PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep learning models for price forecasting of financial time series: A review of recent advancements: 2020-2022

Cheng Zhang, Nilam Nur Amir Sjarif, Roslina Ibrahim

arXiv:2305.04811v2q-fin.STcs.AIcs.LG

TL;DR

Forecasting financial time-series prices is important but challenging. This paper reviews recent deep-learning research on the task; it finds that novel architectures have been adopted for price forecasting, which the authors present as evidence of benefits from deep-learning advances in finance.

  • Problem

    Financial time-series price forecasting is crucial yet challenging, and forecasts can inform financial decisions and risk management.

    Source: §1 · p. 2

  • Method

    The paper reviews recent literature on deep-learning models for financial time-series price forecasting.

    Source: §6 · p. 29

  • Results

    Transformers, GANs, GNNs, and DQNNs have been adopted in price-forecasting tasks.

    Source: §6 · p. 29

  • Takeaways & Limitations

    The authors conclude that these adoptions indicate that the financial sector benefits from recent advances in deep-learning techniques.

    Source: §6 · p. 29

    The authors acknowledge that the review has several limitations.

    Source: §6 · p. 29

Abstract

from arXiv · show

Accurately predicting the prices of financial time series is essential and challenging for the financial sector. Owing to recent advancements in deep learning techniques, deep learning models are gradually replacing traditional statistical and machine learning models as the first choice for price forecasting tasks. This shift in model selection has led to a notable rise in research related to applying deep learning models to price forecasting, resulting in a rapid accumulation of new knowledge. Therefore, we conducted a literature review of relevant studies over the past three years with a view to aiding researchers and practitioners in the field. This review delves deeply into deep learning-based forecasting models, presenting information on model architectures, practical applications, and their respective advantages and disadvantages. In particular, detailed information is provided on advanced models for price forecasting, such as Transformers, generative adversarial networks (GANs), graph neural networks (GNNs), and deep quantum neural networks (DQNNs). The present contribution also includes potential directions for future research, such as examining the effectiveness of deep learning models with complex structures for price forecasting, extending from point prediction to interval prediction using deep learning models, scrutinising the reliability and validity of decomposition ensembles, and exploring the influence of data volume on model performance.

1. INTRODUCTION

Financial time-series price forecasting is crucial yet challenging, and established approaches have limitations in modeling complex patterns. This review surveys 2020–2022 deep-learning advances, focusing on model architectures, applications, advantages, and disadvantages.

1. INTRODUCTION

  • Accurately forecasting financial time-series prices is challenging because the data are intricate and uncertain, despite the value of forecasts for financial decisions and risk management.Price forecasts can inform investors, traders, and financial institutions in managing risks and optimizing investment strategies.Source: §1 · p. 2
  • Statistical methods such as ARIMA and GARCH rely on predefined mathematical structures and assumptions that may miss nonlinear patterns or complex relationships in financial series.The cited discussion associates this limitation with real-world financial time series.Source: §1 · p. 2
  • ANNs and SVR can find patterns in historical data, but depend on feature engineering and can be bottlenecked by model complexity with numerous training data (Jiang, 2021).Source: §1 · p. 3
  • Research on deep-learning price forecasting has risen rapidly, while earlier reviews predominantly covered studies published before 2020 and may omit recent advances.The review responds to the resulting challenge of keeping current with knowledge accumulated over a short period.Source: §1 · p. 3
  • The review covers recent advances from 2020 to 2022, emphasizing forecasting-model details over data collection and preprocessing.Source: §1 · pp. 3–4
  • It treats similar target variables and prediction workflows alongside diverse model architectures, and summarizes their applications, advantages, and disadvantages.Source: §1 · p. 4
  • The review also proposes future research directions intended to contribute to long-term advances in the field.Source: §1 · p. 4

2. FORECASTING BACKGROUND INFORMATION

This section defines financial time series and price forecasting, then outlines the forecasting workflow and the relationship between model complexity and data availability.

