Source-linked AI summary
Leveraging Financial News for Stock Trend Prediction with Attention-Based Recurrent Neural Network
Huicheng Liu
TL;DR
Stock prediction from news is difficult because relevant events and their effects on prices can be indirect. The paper encodes financial news titles with a bidirectional LSTM and applies self-attention; it reports competitive results on index and individual-stock prediction, with individual-stock accuracy higher than S&P 500 index accuracy.
Problem
Identifying relevant news and its implicit, sometimes indirect effects on stock prices is difficult, while existing sentiment signals do not capture events and news context.
Source: §1 · p. 2Method
At-LSTM encodes financial news titles with a bidirectional LSTM and applies self-attention to relevant words, news, and days.
Source: §1 · p. 2, §3.2.1 · p. 9, §3.2.2 · pp. 10, 13Results
At-LSTM is reported as competitive with state-of-the-art approaches; its S&P 500 accuracy is 65.53% max and 63.06% average, lower than KGEB-CNN, while each tested individual stock exceeds 66% accuracy.
Source: §4.3 · pp. 18–19, §5 · p. 19Takeaways & Limitations
The authors conclude that At-LSTM is a promising use of deep-learning representations of financial news for predicting index and individual-stock directional movements.
Source: §5 · p. 19The study leaves prediction across multiple time horizons and how quickly information is absorbed for future work.
Source: §5 · p. 20
Abstract
from arXiv · showhide
Stock market prediction is one of the most attractive research topic since the successful prediction on the market's future movement leads to significant profit. Traditional short term stock market predictions are usually based on the analysis of historical market data, such as stock prices, moving averages or daily returns. However, financial news also contains useful information on public companies and the market. Existing methods in finance literature exploit sentiment signal features, which are limited by not considering factors such as events and the news context. We address this issue by leveraging deep neural models to extract rich semantic features from news text. In particular, a Bidirectional-LSTM are used to encode the news text and capture the context information, self attention mechanism are applied to distribute attention on most relative words, news and days. In terms of predicting directional changes in both Standard & Poor's 500 index and individual companies stock price, we show that this technique is competitive with other state of the art approaches, demonstrating the effectiveness of recent NLP technology advances for computational finance.
1. Introduction
Stock prediction matters because accurate forecasts can yield significant profit, but identifying news relevant to stock movements remains difficult. The paper proposes an attention-based LSTM that uses financial news to predict directional changes in the S&P 500 and individual stocks.
1. Introduction
- Stock prediction is appealing to researchers and practitioners because successful forecasts can yield significant profit.The introduction also describes stock prediction as challenging because financial time series are noisy and volatile.Source: §1 · pp. 1–2
- Daily events can affect prices through implicit correlations and indirect chain reactions, making relevant information difficult to identify.The introduction distinguishes technical analysis, fundamental analysis using external information such as news, and approaches combining financial time series with text.Source: §1 · p. 2
- At-LSTM uses recurrent encoding and self-attention over relevant words, news, and days to predict directional changes in the S&P 500 and individual stock prices.The model uses financial news titles and is described as competitive with state-of-the-art approaches.Source: §1 · pp. 2–3
2. Related work
Prior stock-prediction work has used market data and text, including sentiment and event representations. Neural approaches learn dense text representations, while the paper reviews limitations of earlier text-based methods and challenges in representing information with sparse indicators.
2. Related work
- Many stock-prediction studies analyze recent market prices and trading volumes.The cited examples are.Source: §2 · p. 3
- Earlier text-based stock prediction used news mining, including bursty keywords; these methods had limitations in uncovering market dynamics, leaving models unable to capture the impact of recent trends.Sentiment-based prediction is attributed to Skuza et al. and Sehgal et al..Source: §2 · p. 3
- Neural networks can learn dense text representations for sentiment, events, and factual information that are challenging to encode with sparse indicator features.The discussion frames available stock prices and relevant text as a setting for exploring deep-learning-based prediction.Source: §2 · pp. 3–4
- Related neural work includes event-structure representations, knowledge-graph event embeddings, and news-abstract representations for predicting cumulative abnormal returns.The review also notes other neural-network models proposed to improve prediction accuracy.Source: §2 · p. 4
3. Methodology
The methodology section introduces background and the design of At-LSTM, a model for predicting directional movements in the S&P 500 and individual stock prices using financial news titles.
3. Methodology
- The proposed At-LSTM predicts directional movements in the S&P 500 and individual stock prices using financial news titles.Source: §3 · p. 4
- The section presents background before introducing the proposed prediction model.Source: §3 · p. 4
- The model is named Attention-based LSTM (At-LSTM) and is shown in Figure 1.Source: §3 · p. 4
3.1. Background
The background reviews text and neural-network representations, then describes LSTM memory mechanisms and the optimization and loss choices used by the model.
