Source-linked AI summary
Listening to Chaotic Whispers: A Deep Learning Framework for News-oriented Stock Trend Prediction
Ziniu Hu, Weiqing Liu, Jiang Bian, Xuanzhe Liu, Tie-Yan Liu
TL;DR
Stock trend prediction is challenged by volatile markets and chaotic online news whose quality and reliability vary widely. The paper proposes a Hybrid Attention Network with self-paced learning to model sequential news and diverse influence while learning progressively harder examples. On real-world data, HAN-SPL achieves the best reported accuracy and simulated investment returns, including a 0.611 annualized return versus 0.04 for the Market baseline.
Problem
Stock trend prediction is difficult because markets are volatile and non-stationary, while online financial content varies widely in quality, trustworthiness, and comprehensiveness.
Method
The paper combines Hybrid Attention Networks for sequential related-news modeling with self-paced learning for progressively more difficult training samples.
Results
0.611 annualized return versus 0.04 for the Market performance; HAN-SPL also achieves the best accuracy of 0.478.
Takeaways & Limitations
The framework improves stock trend prediction and produces appreciable profits with highly increased annualized excess return in one-year trading simulation.
Abstract
from arXiv · showhide
Stock trend prediction plays a critical role in seeking maximized profit from stock investment. However, precise trend prediction is very difficult since the highly volatile and non-stationary nature of stock market. Exploding information on Internet together with advancing development of natural language processing and text mining techniques have enable investors to unveil market trends and volatility from online content. Unfortunately, the quality, trustworthiness and comprehensiveness of online content related to stock market varies drastically, and a large portion consists of the low-quality news, comments, or even rumors. To address this challenge, we imitate the learning process of human beings facing such chaotic online news, driven by three principles: sequential content dependency, diverse influence, and effective and efficient learning. In this paper, to capture the first two principles, we designed a Hybrid Attention Networks to predict the stock trend based on the sequence of recent related news. Moreover, we apply the self-paced learning mechanism to imitate the third principle. Extensive experiments on real-world stock market data demonstrate the effectiveness of our approach.
1 INTRODUCTION
Stock trend prediction is difficult because markets are volatile and non-stationary, while online financial content varies widely in quality and reliability. The paper responds with a human-inspired framework that models sequential news context, diverse news influence, and progressively harder learning, achieving stronger prediction and simulated investment results.
- Volatile, non-stationary markets make precise stock trend prediction difficult for investors.
- Online news offers useful market information, but its quality, trustworthiness, and comprehensiveness vary drastically, with much content consisting of low-quality news, comments, or rumors.
- The framework imitates human learning through sequential context dependency, diverse influence, and effective and efficient learning.
- Hybrid Attention Networks process sequences of recent related news with attention mechanisms, while self-paced learning progressively increases training-sample difficulty.
- Extensive real-world experiments show improved stock trend prediction, and simulated trading achieves much better annualized returns than baseline methods.
- The paper contributes human-learning principles, a Hybrid Attention Network with self-paced learning, and experiments including simulated investment performance.
2 RELATED WORK
Prior stock prediction methods primarily use technical or fundamental information, increasingly including online news and sentiment. The paper identifies insufficient attention to news quality and proposes a human-inspired framework for chaotic news.
- Traditional stock prediction approaches mainly use technical market data or fundamental information from outside historical market data.
- Technical-analysis research has expanded from autoregressive and nonlinear models to deep neural networks and recurrent architectures for complex or long-term market patterns.
- Internet content, especially news, has become an important fundamental-analysis source as information availability has expanded.
- Researchers have mined news, social media, sentiments, events, knowledge graphs, volatility, and expressive textual features for stock-related prediction tasks.
- Few prior efforts sufficiently address the quality, trustworthiness, and comprehensiveness of news, despite their strong effect on textual-analysis effectiveness.
3 EMPIRICAL ANALYSIS
The empirical analysis motivates three design principles for prediction from chaotic news: combine sequential context, distinguish news by influence, and learn from easier informative cases before harder ones.
- 3.1 Sequential Context Dependency: A sequence of related news can provide complementary information that supports more reliable stock-trend assessment than relying on one vague or limited report.
- 3.1 Sequential Context Dependency: SINOPEC reform news illustrates that preceding news sequences distinguish otherwise ambiguous reform announcements and provide different signals about subsequent trends.
- 3.1 Sequential Context Dependency: An ideal framework should integrate and interpret news within a sequential temporal context rather than analyze each item separately.
- 3.2 Diverse Influence: Significant news can exert more intensive and durable market influence than trivial news, so prediction should assign them greater attention.
- 3.3 Effective and Efficient Learning: News availability varies substantially: for Jiai Technology, 8.4% of periods had no reported news for more than 10 days, making prediction difficult.
- 3.3 Effective and Efficient Learning: Because prediction difficulty varies, an ideal learner should focus first on informative common cases and then optimize for harder exceptional samples.
4 DEEP LEARNING FRAMEWORK FOR NEWS-ORIENTED TREND PREDICTION
The framework treats stock-trend prediction as three-class classification from recent news sequences, combining Hybrid Attention Networks with self-paced learning. HAN models news-level and temporal-level influence, while SPL gradually introduces difficult samples during training.
- Problem formulation: Stock trend prediction uses news from the previous N days to classify the next-day trend as DOWN, UP, or PRESERVE.Each date contributes a corpus of related news, and the model predicts the class of the next date’s rise percent.
- Hybrid Attention Networks: HAN combines news-level and temporal-level attention to distinguish influential news and critical time periods.The framework aggregates news within each date, encodes the resulting corpus sequence with bidirectional GRUs, and applies temporal attention before classification.
