSE and MAPE are the most often used regression metrics, followed by MAE and MSE.The review presents their changing usage in Figure 10 and Figure 11.Source: §4.4 · p. 58

5.1. Implementation

• Keras and TensorFlow are the dominant frameworks, while Python has become the dominant programming-language choice for deep-learning stock-prediction research in the past three years.Other surveyed implementations use PyTorch, Theano, scikit-learn, R, Matlab, Java, and Node.js; Theano is discontinued and not recommended for further use.Source: §5.1 · pp. 58–59
• Deep-learning stock-prediction models require substantial computation, with GPUs used to accelerate convolutional operations and demand expected to increase as studies process more text data.Cloud computing offers an alternative when local GPUs are unavailable, but surveyed studies rarely adopted it and none explicitly mentioned using a cloud service.Source: §5.1 · p. 62

5.2. Result Reproducibility

• The survey assesses result reproducibility through data and code availability because prior machine-learning studies have raised reproducibility concerns.The authors do not verify every paper’s results, instead focusing on these two aspects as important conditions for reproducibility; shared resources can also serve as baselines for later work.Source: §5.2 · p. 63

5.2.1. Data Availability

• Researchers can draw on free sources such as Yahoo! Finance, which appears in at least 25 of 124 papers, alongside Tushare, official market sites, IMF, World Bank, news and social-media APIs, Wikidata, and financial databases.Commercial providers and databases include Bloomberg, Wind, Quantopian, Investing.com, Interactive Brokers, CSMAR, and WRDS, covering prices, returns, statements, macroeconomic indicators, news, social data, and company relations.Source: §5.2.1 · pp. 63–64
• Kaggle and collaborations with quantitative companies are presented as additional repositories and competition-based sources for stock market prediction data.The Two Sigma Financial Modeling Challenge is cited as an example organized by the hedge fund Two Sigma.Source: §5.2.1 · p. 65
• Although most data sources are available on the Internet, releasing the exact datasets used would make replication more convenient, complementing the public data links listed in Table 13.For mid-price prediction of limit-order-book data, a benchmark dataset provided by Ntakaris et al. (2017) has been used.Source: §5.2.1 · p. 65, Table 13 · p. 65
• Publicly available datasets span stock prices, news, social media, and company relations across multiple markets, periods, frequencies, and prediction tasks.Examples include Qin et al. (2017), Xu & Cohen (2018), Wu et al. (2018), and Feng et al. (2019b).Source: §5.2.1 · pp. 66–67

5.2.2. Code Availability

• Only a small number of stock market prediction studies currently release their source code, despite GitHub being mainstream for hosting computer-science code.Table 14 lists articles with public repositories and provides short method descriptions, with further details in Section 4 and the original documents.Source: §5.2.2 · p. 68

6.1. New Models

• New neural architectures remain underexplored for stock prediction, including Transformers and pre-trained BERT for financial news analysis (Devlin et al., 2018).Attention mechanisms and generative adversarial networks have been covered, but substantial room remains for additional models.Source: §6.1 · pp. 68–69

6.2. Multiple Data Sources

• Future systems should combine multiple, especially less-explored, data sources rather than relying on heavily used single sources such as market data (Zhou et al.,The survey argues that outperforming existing solutions using only one familiar source is challenging.Source: §6.2 · p. 70

6.3. Cross-market Analysis

• Although stock markets differ because of trading rules, different markets may share phenomena that can be leveraged for prediction through transfer learning, and a few studies show positive cross-market results.Source: §6.3 · p. 70

6.4. Algorithmic Trading

• Practical algorithmic trading requires realistic strategies that account for trading rules, transaction costs, market-style adaptation, and time-consuming model training.Deep reinforcement learning is proposed because it can simulate more cases and make faster, better trading choices than humans (Xiong et al., 2018; Lee et al., 2019).Source: §6.4 · pp. 70–71

• The survey examines more than 100 published articles on deep-learning models for stock-market prediction from the past three years.It was motivated by the rapid development and increasing use of deep learning in this field.Source: §7 · p. 71
• It covers the full workflow from raw-data collection and processing through prediction modeling and model evaluation.Source: §7 · p. 71
• By presenting research trends from 2017 to 2019 and outlining future directions, the survey aims to boost future research in related topics.Source: §7 · p. 71
• The review emphasizes deep-learning implementation and reproducibility to accelerate adoption of published models as baselines, potentially with new data inputs.Source: §7 · p. 71