Effective Exploitation of Macroeconomic Indicators for Stock Direction Classification Using the Multimodal Fusion Transformer

Tae-Won Lee, Pawel Teisseyre, Jaesung LeeIEEE Access 2023 · DOI

Introduction

Predicting stock market movements remains one of the most challenging tasks in financial forecasting. The inherent noise, non-linearity, and "random walk" characteristics of price data often render traditional unimodal models—those relying solely on historical price data—insufficient. While recent advancements have integrated external data sources, the methodology of information fusion—how disparate data types are combined—is frequently overlooked.

This paper introduces a Multimodal Fusion Transformer designed to address these challenges. By integrating historical price data with macroeconomic indicators and temporal features through an early fusion strategy, the model captures complex, non-linear correlations that late-fusion models often miss.

Methodology: The Multimodal Fusion Transformer

The core innovation of this research lies in its approach to "intermodality correlations." Rather than processing different data types in isolation, the proposed architecture merges them early in the pipeline to leverage the synergistic relationships between market prices and the broader economic environment.

Data Modalities

The model processes three distinct input streams:

  1. Stock Price Modality: Comprising 17 features, including standard Open-High-Low-Close-Volume (OHLCV) data and technical indicators such as the Relative Strength Index (RSI), Moving Average Convergence Divergence (MACD), and Bollinger Bands.
  2. Temporal Modality: 17 features utilizing one-hot encoding for the month of the year and the day of the week to account for seasonal trends and calendar effects.
  3. Macroeconomic Modality: 6 key global indicators, including the NASDAQ 100 index, US 2/10/30-year bond yields, the US Dollar Index, and WTI Oil prices.

Early Fusion Architecture

Unlike "late fusion" models that merge modalities at the decision layer, this model employs an Early Fusion strategy. All features are normalized and concatenated into a single input matrix before entering the network.

Architecture of the Multimodal Fusion Transformer

As shown in the architecture diagram, the concatenated features pass through a Transformer encoder. The self-attention mechanism within the encoder allows every feature, regardless of its original modality, to "attend" to every other feature. This enables the model to learn, for instance, how a specific shift in bond yields might influence a stock's RSI in real-time. The output is then processed through feature and sequence compression layers before a final classification layer determines the stock's direction as "Up" or "Down."

Experimental Setup and Datasets

The model's performance was evaluated using the KDD17 Public Dataset, which contains historical price data for 50 major US stocks. This was augmented with macroeconomic data sourced from Investing.com. The combined dataset spans from January 1, 2007, to January 1, 2016, covering 2,518 trading days. To ensure the integrity of the results and prevent "data leakage," the authors employed blocked time-series cross-validation, ensuring the model never "sees" future data during the training phase.

Key Insights and Results

The primary finding of the study is that the relationship between modalities is often more predictive than the individual modalities themselves. Early fusion outperformed late fusion in 30 out of 50 datasets, demonstrating its superior ability to model cross-modal dependencies early in the learning process.

Comparative Performance

The Multimodal Fusion Transformer was benchmarked against three state-of-the-art models: CNN-Attention BiLSTM, Adversarial LSTM, and DTML.

  • Average Rank: The proposed model achieved the best average rank of 1.72 across all 50 stocks, significantly ahead of the next best model (DTML) at 2.68.
  • Statistical Significance: Performance gains were validated using Friedman and Bonferroni-Dunn tests, confirming that the improvements were statistically significant.
  • Sector Sensitivity: The model showed particular strength in the Energy and Financial sectors. This is likely because stocks in these sectors are highly sensitive to the specific macroeconomic indicators used, such as oil prices and interest rates.

Conclusion and Future Directions

The Multimodal Fusion Transformer represents a significant step forward in stock direction classification by effectively exploiting intermodality correlations. By treating concatenated features as a sequence for a Transformer encoder, the model acts as a sophisticated feature-selection and correlation-mapping tool.

Future research could expand this framework by incorporating a wider array of global indices or commodities. Additionally, integrating unstructured "soft" data, such as news sentiment or social media feeds, could further enhance the model's ability to navigate short-term market volatility. Finally, hyperparameter tuning tailored to specific market sectors may unlock even higher levels of predictive accuracy.

Summary updated

BibTeX
@article{lee2023effective,
  author = {Tae-Won Lee and Pawel Teisseyre and Jaesung Lee},
  title = {{Effective Exploitation of Macroeconomic Indicators for Stock Direction Classification Using the Multimodal Fusion Transformer}},
  journal = {IEEE Access},
  year = {2023},
  doi = {10.1109/access.2023.3240422},
}
Figure