News and text sentiment
SentimentAdvanced
Sentiment strategies turn news and filings into a score and trade the score. Early work counted pessimistic words in a widely read column and found a short-lived pressure on prices. Later work trains the word weights on returns instead of using a fixed dictionary. The text is an input. It is not a trading system until it has a timestamp, a horizon, and a cost assumption.
Not investment advice. Nothing here is a recommendation to buy or sell anything. Backtests and historical studies are not guarantees, and a rule that worked in one sample can fail in the next. Trading can lose money. Links are for education, not endorsements. Do your own research.
- Rules
- Ingest items with the time they were actually available. Score them with a dictionary or a supervised model, aggregate to a stock or to the index, and trade in the direction of the score over a horizon you chose in advance, often a day to a few days. Neutralize market and sector exposure if you want a stock-specific bet. Ignore text the model could not have seen, including headlines that were edited later.
- Where AI fits
- A modern language model can extract events, guidance changes, and uncertainty better than a word list, and it can explain the score in plain language. The model should not place the order. A separate, tested mapping from score to position should do that, with the LLM's output logged. Prompt changes are parameter changes and belong in the backtest.
- Markets
- Liquid stocks with a real news feed. Index-level sentiment is cleaner and slower. Social-media text is noisier and easier to manipulate.
- Timeframes
- Intraday to a few days for news. Filings can justify a longer hold. The edge, where research finds one, decays quickly.
- Risks
- Most price reaction is already in the print by the time a retail feed delivers the story. Sarcasm, boilerplate, and recycled articles fool word counts. A model trained on today's phrasing will not match the phrasing of ten years ago, so a naive backtest leaks the future.