KILLED · noise + bots
Hypothesis
Aggregate sentiment polarity from crypto X posts (transformer classifier) predicts short-horizon returns.
Math — net sentiment
$$ S_t = \frac{1}{N_t}\sum_{i=1}^{N_t}\,\text{polarity}(\text{post}_i)\in[-1,1] $$
Method
FinBERT-style polarity on filtered cashtag streams, hourly aggregate vs next-hour return. Bot/spam filtering attempted.
Results
| Information coefficient | ~0.01 |
| Bot / shill contamination | severe |
| Endogeneity (sentiment = price) | high |
Sentiment is dominated by bots, shills, and reflexivity (sentiment is mostly a lagged function of price). After de-botting, IC ~0.01 — indistinguishable from zero. The clean signal, if any, is buried under manufactured noise.
Social sentiment in crypto is an adversarial, reflexive dataset — much of it is paid or automated, and the rest just echoes the price. Treat it as noise until proven otherwise.