Sentiment Tracker
Public-opinion sentiment scored into a YES/NO probability, compared against the live Polymarket implied probability. When our number diverges from the market with a high enough sample size, we mark the row signal. The mapping function from sentiment-index โ probability is uncalibrated: it is a logistic placeholder until enough resolved-outcome data accumulates for a Platt-scaling refit.
Pipeline: Reddit + Jina web + Polymarket comments โ Claude per-item scoring (relevance, stance toward market YES, confidence) โ weighted aggregation (relevance ร confidence ร recency-decay ร log(1+engagement)) โ logistic mapping to probability.
| Subject | Our P(YES) | Polymarket | ฮ | n | Status |
|---|---|---|---|---|---|
Eurovision 2027 host city will be Sofia Market: eurovision-2027-host-city ยท YES = Sofia | โ | โ | โ | 0 | no data |
South Korea wins Eurovision Asia 2026 Market: eurovision-asia-2026-winner-south-korea ยท YES = South Korea | โ | โ | โ | 0 | inactive |
Thailand finishes in top 3 of Eurovision Asia 2026 Market: eurovision-asia-2026-top-3-thailand ยท YES = Thailand top 3 | โ | โ | โ | 0 | inactive |
Uncalibrated probabilities. The mapping from raw weighted-stance index to probability is a logistic placeholder (k = 2.5, offset = 0). It will be re-fit against resolved-outcome data once we have โฅ50 closed snapshots. Until then, treat divergence as a discussion prompt, not a price call.
See data/sentiment/README.md for the full pipeline, tunables, and refresh schedule.