Paper: arXiv 2410.03707

Authors: Ali Mehrabian, Ehsan Hoseinzade, Mahdi Mazloum, Xiaohong Chen

Abstract

Stock markets play an important role in the global economy, where accurate stock price predictions can lead to significant financial returns. While existing transformer-based models have outperformed long short-term memory networks and convolutional neural networks in financial time series prediction, their high computational complexity and memory requirements limit their practicality for real-time trading and long-sequence data processing. To address these challenges, we propose SAMBA, an innovative framework for stock return prediction that builds on the Mamba architecture and integrates graph neural networks. SAMBA achieves near-linear computational complexity by utilizing a bidirectional Mamba block to capture long-term dependencies in historical price data and employing adaptive graph convolution to model dependencies between daily stock features. Our experimental results demonstrate that SAMBA significantly outperforms state-of-the-art baseline models in prediction accuracy, maintaining low computational complexity. The code and datasets are available at github.com/Ali-Meh619/SAMBA.

Complexity vs Empirical Score

  • Math Complexity: 7.0/10
  • Empirical Rigor: 8.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper uses advanced mathematics including state space models, discretization, and bidirectional Mamba blocks, while also providing specific empirical results with comparisons to state-of-the-art baselines and publicly available code and datasets for reproducibility.

Research Flowchart

  flowchart TD
  A["Research Goal:<br>Accurate Stock Price Prediction<br>Low Computational Complexity"] --> B["Input Data:<br>Historical Stock Price Data"]
  B --> C["Methodology 1:<br>Bidirectional Mamba Block<br>for Long-Term Dependencies"]
  B --> D["Methodology 2:<br>Adaptive Graph Convolution<br>for Feature Dependencies"]
  C & D --> E["Computational Process:<br>Integrated Graph-Mamba Framework"]
  E --> F["Key Finding 1:<br>SOTA Prediction Accuracy"]
  E --> G["Key Finding 2:<br>Near-Linear Computational Complexity"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes SAMBA, a stock return prediction framework that combines a bidirectional Mamba block with adaptive graph convolution to capture temporal and feature dependencies.

Source sentence
we propose SAMBA, an innovative framework for stock return prediction that builds on the Mamba architecture and integrates graph neural networks (GNNs). SAMBA achieves near-linear computational complexity by utilizing a bidirectional Mamba block to capture long‑term dependencies in historical price data and employing a

Assumes 3

  • The data split assumes 80% training, 5% validation, and 15% testing. data
    Source sentence
    We use the first 80% of the data for training, the next 5% for validation, and the rest 15% for testing.
  • The model assumes that daily stock features can be represented as nodes in a graph whose edges are learned end‑to‑end. model
    Source sentence
    We propose a task‑specific adaptive graph construction method that allows the model to learn dependencies between daily stock features as a graph structure in an end‑to‑end manner.
  • The adaptive graph uses a Gaussian kernel with a learnable scaling factor to capture non‑linear relationships. model
    Source sentence
    We use the Gaussian kernel to build the dependency graph between daily stock features as follows:

Data

  • NASDAQ market-prices unknown
    January 2010 to November 2023 · 82 daily stock features
    Source sentence
    We apply our proposed model to three real‑world datasets from the US stock market with N = 82 daily stock features explained in [5]: NASDAQ, New York Stock Exchange (NYSE), and Dow Jones Industrial Average (DJIA), covering the period from January 2010 to November 2023.
  • NYSE market-prices unknown
    January 2010 to November 2023 · 82 daily stock features
    Source sentence
    We apply our proposed model to three real‑world datasets from the US stock market with N = 82 daily stock features explained in [5]: NASDAQ, New York Stock Exchange (NYSE), and Dow Jones Industrial Average (DJIA), covering the period from January 2010 to November 2023.
  • DJIA market-prices unknown
    January 2010 to November 2023 · 82 daily stock features
    Source sentence
    We apply our proposed model to three real‑world datasets from the US stock market with N = 82 daily stock features explained in [5]: NASDAQ, New York Stock Exchange (NYSE), and Dow Jones Industrial Average (DJIA), covering the period from January 2010 to November 2023.

Costs

Not applicable The paper reports no trading or execution result.