- Hybrid Attention Networks: News-level attention weights individual news items before producing a weighted corpus vector for each date.The attention layer is trained end-to-end to assign greater attention to reliable and informative news based on content.
- Hybrid Attention Networks: Bidirectional GRUs encode the temporal sequence of daily corpus vectors, incorporating surrounding context before temporal attention and classification.The GRU uses reset and update gates to control how past and new information contribute to each state.
- Self-paced Learning: Self-paced learning jointly optimizes model parameters and sample weights, beginning with easier samples and progressively incorporating harder ones.A pace parameter controls when samples with larger losses enter training; Alternative Convex Search alternates optimization over model parameters and latent weights.
- Self-paced Learning: Self-paced learning addresses scarce or vague news by skipping challenging training samples early and increasing training complexity over time.The method uses a self-paced regularizer and a pace-controlling hyper-parameter to organize sample selection during learning.
5 EVALUATION
The evaluation combines experiments on the proposed deep learning framework with a trading simulation on real-world market data. The section examines predictive performance and simulated investment effectiveness.
- The evaluation conducts comprehensive experiments to assess the proposed deep learning framework.
- A systematic trading simulation examines the framework’s effectiveness on the real-world market.
5.1 Experimental Setup
The experiments use Chinese stock and economic-news data, tri-label trend classification, temporal splitting, and comparisons against conventional, recurrent, and attention-based methods.
- The dataset covers 2,527 Chinese stocks with daily price and trading-volume time series from 2014 to 2017.
- 1,271,442 economic news items were collected, linked to stocks, filtered to 425,250 related items, and aggregated into daily news corpora.
- Stock trends are classified as DOWN below −0.41%, UP above 0.87%, or PRESERVE between those thresholds.
- The data are split chronologically into 66.7% training data from September 2014–May 2016 and 33.3% test data from May 2016–March 2017.
- Baselines include Random Forest, MLP, News-RNN, One-RNN, Temporal-Attention-RNN, and News-Attention-RNN, alongside HAN and HAN-SPL.
5.2 Effects of Two Attention Mechanisms
The attention analyses indicate that HAN assigns different importance to individual news items and dates, with recent news generally more influential but context-dependent variation remaining.
- News attention assigns low weights to price or code-only items, middle weights to prior market-performance reports, and high weights to analyst predictions or significant events.
- Overall temporal attention gives greater influence to recently reported news than to earlier news.
- Five temporal-attention clusters show that attention distributions vary substantially across contexts and news contents.
- Informative news from several days earlier can still influence the temporal-attention layer.
5.3 Effects of Self-paced learning
Self-paced learning initially trains more slowly because it can defer difficult samples, but later achieves higher testing accuracy than standard training.
- SPL starts more slowly than standard training, then eventually converges to a better testing-accuracy result.
5.4 Overall Performance Experiments
HAN-SPL achieves the strongest classification and simulated-trading performance across the reported settings, with the highest annualized return when selecting 40 stocks.
- HAN achieves the best accuracy among all baseline methods in the tri-label classification experiment.
- The trading simulation scores stocks by rising-trend probability minus declining-trend probability before portfolio construction.
- Annualized return measures cumulative profit per year, with portfolios evenly investing in the top 20, 40, 60, or 80 stocks.
- Selecting only 20 stocks produces less obvious improvement because higher turnover raises transaction costs and offsets profits.
- With 40 stocks, HAN-SPL's cumulative profit curve is best among the baseline methods.
5.5 Performance Discussion
HAN-SPL achieves the strongest reported accuracy and investment returns, while the ablations indicate that sequential modeling, attention, and self-paced learning each contribute to performance.
- Performance results: 0.611 annualized return is achieved by HAN-SPL when investing in the top 40 stocks, versus 0.04 for the Market baseline.HAN-SPL also achieves the best accuracy, 0.478, and the highest profits across all tested settings.
- RNN setting: MLP and RF perform worse than sequential models, indicating that organizing news into sequential contexts is important for prediction.The discussion attributes this difference to the lack of sequential organization in MLP and RF inputs.
- Bi-directional GRU setting: News-RNN outperforms One-RNN, indicating that the bi-directional GRU setting benefits prediction by using information from both past and future.The two models have similar structures except for their bi-directional versus one-directional GRU settings.
- Attention mechanisms: Temp-ATT and News-ATT outperform News-RNN, while HAN outperforms these models by combining two attention layers.The comparison supports distinguishing the diverse influences of different news items and dates.
- Performance results: HAN-SPL significantly outperforms all other methods throughout the test period, with profit gaps remaining nearly constant over time.The accuracy and annualized-return trends are consistent across methods.
- Self-paced learning: HAN-SPL outperforms HAN trained conventionally, supporting the importance of effective and efficient learning through self-paced training.The comparison isolates the contribution of self-paced learning within the HAN framework.
6 CONCLUSION AND FUTURE WORK
The paper proposes HAN with self-paced learning for news-oriented stock trend prediction and reports improved accuracy and profitable back-testing results. Future work will connect news across stocks and combine news-oriented modeling with technical analysis.
- Conclusion: The framework combines Hybrid Attention Network with self-paced learning to model sequential context, diverse influence, and effective learning from online news.These three principles are presented as the basis for the proposed framework.
- Conclusion: Experiments on real stock-market data show significant improvement in stock-trend prediction accuracy.The conclusion reports the result without specifying a numerical accuracy value.
- Conclusion: Back-testing produces appreciable profits and highly increased annualized excess return in a one-year trading simulation.The reported investment evaluation uses a one-year round trading simulation.
- Future work: Future work will model relationships among news for different stocks through industrial connections and integrate news-oriented modeling with technical analysis.Both directions are proposed to support more accurate stock-trend prediction.