Bridging Univariate Foundation Models and Multivariate Financial Signals for Stock Price Prediction

Document Type : Research Article

Authors

School of Intelligent Systems, University of Tehran, Tehran, Iran.

10.22108/jcs.2025.144738.1160

Abstract

This paper presents a hybrid forecasting framework that couples a pre-trained univariate foundation model for time-series analysis with a lightweight multivariate predictor to forecast next-day stock prices on the Tehran Stock Exchange. We design two complementary fusion strategies. In the residual-correction strategy, a conventional multivariate regressor first produces a price estimate; its day-ahead error sequence is then modelled by the foundation model, whose predicted residual is added back to refine the original forecast. In the relational strategy, the foundation model generates independent one-step forecasts for each financial variable, and these preliminary outputs are concatenated and fed into a multivariate neural network that delivers the final price prediction.
The framework is evaluated on ten highly liquid Iranian stocks across the 2000-2023 period. Empirical results show that the hybrid approach lowers Mean Squared Error by as much as 30 percent and raises the macro-averaged F1-score for price-direction classification by up to 25 percent when compared with strong foundation-model baselines. These improvements underscore the value of combining pre-trained temporal representations with auxiliary market signals for more accurate and robust stock forecasting.

Keywords

Main Subjects

References

[1] Han, Z. and Zhao, J. and Leung, H. and Ma, K. F. and Wang, W.. A Review of Deep Learning Models for Time Series Prediction. IEEE Sensors Journal. 1--8, 2019. [DOI ]
[2] Salinas, D. and Flunkert, V. and Gasthaus, J. and Januschowski, T.. DeepAR: Probabilistic Forecasting with Autoregressive Recurrent Networks. International Journal of Forecasting. 2019. [DOI ]
[3] Brown, T. B. and Mann, B. and Ryder, N. and Subbiah, M. and Kaplan, J. and Dhariwal, P. and Neelakantan, A. and Shyam, P. and Sastry, G. and Askell, A. and Agarwal, S. and Herbert-Voss, A. and Krueger, G. and Henighan, T. and Child, R. and others. Language Models are Few-Shot Learners. Advances in Neural Information Processing Systems. 2020. [DOI ]
[4] Zhou, H. and Zhang, S. and Peng, J. and Zhang, S. and Li, J. and Xiong, H. and Zhang, W.. Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting. AAAI Conference on Artificial Intelligence. 2021. [DOI ]
[5] Nie, Y. and Nguyen, N. H. and Sinthong, P. and Kalagnanam, J.. A Time Series is Worth 64 Words: Long-term Forecasting with Transformers. International Conference on Learning Representations (ICLR). 2023. [DOI ]
[6] Zeng, A. and Chen, M. and Zhang, L. and Xu, Q.. Are Transformers Effective for Time Series Forecasting. AAAI Conference on Artificial Intelligence. 2023. [DOI ]
[7] Louvron, H. and Lavril, T. and Izacard, G. and Martinet, X. and Lachaux, M.-A. and Lacroix, T. and Rozière, B. and Goyal, N. and Hambro, E. and Azhar, F. and Rodriguez, A. and Joulin, A. and Grave, E. and Lample, G.. Llama: Open and Efficient Foundation Language Models. arXiv preprint arXiv:2302.13971. 2023. [DOI ]
[8] Lee, J. and Youn, H. L. and Poon, J. and Han, S. C.. StockEmotions: Discover Investor Emotions for Financial Sentiment Analysis and Multivariate Time Series. AAAI Bridge Conference on AI for Financial Services. 2023. [DOI ]
[9] Liu, X.-Y. and Wang, G. and Yang, H. and Zha, D.. FinGPT: Democratizing Internet-scale Data for Financial Large Language Models. NeurIPS Instruction Workshop. 2023. [DOI ]
[10] Das, A. and Kong, W. and Sen, R. and Zhou, Y.. A Decoder-Only Foundation Model for Time-Series Forecasting. International Conference on Machine Learning (ICML). 2024. [DOI ]
[11] Li, X. and Shen, X. and Zeng, Y. and Xing, X. and Xu, J.. FinReport: Explainable Stock Earnings Forecasting via News Factor Analyzing Model. The Web Conference. 2024. [DOI ]
[12] Jia, F. and Wang, K. and Zheng, Y. and Cao, D. and Liu, Y.. GPT4MTS: Prompt-based Large Language Model for Multimodal Time-series Forecasting. AAAI Conference on Artificial Intelligence. 2024. [DOI ]
[13] Rasul, K. and Ashok, A. and Williams, A. R. and Ghonia, H. and Bhagwatkar, R. and Khorasani, A. and Darvishi Bayazi, M. J. and Adamopoulos, G. and Riachi, R. and Hassen, N. and BiloÅ¡, M. and Garg, S. and Schneider, A. and Chapados, N.. Lag-Llama: Towards Foundation Models for Probabilistic Time Series Forecasting. arXiv preprint. 2024. [DOI ]
[14] Li, T. and Liu, Z. and Shen, Y. and Wang, X. and Chen, H. and Huang, S.. Master: Market-guided Stock Transformer for Stock Price Forecasting. AAAI Conference on Artificial Intelligence. 2024. [DOI ]
[15] Fan, J. and Shen, Y.. StockMixer: A Simple yet Strong MLP-Based Architecture for Stock Price Forecasting. AAAI Conference on Artificial Intelligence. 2024. [DOI ]
[16] Cao, D. and Jia, F. and Arik, S. Ö. and Pfister, T. and Zheng, Y. and Ye, W. and Liu, Y.. TEMPO: Prompt-based Generative Pre-trained Transformer for Time Series Forecasting. International Conference on Learning Representations (ICLR). 2024. [DOI ]
[17] Lin, M. and Wang, S. and Ma, L. and Chu, Z. and Zhang, J. Y. and Shi, X. and Chen, P.-Y. and Liang, Y. and Li, Y.-F. and Pan, S. and Wen, Q.. Time-LLM: Time Series Forecasting by Reprogramming Large Language Models. International Conference on Learning Representations (ICLR). 2024. [DOI ]
[18] Ekambaram, V. and Jati, A. and Dayama, P. and Mukherjee, S. and Nguyen, N. H. and Gifford, W. M. and Reddy, C. and Kalagnanam, J.. Tiny Time Mixers (TTMs): Fast Pre-trained Models for Enhanced Zero/Few-Shot Forecasting of Multivariate Time Series. arXiv preprint arXiv:2401.03955. 2024. [DOI ]

Files

History

  • Receive Date: 19 March 2025
  • Revise Date: 01 June 2025
  • Accept Date: 30 June 2025

Share

How to cite

Statistics