2.1 Concept of financial time series

  • Financial series commonly track daily closing prices, while multivariate inputs can add technical or sentiment indicators as exogenous features.The reviewed background also identifies stock indices, individual stocks, commodities, cryptocurrencies, and forex rates as financial assets.Source: §2.1 · p. 5

2.2 Characteristics of financial time series

  • Volatility, nonlinearity, heteroscedasticity, autocorrelation, and nonstationarity challenge precise modeling and accurate forecasting.Traditional statistical methods assume stable connections between past and future values, an assumption that can fail for nonstationary series.Source: §2.2 · p. 5
  • Deep-learning forecasters need not require stationarity because they can discern irregular patterns embedded in the data.The background contrasts this capacity with traditional methods’ stable past-to-future assumption.Source: §2.2 · pp. 5–6

2.3 Definition of price forecasting

  • Price forecasting maps historical target-series observations and related exogenous sequences to a future asset price.Related inputs can include technical, sentiment, or macroeconomic variables.Source: §2.3 · p. 6

2.4 Workflow of price forecasting

  • The workflow checks and processes raw data, converts series into labeled sliding windows for supervised training, and selects a model to meet task requirements.Processing may include denoising or feature engineering, while model selection depends partly on data availability, computing resources, and domain knowledge.Source: §2.4 · p. 6
  • Complex models generally need more training data for optimal performance, and they do not consistently outperform simpler models.A small dataset can leave a complex model prone to high variance or overfitting, whereas a large dataset can leave a simple model underfit.Source: §2.4 · p. 7

3. OVERVIEW

The review searched studies published from 2020 to 2022 and included 72 articles on deep-learning financial price forecasting. It summarizes the asset prices studied and the characteristics and sources of their raw datasets.

3. OVERVIEW

  • 72 studies were included after screening 711 search results, removing 167 duplicates, and excluding articles that failed the review criteria.A further 59 articles were excluded because their full texts could not be retrieved.Source: Figure 3 · p. 8, §3 · p. 8
  • Fund forecasting had not received significant attention among the reviewed asset-price categories.Figure 4 summarizes the types of asset prices predicted in the reviewed studies.Source: §3 · pp. 8–9, Figure 4 · p. 9
  • Figure 5 summarizes the specific financial assets predicted across the reviewed studies.The overview identifies stock indices, individual stock prices, commodity prices, cryptocurrency rates, and forex rates as financial asset types.Source: §3 · p. 9, Figure 5 · p. 10
  • Most included studies used daily historical trading data, and datasets exceeding five years were typically considered necessary for sufficient model training.Yahoo Finance was the most widely used raw-data source because it provides an accessible database for queries and downloads.Source: §3 · p. 9, Figure 6 · p. 10

4. DEEP LEARNING MODELS FOR PRICE FORECASTING

This section reviews deep learning models for price forecasting, covering individual models and ensembles as well as recurrent, convolutional, hybrid, autoencoder, Transformer, graph, adversarial, and quantum-classical approaches. It describes their modeling mechanisms and applications alongside practical limitations.

4.1.1 Deep neural networks (DNNs)

  • DNNs outperformed LSTM and CNN-LSTM models for BTC prices using technical indicators, and a DNN on denoised inputs outperformed ARIMA.The reported comparisons concern specific BTC-price inputs and a denoised-input setting, respectively.Source: §4.1.1 · p. 11
  • Conventional DNNs flatten input sequences and do not explicitly model relationships among elements, which may limit their capture of temporal or spatial dependencies.DNNs can learn hierarchical representations, with lower layers capturing simple features and higher layers forming more complex abstractions.Source: §4.1.1 · p. 12

4.1.2 One-dimensional convolutional neural networks (1D CNNs)

  • One-dimensional convolutional filters capture local sequence patterns; Durairaj and Mohan (2022) proposed a price-forecasting model incorporating chaos theory, a 1D CNN for initial predictions, and polynomial regression to predict errors from the error series.The 1D CNN may struggle to model long-term dependencies and relationships between distant data points.Source: §4.1.2 · pp. 12–13