3.1.1. Bag of Words Model and Word Embedding
- Bag-of-words represents text by word occurrence while disregarding grammar and word order; word embeddings map words or phrases to lower-dimensional real-valued vectors.The passage notes that embeddings have been used as input representations in NLP tasks including syntactic parsing and sentiment analysis.Source: §3.1.1 · pp. 4–5
3.1.2. Convolutional Neural Network
- CNNs use learnable filters and layers of convolution, subsampling, activation, and dropout; sparse interaction and parameter sharing are two key filter ideas.The illustrated architecture alternates convolutional and subsampling layers before fully connected classifier layers, typically ending with softmax.Source: §3.1.2 · p. 5, Figure 1 · p. 6
3.1.3. Long-Short Term Memory
- LSTM cells are designed to address long-term dependencies by retaining values over time, with gates controlling what is reserved, updated, or filtered from the cell state.The LSTM figure labels the input, output, forget-gate activation vectors, cell-state vector, and unit input and output.Source: Figure 2 · p. 6, §3.1.3 · p. 7
3.1.4. Optimization Algorithm and Loss Function
- The model uses Adadelta to update weights and binary cross-entropy to compare predictions with true outputs.The text describes Adadelta as requiring no manual learning-rate tuning and identifies cross-entropy as the model's loss function.Source: §3.1.4 · p. 8
3.2. Model design
At-LSTM has four stages: input and embedding, news-level encoding and attention, day-level encoding and attention, and output prediction. Its design turns financial news titles into representations and aggregates information across news and dates.
3.2.1. Input and Embedding Layer
- Whole news titles are embedded using word vectors concatenated with CNN-derived character-composition vectors.The model uses the full title content rather than event tuples; its word embeddings are trained on financial news text.Source: §3.2.1 · pp. 9–10
3.2.2. News Level Bi-LSTM and Self-Attention Layer
- A Bi-LSTM encodes each news title in both directions, concatenating forward and backward hidden states to represent the title in context.The resulting title representation is the input to the news-level attention layer.Source: §3.2.2 · p. 10
- Multi-hop self-attention assigns weights to Bi-LSTM-encoded news, with different groups attending to sentiment-related news and mentions of major S&P 500 companies.The attention layer is trained end-to-end to assign attention based on news content.Source: §3.2.2 · p. 13
3.2.3. Day level Bi-LSTM and self-attention layer
- A day-level Bi-LSTM encodes the sequence of daily news vectors, and self-attention weights dates within a delay window before prediction.The final vector represents news published before the prediction date.Source: §3.2.3 · p. 14
3.2.4. Output and Prediction Layer
- The final fully connected layer with softmax predicts whether the daily stock-price direction is up or down.The labels are [1,0] for an increase and [0,1] for a decrease.Source: §3.2.4 · p. 14
4. Experiments
The experiments evaluate At-LSTM on S&P 500 index and individual-company price movements using financial news, then compare it with prior models and targeted variants. Results report competitive index accuracy and higher accuracy for the listed individual stocks.
4.1.1. Data
- The index task predicts next-day S&P 500 closing-price direction from daily news headlines; the collected data span October 2006 to March 2018.The study used Reuters and Bloomberg news from October 2006 to November 2013, then Reuters news for 473 S&P 500 companies through March 2018, with historical stock prices from Yahoo Finance.Source: §4.1.1 · p. 15
4.1.2. Implementation Details
- The implementation used fine-tuned 100-dimensional skip-gram word embeddings, 15-dimensional character embeddings, and news- and day-level Bi-LSTMs with 300 hidden units.The day-level window was 7, and the model was trained for 200 epochs.Source: §4.1.2 · pp. 15–16
4.2. Base Lines and Proposed Model
- The comparisons included event-embedding CNNs, SVM, At-LSTM variants, document attention, technical indicators, and a CNN replacement for news-level attention.The variants tested sentence encoding, character composition, headline versus abstract or document input, and alternative news-level modeling.Source: §4.2 · pp. 16–18
4.3. Result and Discussion
- Deep neural models beat SVM; sentence encoding and character composition helped, technical indicators reduced accuracy, and news-level attention outperformed CNN substitution.Using abstracts or full documents also yielded lower accuracy than using headlines.Source: §4.3 · p. 18
- 65.53% max accuracy and 63.06% average accuracy were reported for At-LSTM, below KGEB-CNN.The table caption notes that maximum accuracies for [17, 18, 19] were unknown, so their reported accuracies were treated as averages.Source: §4.3 · p. 18, §4.2 · p. 18, Table 3 · p. 19
- Each listed individual stock exceeded 66% accuracy; WALMART reached 70.36% average and 72.06% maximum accuracy.The authors attribute higher individual-stock accuracy than index accuracy to more relevant company-specific news, versus noise from the full corpus used for the index.Source: §4.3 · p. 19
5. Conclusion
The paper concludes that At-LSTM, which uses financial news titles, is promising and competitive with a state-of-the-art knowledge-graph event-embedding model for directional stock prediction. It identifies headline/body content, prediction horizons, and noisy small price variations as directions for further work.
5. Conclusion
- At-LSTM uses financial news titles to predict directional movements in the S&P 500 index and individual-company stock prices, with results described as competitive with knowledge-graph event-embedding work.The cited individual-stock results list major S&P 500 companies only.Source: §5 · p. 19, Table 4 · p. 20
- Although body text performed worse than headlines in prior results, the authors leave open whether it contains useful information to extract.They also propose testing multiple prediction horizons to study how quickly information is absorbed into the market.Source: §5 · p. 20
- Because small movements in volatile financial time series may be noise, eliminating them could focus the model on news linked to significant price variation and might increase accuracy.Source: §5 · p. 20