LLM news-sentiment sector grader
A daily engine that grades every sector A–F from the news flow, and the 2.5-year backtest that shows where that signal actually works.
What the chart shows
Every trading day the grader pulls the pre-market news flow from Polygon (Benzinga, Dow Jones, Reuters, AP), scores each headline with a rules-based keyword/lexicon model (tanh-compressed, with a two-keyword-hit threshold and word-boundary guards), maps each article to its GICS sector through a ~5,800-name lookup, and rolls the day’s scores into a letter grade per sector.
The real question is whether those grades predict returns. On average, barely, across the full 2024–2026 window the overall link between sentiment and forward moves is close to zero. But the average hides the result: the response is concentrated by sector. After positive (A/B) news, Health Care and Information Technology lead (+0.6 to +0.7% over five days), Materials and Industrials follow, while Consumer Staples and Real Estate lag the market. Most of the lift is simply stocks rising on good-news days, the dispersion across sectors is the real, if modest, signal.
Measuring across 2.5 years matters: two flashy 2024 “findings”, an apparent Energy weakness and a Financials edge, both reversed by 2026, a reminder that one-month sector rankings are mostly noise. What held up is the simple shape: growth and cyclical sectors respond to good news; classic defensives don’t. A raw sentiment score is noise until you know which sector it describes, and over what window.
- Sources: Polygon news (Benzinga, Dow Jones, Reuters, AP), pre-market window [prior 3:00pm → 9:30am].
- Mapping & scoring: tickers resolved to GICS sectors via a ~5,800-name lookup (Symbols table + Yahoo); rules-based keyword/lexicon sentiment, tanh-compressed, two-keyword-hit threshold, word-boundary guards.
- Validation: 614 trading days (Jan 2024 – June 2026); forward 5-day return taken from the first open after the news window (no look-ahead), positive grades only.