4.1.3 Recurrent neural networks (RNNs)

  • RNNs pass information from earlier steps to capture temporal dependencies; LSTMs use gated memory for long-term patterns, while GRUs have fewer parameters and ConvLSTMs capture spatiotemporal dependencies.Forecasting layers include many-to-one, bidirectional, and stacked forms.Source: §4.1.3 · pp. 13–14
  • LSTM models have limited memory capacity and may struggle to distinguish intricate patterns from random fluctuations; under these conditions, denoising via filtering or smoothing and feature engineering can enhance model performance.Source: §4.1.3 · p. 15
  • The section lists varied RNN-based forecasting models, including MRC-LSTM, DBGRUNN, DP-MAELS, ST-GRU, PSR-Stacked-BiLSTM, and RNN-LSTM.The supplied passages identify these model names but do not provide their comparative results.Source: §4.1.3 · p. 16

4.1.4 Convolutional-recurrent neural networks (CRNNs)

  • CRNNs combine convolutional layers for spatial features with recurrent layers for temporal dependencies, but their resource demands, training time, tuning difficulty, and limited interpretability are drawbacks.The recurrent layers may precede convolution when temporal dependencies must be modeled before spatial feature extraction.Source: §4.1.4 · pp. 16–17

4.1.5 Autoencoders

  • Autoencoders encode historical time series into compressed representations and decode them into future predictions, enabling feature learning but risking poor interpretability and overfitting.They require sufficient training data.Source: §4.1.5 · p. 18

4.1.6 Transformers

  • Transformers use multi-head self-attention to capture relationships between distant sequence positions, but their quadratic time and memory requirements reduce efficiency on extremely long sequences.A described price-forecasting architecture divides multivariate input into fixed-length patches, embeds them, and feeds them to a Transformer encoder.Source: §4.1.6 · pp. 19–20

4.1.7 Graph neural networks (GNNs)

  • GNN forecasting represents related assets as graph nodes, uses relationships such as correlation or sector similarity as edges, and applies graph convolution to update node features.Node inputs include historical trading attributes and technical indicators; a GNN block can combine a graph convolution network with a CNN or LSTM.Source: §4.1.7 · pp. 20–21
  • GNNs model interconnections among assets that may improve forecasting, but require carefully constructed graphs and can struggle with limited history, market shifts, and computational demands.Graph structure encodes relationships among nodes and remains fixed during graph convolution.Source: §4.1.7 · pp. 21–22

4.1.8 Generative adversarial networks (GANs)

  • GAN price forecasters use an LSTM-like generator and a neural-network discriminator (Li et al., 2021; Staffini, 2022), but training can be data-hungry and unstable.GANs can generate synthetic price sequences resembling real market data, while limited data can impede effective training.Source: §4.1.8 · p. 22

4.1.9 Deep quantum neural networks (DQNNs)

  • Paquet and Soleymani (2022) proposed QuantumLeap, which uses a QNN to predict an output density matrix from high/low prices and a classical DNN to measure the day's maximum price.The QNN is trained on input-output density-matrix pairs, while the classical DNN learns to measure the maximum price from the output matrix.Source: §4.1.9 · pp. 22–23
  • DQNNs remain early-stage: practical use requires specialised packages and hardware, and QNN training and optimisation are difficult.The review notes that classical optimisation algorithms cannot be directly applied and that DQNNs have no definitive advantage over classical models yet.Source: §4.1.9 · p. 23

4.2. Ensemble models

The review covers stacking and decomposition ensembles for price forecasting, describing their construction, benefits, and limitations. It also notes LSTM networks’ continuing significant position in time-series forecasting.

