Deep learning models for price forecasting of financial time series: A review of recent advancements: 2020-2022
Deep learning models for price forecasting of financial time series: A review of recent advancements: 2020-2022
This paper provides a systematic review of deep learning models applied to financial time series price forecasting, focusing on advancements from 2020 to 2022. The authors categorize models into traditional deep learning architectures (e.g., CNNs, RNNs, LSTMs) and advanced models (Transformers, GANs, GNNs, DQNNs), discussing their architectures, applications, advantages, and limitations. Key findings indicate that deep learning models, particularly Transformers and attention-based models, have significantly improved forecasting accuracy by capturing complex temporal dependencies. GANs are highlighted for their ability to generate realistic synthetic data, addressing data scarcity issues, while GNNs effectively model inter-asset relationships. The review also identifies critical gaps and future research directions, such as the need for interval prediction over point forecasts, the reliability of decomposition ensembles, and the impact of data volume on model performance. This comprehensive survey serves as a valuable resource for researchers and practitioners aiming to leverage deep learning for financial forecasting.
Highlights
- 1Comprehensive review of deep learning models for financial time series price forecasting from 2020-2022.
- 2Detailed analysis of advanced models including Transformers, GANs, GNNs, and deep quantum neural networks.
- 3Identification of future research directions such as interval prediction and data volume impact.
- 4Comparison of deep learning models with traditional statistical and machine learning approaches.
Methods
- MLiterature review methodology covering 2020-2022 publications.
- MTaxonomy of deep learning architectures (CNNs, RNNs, Transformers, GANs, GNNs, DQNNs).
- MEvaluation of model performance metrics (e.g., RMSE, MAE, MAPE).
Results
- RDeep learning models outperform traditional methods in price forecasting accuracy.
- RTransformers and attention mechanisms show promise for capturing long-term dependencies.
- RGANs are effective for generating synthetic financial data and improving robustness.
- RGNNs excel in modeling relational dependencies among multiple assets.
- RDeep quantum neural networks are emerging but require further validation.
Analyze Paper
Generate insights from "Deep learning models for price forecasting of financial time...".