Deep Learning Methods for Time Series Forecasting: A Comprehensive Review

This Survey has consolidated deep learning methods for time series forecasting into five architectural families: recurrent, convolutional, Transformer-based, specialized/hybrid and foundation models. From vanilla RNNs to gated recurrent architectures (LSTM, GRU), attention-augmented (DA-RNN, TFT), convolutional alternatives (TCN), to the Transformer revolution triggered by Informer, Autoformer and FEDformer. PatchTST and iTransformer. In addition, specialized architectures such as TimesNet introduced 2-D tensors into the design space and pure-MLP architectures (N-BEATS, N-HiTS, TimeMixer) demonstrated that simple baselines continue to be very competitive. A key recent development that has had the biggest impact is the emergence of time series foundation models. TimeGPT, Chronos, Moirai, TimesFM, Sundial, Timer-XL, UniTS, ChatTime and Lag-Llama are instances of a paradigm shift from task-specific supervised models to general-purpose pre-trained models that can make zero-shot forecasts. Chronos's finding-that a vanilla T5 trained on tokenized series can match or exceed supervised suggests a consensus on the language model paradigm, where innovations specific to time-series are mostly related to tokenization, pre-training data, adaptation, and not the underlying architecture. Linguistic knowledge is transferred to numerical series with few adaptations, as shown by LLM-reprogramming approaches (Time-LLM, GPT4TS, TEMPO) and by Sundial's flow-matching TimeFlow Loss towards native continuous-valued pre-training. Incorporating open challenges, such as data scarcity, distribution shift, interpretability, calibration, multivariate heterogeneity, benchmarking inconsistencies, and edge deployment, necessitates concerted efforts in data curation, model architecture, model evaluation, and deployment infrastructure. Time series forecasting is now a commodity service available in nearly all scientific and industrial fields, as the foundation models evolve, just as text generation and image recognition did before it.

1.       W. Liao, S. Wang, D. Yang, Z. Yang, J. Fang, C. Rehtanz, and F. Porte-Agel, “TimeGPT in Load Forecasting: A Large Time Series Model Perspective,” arXiv preprint arXiv:2404.04885, 2024. 2.       Y. Liu, T. Hu, H. Zhang, H. Wu, S. Wang, L. Ma, and M. Long, “iTransformer: Inverted Transformers Are Effective for Time Series Forecasting,” in Proc. Int. Conf. Learn. Representations (ICLR), 2024. 3.       P. Hewage, A. Behera, M. Trovati, E. Pereira, M. Ghahremani, F. Palmieri, and Y. Liu, “Temporal Convolutional Neural (TCN) Network for an Effective Weather Forecasting Using Time-Series Data From the Local Weather Station,” Soft Comput., vol. 24, pp. 16453–16482, 2020. 4.       H. Wu, T. Hu, Y. Liu, H. Zhou, J. Wang, and M. Long, “TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis,” in Proc. Int. Conf. Learn. Representations (ICLR), 2023. 5.       M. R. Kabir, D. Bhadra, M. Ridoy, and M. Milanova, “LSTM–Transformer-Based Robust Hybrid Deep Learning Model for Financial Time Series Forecasting,” Sci, vol. 7, no. 1, Art. no. 7, 2025, doi: 10.3390/sci7010007. 6.       Y. Liang, H. Wen, Y. Nie, Y. Jiang, M. Jin, D. Song, S. Pan, and Q. Wen, “Foundation Models for Time Series Analysis: A Tutorial and Survey,” in Proc. 30th ACM SIGKDD Conf. Knowledge Discovery and Data Mining (KDD), pp. 6555–6565, 2024. 7.       G. Ke, Q. Meng, T. Finley, T. Wang, W. Chen, W. Ma, Q. Ye, and T.-Y. Liu, “LightGBM: A Highly Efficient Gradient Boosting Decision Tree,” in Advances in Neural Information Processing Systems, vol. 30, pp. 3146–3154, 2017. 8.       S. Makridakis, E. Spiliotis, and V. Assimakopoulos, “The M4 Competition: Results, Findings, Conclusion and Way Forward,” Int. J. Forecasting, vol. 34, no. 4, pp. 802–808, 2018. 9.       S. Hochreiter and J. Schmidhuber, “Long Short-Term Memory,” Neural Computation, vol. 9, no. 8, pp. 1735–1780, 1997. 10.    K. Cho, B. van Merriënboer, C. Gülçehre, D. Bahdanau, F. Bougares, H. Schwenk, and Y. Bengio, “Learning Phrase Representations Using RNN Encoder–Decoder for Statistical Machine Translation,” in Proc. Conf. Empirical Methods in Natural Language Processing (EMNLP), pp. 1724–1734, 2014.