The hypothesis that news sentiment predicts short-term stock price movements is not new. Academic research dating back to the early 2000s has documented statistically significant relationships between news tone and same-day and next-day returns. But the practical application of news sentiment as a trading signal is considerably more nuanced than the research headlines suggest.
How Sentiment Analysis Works
Modern sentiment analysis of financial text typically involves three approaches:
- Rule-based / lexicon: VADER (Valence Aware Dictionary for sEntiment Reasoning) and the Loughran-McDonald financial wordlist assign positive, negative, and neutral scores to words. A sentence is scored by summing word-level scores with adjustments for negation and intensifiers. Fast, transparent, but limited in handling complex financial language.
- Machine learning classifiers: Models trained on labeled financial news (human annotators assign sentiment) using features like bag-of-words or TF-IDF. Better at capturing domain-specific language but requires large labeled datasets.
- Large Language Models (LLMs): GPT-class models can understand financial context, sarcasm, and implicit sentiment with remarkable accuracy. The drawback is cost and latency -- LLM inference at scale is expensive.
CWOS uses a hybrid approach: VADER for real-time, high-volume news scoring; LLM-based analysis for longer-form earnings call summaries and management commentary where context matters more.
What Sentiment Signals Actually Predict
Research-backed findings:
- Strong negative sentiment spikes (3+ standard deviations from baseline) are associated with same-day price declines, with varying persistence -- often 1-3 days for liquid large-caps.
- Sentiment is more predictive for direction changes than for magnitude -- knowing "this is bad news" is easier for a model than knowing "how bad."
- Sentiment signals are stronger for mid and small-caps where institutional analysts cover fewer companies, leaving more information asymmetry for sentiment tools to exploit.
- Earnings call sentiment (management tone) has been shown to predict the next quarter's earnings surprise direction with moderate accuracy.
Where Sentiment Analysis Breaks Down
- Saturation: As more quant funds run sentiment strategies, the alpha from simple sentiment signals decays. The signal-to-noise ratio deteriorates as the trade becomes crowded.
- Ambiguity: "TCS announces major restructuring" could be positive (efficiency) or negative (layoffs signal softening demand). Simple lexicon models often get this wrong.
- Speed: By the time news reaches a sentiment model, the fastest market participants (HFT algorithms with direct exchange feeds) have already traded on the information. Retail-facing sentiment tools are capturing the delayed price impact, not the initial reaction.
- Fake news and market manipulation: Coordinated social media sentiment manipulation (pump-and-dump via WhatsApp groups or Telegram channels) can generate false sentiment signals that look real to naive models.
- Regulatory events: SEBI orders, income tax raids, ED investigations generate negative sentiment but also trading halts -- the price impact cannot be captured until trading resumes, and the first-day reaction is often extreme.
Practical Use in a Research Workflow
The most effective use of news sentiment for non-HFT investors is not as a standalone trading trigger but as a filter and research prioritization tool:
- Sentiment spike alerts direct your attention to companies worth examining further -- they don't tell you what to do.
- Persistent negative sentiment (5+ days of below-baseline scores) on a stock you hold warrants a deeper review of the underlying news, not an automatic sell.
- Sentiment combined with price and volume action (sentiment negative + price holding up = potential "stock climbing a wall of worry"; sentiment positive + price declining = potential distribution) is more informative than sentiment alone.
CWOS's news sentiment scores on the Research Terminal are updated multiple times daily and include the headline count driving the score -- helping you distinguish between one sensational article and a genuine wave of coverage.