Finds 3

  • SAMBA outperforms all baselines in RMSE, IC, and RIC, achieving 11.72% RMSE improvement for NASDAQ, 10.07% for NYSE, and 6.90% for DJIA over the second‑best models. performance
    Source sentence
    Our proposed SAMBA model outperforms all the baselines in RMSE, IC, and RIC metrics by a significant margin. In particular, SAMBA achieves improvements in RMSE, with gains of 11.72% for NASDAQ, 10.07% for NYSE, and 6.90% for DJIA compared to the second‑best models, respectively.
  • SAMBA also improves IC and RIC substantially, e.g., 85.38% and 80.30% better IC/RIC for NASDAQ versus the second‑best models. predictive-accuracy
    Source sentence
    The proposed SAMBA model also shows improvements in IC and RIC, providing 85.38% and 80.30% better performance for NASDAQ, 36.43% and 50.32% for NYSE, and 33.62% and 53.54% for DJIA compared to the second‑best models, respectively.
  • SAMBA reduces training time per epoch by 6.70% compared to AGCRN and requires only 0.11 Million MACs, far less than Transformer and AGCRN. performance
    Source sentence
    In particular, the proposed SAMBA model improves the training time per epoch 6.70% compared to the AGCRN model, and requires 0.11 Million MACs, which is significantly less than that of the Transformer and AGCRN models.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 46 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 1 paper in the archive (1 by independent authors)

Code

https://github.com/Ali-Meh619/SAMBA — the authors' own repository

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (39)

  1. A Spatial-Temporal Transformer Network for City-Level Cellular Traffic Analysis and Prediction (2023)
  2. Mamba: Linear-time sequence modeling with selective state spaces (2023) — link
  3. CNN pred: CNN -based stock market prediction using a diverse set of variables (2019)
  4. U- CNN pred: A universal CNN -based predictor for stock markets (2019) — link
  5. Deep learning networks for stock market analysis and prediction: Methodology, data representations, and case studies (2017)
  6. Cellular Network Traffic Prediction Based on Correlation C onv LSTM and Self-Attention Network (2023)
  7. ST - T ran: S patial-Temporal Transformer for Cellular Traffic Prediction (2021)
  8. M amba: L inear-Time Sequence Modeling with Selective State Spaces (2023) — link
and 31 more
  1. Financial time series forecasting with multi-modality graph neural network (2022)
  2. Attention is All you Need (2017)
  3. Financial time series forecasting with multi-modality graph neural network (2022)
  4. A Dynamic B ernstein Graph Recurrent Network for Wireless Cellular Traffic Prediction (2023)
  5. Adaptive Graph Convolutional Recurrent Network for Traffic Forecasting (2020)
  6. Frequency-domain MLP s are More Effective Learners in Time Series Forecasting (2023)
  7. Fourier GNN : R ethinking Multivariate Time Series Forecasting from a Pure Graph Perspective (2023)
  8. A multi-source dataset of urban life in the city of M ilan and the P rovince of T rentino (2015)
  9. Bayesian Meta-Learning for Adaptive Traffic Prediction in Wireless Networks (2024)
  10. Combining recurrent, convolutional, and continuous-time models with linear state space layers (2021)
  11. Hi PPO : R ecurrent memory with optimal polynomial projections (2021)
  12. Is M amba Effective for Time Series Forecasting? (2024) — link
  13. Semi-Supervised Classification with Graph Convolutional Networks (2017)
  14. Joint spectrum, precoding, and phase shifts design for RIS -aided multiuser MIMO THz systems (2024)
  15. Adaptive Bandwidth Allocation in multiuser MIMO TH z systems with graph-transformer networks (2024)
  16. Are Transformers Effective for Time Series Forecasting? (2023)
  17. S tock M ixer: A Simple Yet Strong MLP -Based Architecture for Stock Price Forecasting (2024)
  18. Frequency-domain MLP s are More Effective Learners in Time Series Forecasting (2023)
  19. F ourier GNN : R ethinking Multivariate Time Series Forecasting from a Pure Graph Perspective (2023)
  20. TSM ixer: A n All- MLP Architecture for Time Series Forecasting (2023)
  21. Stock market prediction using LSTM recurrent neural network (2020)
  22. Accurate multivariate stock movement prediction via data-axis transformer with multi-level contexts (2021)
  23. Stock market index prediction using deep Transformer model (2022)
  24. Py T orch: An imperative style, high-performance deep learning library (2019)
  25. Adam: A Method for Stochastic Optimization (2015)
  26. Master: M arket-guided stock transformer for stock price forecasting (2024)
  27. CI-STHPAN : P re-trained Attention Network for Stock Selection with Channel-Independent Spatio-Temporal Hypergraph (2024)
  28. Empirical asset pricing via machine learning (2020)
  29. Semiparametric ARX neural-network models with an application to forecasting inflation (2001)
  30. Mamba S tock: S elective state space model for stock prediction — → 2402.18959
  31. Economic prediction using neural networks: T he case of IBM daily stock returns (1988)

1 of 39 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.