Context
This started as a hiring assignment for Primetrade.ai. The brief was open: take real Hyperliquid trade data, take the Bitcoin Fear and Greed index, and find out if the two say anything to each other. Sentiment indices are everywhere in crypto and mostly used as vibes, so the goal was to put a number on it.
Problem
The trap with this kind of analysis is reading a story into noise. Markets are noisy, sentiment labels are coarse, and it is easy to find a pattern that vanishes the moment you split the data differently. I needed findings that survived being measured properly, not just a chart that looked suggestive.
Approach
I joined per-trade records to the daily sentiment regime, then computed the metrics that actually matter to a trader for each regime: win rate, profit factor, Sharpe, max drawdown, ROI. The helper code is split into a loader, a feature builder, and a metrics module so the numbers are reproducible rather than hand-tuned in a notebook cell. Where a claim needed a significance check, I ran one instead of eyeballing it.
Outcome
A few results held up cleanly. Extreme Greed was the best regime for this cohort, 89.2 percent win rate with a profit factor of 11.0. Plain Greed, oddly, was the worst, carrying the lowest Sharpe at 3.41 and the deepest drawdown near 419k, which is a reminder that "greed" is not one thing. Positioning shifted too: long share fell from 51.1 percent in Extreme Fear to 44.9 percent in Extreme Greed, and that shift was nowhere near chance with a chi-square p below 1e-68. Twenty-eight traders qualified for a closer contrarian cut, some with fifty point win-rate gaps between regimes.
Next Iteration
- Track each trader through time instead of pooling, to separate skill from regime luck.
- Build a simple regime-aware position signal and backtest it honestly with costs.