Multimodal deep learning for short-term stock volatility prediction

Abstract: Stock market volatility forecasting is a task relevant to assessing market risk. We investigate the interaction between news and prices for the one-day-ahead volatility prediction using state-of-the-art deep learning approaches. The proposed models are trained either end-to-end or using sentence encoders transfered from other tasks. We evaluate a broad range of stock market sectors, namely Consumer Staples, Energy, Utilities, Heathcare, and Financials. Our experimental results show that adding news improves the volatility forecasting as compared to the mainstream models that rely only on price data. In particular, our model outperforms the widely-recognized GARCH(1,1) model for all sectors in terms of coefficient of determination $R^2$, $MSE$ and $MAE$, achieving the best performance when training from both news and price data.

• The paper demonstrates that integrating textual news with price data using multimodal deep learning significantly improves daily volatility predictions compared to traditional models.
• The approach employs a hierarchical neural architecture with advanced attention mechanisms to effectively capture time-aligned financial news relevance.
• Experiments reveal robust performance across sectors by leveraging end-to-end BiLSTM encoders and transfer learning, achieving notable R² improvements over GARCH(1,1).

Multimodal Deep Learning for Short-Term Stock Volatility Prediction: A Technical Assessment

Problem Context and Motivation

This work addresses the challenge of accurate short-term (one-day-ahead) stock volatility prediction—a fundamental problem for risk assessment, regulatory compliance, and dynamic portfolio management. The authors contend that traditional models, particularly those relying solely on price information, systematically discard informative patterns imbued within textual news, especially corporate headlines. Prior literature predominantly focuses on long-horizon volatility, using unsophisticated text representations (e.g., bag-of-words, sentiment lexica) and late fusion paradigms, circumventing the need to capture fine-grained, word-order-dependent semantics required for high-frequency risk forecasting. This manuscript fills critical gaps by leveraging end-to-end, multimodal deep learning that fuses time-aligned, factual financial news with historical price data, thus enabling joint learning of volatility-relevant representations.

Methodological Innovations

Data Construction

A substantial corpus was curated, comprising approximately 147,000 financial news headlines from Reuters, dated 2007–2017, covering 50 large-cap US stocks across diverse sectors (Consumer Staples, Energy, Utilities, Healthcare, Financials). News items were algorithmically associated to stocks using surface form expansions (via DBpedia) and then time-binned to synchronize with market hours, aligned to Eastern Daylight Time for granular correspondence with trading activity.

Volatility Proxy

Rather than target raw squared returns (high variance, low efficiency), the methodology adopts the Garman-Klass and Parkinson range estimators as the volatility targets. These proxies—computed from open, high, low, and close prices—are theoretically efficient and practical alternatives when high-frequency intraday data is unavailable, following the recommendations in the financial econometrics literature.

Neural Architecture

A hierarchical, multimodal network is developed with explicit modules:

• Sentence Encoder: Both end-to-end and transfer learning (TL)-based sentence encoders are explored. End-to-end models employ BiLSTM with either attention or max-pooling aggregation. TL encoders are trained on auxiliary NLP tasks—Reuters RCV1 text categorization (domain-specific) and SNLI natural language inference (general, high semantic complexity). Word-level attention is included as a baseline.
• Hierarchical Attention Mechanism: News relevance is distilled via a daily news relevance attention (NRA) layer, which adaptively weights multiple news released per day. This module is empirically shown to outperform naive averaging.
• Temporal Context Encoding: Sequences of daily news embeddings undergo BiLSTM-based temporal encoding to capture novelty and persistence.
• Price Encoder: Two stacked LSTMs process price feature sequences.
• Stock Embedding: A categorical indicator passed through a trainable embedding layer permits stock-specific idiosyncrasy modeling within a global (cross-stock) prediction model.

All modalities are merged at the representation level and forecast the one-day-ahead Garman-Klass volatility via a shallow fully connected head.

Handling Missing Modalities

For days without news, a zero-imputation mechanism with a missingness indicator (cf. [Baltrusaitis et al., 2017], [Lipton et al., 2016]) is utilized, preventing loss of potentially informative “news absence” and supporting robust multimodal aggregation.

Experimental Protocol and Results

Baselines and Metrics

Comparisons are made against the classical GARCH(1,1) model (using conditional one-day-ahead volatility), unimodal price-only models, and ablated variants (e.g., without NRA or without news). Metrics include mean squared error (MSE), mean absolute error (MAE), and the coefficient of determination ($R^2$), with rigorous separation of training, validation, and out-of-sample test periods (2007–2017; test: 2016–2017).

Key Findings

• Numerical Performance: The best end-to-end model with BiLSTM+attention and NRA delivers $R^2 = 0.455$ (Garman-Klass), a substantial improvement over both GARCH(1,1) ($R^2 = 0.357$) and the price-only deep model ($R^2 = 0.384$). Relative MSE and MAE errors decrease by 11% and 9% respectively compared to price-only.
• Effect of News: Addition of news significantly increases forecasting performance and is complementary to price data rather than redundant.
• Sentence Encoder Comparison: End-to-end SOTA encoders (BiLSTM+Att/MP) outperform TL-based sentence encoders. Among TL encoders, those trained on the SNLI NLI dataset transfer better than those trained on the domain-specific Reuters RCV1 classification, consistent with [Conneau et al., 2017].
• Attention Mechanisms: The NRA module achieves measurable gain over per-day naive averaging, even when controlling for sentence encoder strength.
• Ablations: Simpler news encoders (e.g., word-level attention) benefit substantially from TL initialization but still do not exceed end-to-end architectures.
• Sector Generalization: Outperformance over GARCH(1,1) is robust across sectors, with improvements ranging from 0.225 to 0.538 in $R^2$, strongest in Energy and weakest in Utilities and HealthCare. The model does not suffer from sector overfitting due to the global cross-stock training approach.

Theoretical and Practical Implications

The empirical results empirically refute claims that news is redundant for short-horizon volatility prediction at the daily granularity. This contradicts much of the traditional literature that views price processes as efficiently incorporating public news at short time scales. The findings strengthen the argument for multimodal, end-to-end representation learning, demonstrating that the structure and timing of textual signals—especially when processed with sophisticated attention—can be harnessed for actionable risk estimation.

From a practical standpoint, these results suggest that asset managers and financial institutions should integrate dynamic, real-time NLP-driven news analysis into existing risk frameworks, moving beyond univariate, econometric models. The architecture also generalizes well to other high-frequency forecasting tasks with missing modalities and heterogeneous side information.

Future Directions

• Intraday Volatility Proxies: Although daily range estimators are efficient, leveraging high-frequency intraday data may further enhance both ground-truth volatility measurement and the capacity to temporally align news and price impact more precisely.
• Cross-market Transferability: The global model and hierarchical attention mechanism warrant further validation on other asset classes and international markets.
• Architecture Extension: Incorporating more advanced transformer-based sequence encoders and cross-modal fusion layers could capture higher-order interactions and emerge as valuable extensions.

Conclusion

This study establishes that integrating end-to-end deep multimodal networks, with explicit attention mechanisms and transfer learning from strong NLP tasks, materially surpasses both classical econometric and unimodal deep benchmarks for daily volatility prediction. Multimodal architectures with advanced textual encoders should be considered foundational for contemporary financial time series modeling, with implications for both practical forecasting systems and theoretical models of market microstructure and information diffusion.

---

Reference

"Multimodal deep learning for short-term stock volatility prediction" (1812.10479)