[Paper Review] Listening to Chaotic Whispers: A Deep Learning Framework for News-oriented Stock Trend Prediction

Ziniu Hu, Weiqing Liu+3 more|arXiv (Cornell University)|Dec 6, 2017

Stock Market Forecasting Methods58 citations

TL;DR

Introduces Hybrid Attention Networks with self-paced learning to predict stock trends from chaotic online news, improving accuracy and simulated trading returns.

ABSTRACT

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.

Motivation & Objective

• Motivate stock trend prediction using chaotic online news and address data quality issues.
• Model sequential context and diverse influence of news using attention mechanisms.
• Incorporate an effective and efficient learning paradigm via self-paced learning to handle hard samples.

Proposed method

• Proposes Hybrid Attention Networks (HAN) with two attention layers: news-level attention within a date and temporal attention across dates.
• Uses a news embedding layer (pre-trained Word2Vec) to create news vectors, followed by a news-level attention to form a date corpus vector.
• Employs a bi-directional GRU to encode the sequence of date vectors and a temporal attention layer to aggregate across time.
• Implements self-paced learning (SPL) to progressively incorporate harder samples by optimizing model parameters and sample weights iteratively.
• Evaluates using real-world Chinese stock and news data, plus a simple trading simulation; compares HAN to RF, MLP, News-RNN variants, and HAN variants with SPL.
• Key equations include the news-level attention u_ti = sigmoid(W_n n_ti + b_n), alpha_ti = exp(u_ti) / sum_j exp(u_tj), d_t = sum_i alpha_ti n_ti, GRU state equations, and the temporal attention o_i and beta_i for V construction.

Experimental results

Research questions

• RQ1 Can a deep learning framework effectively predict next-day stock trend using chaotic, varying-quality news?
• RQ2 Do hierarchical attention mechanisms capturing sequential context and news-level importance improve prediction over baselines?
• RQ3 Does self-paced learning enhance performance when data include noisy or sparse news periods?

Key findings

• HAN with SPL achieves the best accuracy among baselines in tri-label classification.
• Temporal attention captures that recently published news generally have more impact, with diverse patterns across contexts.
• Self-paced learning improves final accuracy and convergence, especially on harder samples.
• Trading simulation shows HAN-SPL yields the highest annualized return across various top-K stock selections (e.g., 0.611 annualized return with top 40 stocks).
• News attention can distinguish informative articles from uninformative ones, aligning with observed high-attention examples.
• The framework demonstrates significant performance gains over RF, MLP, News-RNN, and Temp-ATT variants.