4.2.1 Stacking ensemble

  • Stacking combines diverse base-model predictions through meta-learning or arithmetic averaging, capturing broad patterns and complex relationships and yielding robustness and greater reliability than individual models.Source: §4.2.1 · p. 24
  • Stacking requires additional computation, depends on base-model selection and configuration, and can obscure which underlying factors contribute to forecasts.Source: §4.2.1 · p. 24

4.2.2 Decomposition ensemble

  • Decomposition ensembles split the original series into intrinsic mode functions, train an independent deep learning model on each, and sum their predictions.Source: §4.2.2 · p. 25
  • The reviewed decomposition techniques include EMD, EEMD, MEMD, and CEEMD.Source: §4.2.2 · p. 25
  • Decomposition can capture trends, seasonality, and residuals, improving predictions and interpretability through component separation.Source: §4.2.2 · p. 26
  • Most time-frequency decomposition ensembles decompose once, introducing future information into training; the review says this mixes in-sample and out-of-sample inputs and reports higher forecasting accuracy.Source: §4.2.2 · p. 26

4.3. Discussion

  • The review presents the distribution of deep learning models in Figure 19 and notes that LSTM networks continue to hold a significant position in time-series forecasting.Source: §4.3 · p. 26, Figure 19 · p. 27

5. FUTURE DIRECTIONS

The review identifies four priorities: evaluate complex deep-learning architectures, extend point forecasts to intervals, test decomposition-ensemble reliability, and study data-volume effects. It notes specific concerns including Transformers’ computational cost and possible future-data leakage in decomposition.

Complex architectures

  • Transformer models remain limited in price forecasting, and their self-attention requires quadratic time and memory as input length grows.The review also calls for further investigation of complex architectures and DQNN effectiveness using more advanced software and hardware.Source: §5 · p. 28

Interval forecasting

  • Deep-learning price forecasting should extend from point forecasts to prediction intervals with upper and lower bounds for future values.The review describes interval forecasting as a promising direction and notes that existing approaches can use deep-learning point forecasts to construct intervals.Source: §5 · p. 28

Decomposition ensembles

  • One-time decomposition of the entire price series can mix in-sample and out-of-sample data, potentially exposing future characteristics and inflating accuracy.The review therefore calls for testing decomposition ensembles’ reliability and their effectiveness in real-world forecasting.Source: §5 · p. 28

Data volume

  • Data-volume effects warrant study because reviewed papers often do not establish whether training datasets are excessive or inadequate.The review suggests suitably sized datasets may reduce computational costs while achieving optimal model performance.Source: §5 · pp. 28–29

6. CONCLUSION

This review synthesizes recent deep-learning price-forecasting research from theoretical and practical perspectives, covering architectures and applications. It reports adoption of newer architectures alongside LSTM and identifies limitations in the review’s coverage and several future research directions.

Conclusion

  • The review examines model architectures to illuminate input processing and decision-making, and summarizes practical applications to encourage further adoption.It describes these as theoretical and practical perspectives, respectively.Source: §6 · p. 29
  • LSTM remains prominent, while Transformers, GANs, GNNs, and DQNNs have also been adopted for price forecasting.The review interprets this adoption as evidence that the financial sector benefits from recent deep-learning advances.Source: §6 · p. 29

Limitations

  • Some relevant studies may have been missed because they were absent from selected databases or used less common search terms.The review notes that comprehensive keywords were used, but this possibility remains.Source: §6 · p. 29
  • The review does not provide an exhaustive experimental comparison between deep learning and other methods or among the included studies.The authors state such comparisons would require significant computational resources and remain future work.Source: §6 · p. 29

Future research

  • Future work should examine complex architectures, interval forecasting, decomposition-ensemble reliability, and the effects of data quantity on performance.The review links data-quantity exploration to the long-term application of forecasting models.Source: §6 · p. 29

Abbreviations

The abbreviations list expands model, method, and evaluation terminology used in the review.

Model abbreviations

  • The glossary includes neural architectures such as CNN, GRU, LSTM, GAN, GNN, and DQNN.It also lists recurrent and attention-based variants.Source: Abbreviations · pp. 30–31

Method abbreviations

Metric abbreviations

Funding Information

The study reports no specific grant funding from public, commercial, or non-profit agencies.

Funding

Loading 2305.04811v2…