When Sentiment Diverges From Score
Constituents where 60-day sentiment moved opposite to the underlying CIP signal — and the historical hit rate.
Abstract
When the 60-day NLP sentiment signal moves opposite to the underlying CIP score by more than 15 points, we call it a divergence. We document the current eighteen divergences and the historical mean reversion pattern from 2019-2025.
Key Findings
- 01Eighteen current divergences: twelve negative-sentiment / positive-score, six positive-sentiment / negative-score.
- 02Historical hit rate: score wins 68% of episodes over a six-month horizon.
- 03Median time-to-convergence: 84 trading days.
- 04The asymmetry is real — negative-sentiment / positive-score episodes mean-revert faster than the inverse.
1. Defining divergence
We standardize both signals to a 0-100 scale and call a divergence when the absolute gap exceeds 15 points and the gap has persisted for at least 20 trading days. The 15-point threshold was chosen to keep the false-positive rate below 10% in our 2019-2025 backtest.
2. The asymmetry
Markets are faster to punish bad news than to reward quiet improvement. Negative-sentiment / positive-score episodes converge in a median 71 days; positive-sentiment / negative-score episodes take 96.
Implications
- →Divergence is a signal, not a trade. Use it to prioritize which constituents to re-research, not as a stand-alone factor.
- →We do not publish the eighteen names ahead of the next ESG Research Drop on 2026-04-15.
Methodology
- • Sentiment: 60-day rolling NLP over filings, NGO reports, and litigation databases.
- • Score: composite CIP, refreshed monthly.
- • Backtest: 2019-01 to 2025-12, monthly snapshots.
Sources & References
- CIP Sentiment Engine v2.3 (internal)
- Refinitiv News Archive
This research note is published by the Carbon Index Protocol research desk for informational purposes only. It does not constitute investment advice, an offer to sell, or a solicitation to buy any security or carbon instrument. C100 scores and weights are derived from public filings, registry data, and proprietary models; figures may be revised as new information becomes available.