Readme👻

News🎇This paper has been accepted by ESWA

Weijie Zhu, Liang Xie, Hanyu Fu, Jiangling Zhang,
GHOST: Sentiment-gated mamba and stock-wise tokenization attention for enhanced stock prediction,
Expert Systems with Applications,
Volume 301,
2026,
130474,
ISSN 0957-4174,
https://doi.org/10.1016/j.eswa.2025.130474.

https://doi.org/10.1016/j.eswa.2025.130474

Abstract

Stock trend prediction faces significant challenges due to market sentiment influences, long-term dependencies, and time-varying stock correlation. Specifically, (1) Current sentiment analysis methods exhibit insufficient market sentiment quantification and lack dynamic adaptive integration mechanismsfor market fluctuations. (2) Transformer-based models’ quadratic complexity limits long-term stock prediction while lacking finance-specific temporal inductive biases. (3)Traditional temporal tokenization paradigms forcibly merge multi-stock features, weakening stock correlation modeling while dramatically increasing computational costs. To address these constraints, we present GHOST (Gated Hybrid Organization with Sentiment-guided Temporal Mamba and Stock-wise Tokenization Attention). In particular, we leverage GDELT(Global Database of Events, Language and Tone) sentiment quantification through a Hierarchical Sentiment-Gated Layer for dynamic fusion of affective features with trading data. Additionally, Intra-Stock Mamba Selection Layer achieves time-series linear complexity for long-term forecasting by combining a dynamically parameterized shared state space model to provide specialized financial inductive biases. Moreover, our Stock-wise Tokenization Layer converts temporal tokens into stock tokens while preserving data integrity, enabling Inter-Stock Attention Layer to capture stock correlation via attention between stock tokens and reducing attention computation complexity. Experimental results on real-world stock datasets demonstratethe effectiveness of our model

Figure 1: Architecture of GHOST: Sentiment-Gated Mamba with Stock-wise Tokenization for Multi-Stock Prediction.

Figure 2: GDELT-based Sentiment Quantification

Figure 3: Backtesting Pipeline

Usage

i. First, configure the environment according to the requirement.txt file

Ensure that the following libraries have aligned versions:

• causal-conv1d==1.1.0
• mamba-ssm==1.1.1
• torch==2.1.1+cu118
• torchvision==0.16.1+cu118
• torchaudio==0.16.1+cu118

ii. Download stock data to .\dataset\stock_data

market sentiment data to .\dataset

iii. Finally, python run.py .

💡💡💡 Don't forget to modify the number of input stocks and the number of features.Considering that mamba-ssm depends on a Linux environment, we used Ubuntu 22.04 with miniconda as the basic environment. I have also placed the wheels of these two libraries in the corresponding links, hoping this will make the reproduction of this work more convenient and quick.

❄[Update opensource wheel]:Baidu

❄[Update opensource wheel]:OneDrive

Dataset

We provide stock data from two markets: CSI300 and NASDAQ100. After data preprocessing, 189 and 64 stocks remain respectively, along with corresponding market sentiment data: CHN_NEWS_sentiment.csv and USA_NEWS_sentiment.csv.

🔥[Update opensource data]: Baidu

🔥[Update opensource data]:OneDrive

The market sentiment information of China and the United States is time-step aligned with the stock data of CSI300 and NASDAQ100 respectively. You can choose any connection to download the data.

At the same time, you need to modify sentiment_path="your root" in .\data_provider\data_loader to adapt to the file path of sentiment data.Also, note that there are some differences in the feature engineering of NASDAQ100 and CSI300 stock data, so you need to modify the feature_columns in .\data_provider\data_loaderaccordingly.

🚀Finally, the complete code for market sentiment data will also be open-sourced in the future.<br>

Build by Weijie Zhu and Jiangling Zhang