Author: Hiral Sarkar | MSc Global Financial Markets | PG Data Science & AI
Dark-theme analytics dashboard -- 6 KPI cards, 8 interactive charts, regime breakdown bars, account performance tables. Runs instantly in any browser, no install required.
Retail and professional crypto traders make sizing and directional decisions continuously across varying market sentiment regimes. A common assumption is "buy the fear, sell the greed" -- but does the data actually support regime-based trading edge?
This project merges 211,224 Hyperliquid trades across 32 accounts (2023-2025) with the Bitcoin Fear & Greed Index to quantify how sentiment regime affects win rate, average PnL per trade, position sizing, and fee efficiency. The goal: extract actionable rules a trader or fund manager can act on immediately.
Derived directly from 211K+ trades across 5 sentiment regimes:
- Extreme Greed is the alpha regime: highest win rate (46.5%) and highest avg PnL/trade ($67.9) with the smallest average position size ($3.1K). The edge is concentrated in the euphoria tail -- not broad "Greed."
- Extreme Fear is the worst regime: lowest win rate (37.1%) and lowest avg PnL/trade ($34.5). The data offers no "buy the dip" alpha here.
- Inverted sizing in Fear is the #1 problem: plain Fear shows the largest average trade size ($7.8K) despite only 42.1% win rate -- a capital-destructive mismatch.
- The 0-100 score is nearly useless: raw Fear & Greed value correlates with daily PnL at only r = -0.08. The categorical regime label -- especially "Extreme" buckets -- is what actually matters.
- PnL is heavily concentrated: top 5 of 32 accounts generated $6.36M (~62%) of the $10.25M total. Aggregate-level win rates mask enormous per-account divergence.
- Greed increases sell pressure: sell mix rises to 52.9% in Greed vs ~50% in Fear -- traders are instinctively scaling out into strength, which is correct directionally.
- Fee drag hits hardest in Fear: avg fee per trade is $1.50 in Fear vs $0.68 in Extreme Greed -- over-trading in bad regimes compounds the loss.
6 KPI cards -- Total PnL, best/worst regime win rates, inverted sizing risk, PnL concentration
Daily Closed PnL vs Fear & Greed Index value -- correlation barely -0.08; regime label matters, raw score does not
Top and bottom 10 accounts -- top 5 drive 62% of total PnL; bottom 5 are consistent drags
| Priority | Action | Rationale |
|---|---|---|
| 1 -- Immediate | Size up 30-50% in Extreme Greed | Win rate 46.5%, avg PnL/trade $67.9, current size already smallest ($3.1K) -- clear room to add |
| 2 -- Immediate | Cut size 30-50% in Extreme Fear | Worst win rate (37.1%) and worst PnL/trade ($34.5) -- less is more |
| 3 -- Critical | Fix inverted sizing in plain Fear | $7.8K avg size at 42.1% WR is the single biggest capital efficiency problem |
| 4 -- Tactical | Reduce trade frequency in Fear | $1.50 avg fee in Fear vs $0.68 in Extreme Greed -- over-trading in bearish regimes destroys net PnL |
| 5 -- Analytical | Build per-coin, per-regime playbooks | TRUMP: +$81 avg PnL in Fear vs -$454 in Greed. FARTCOIN: opposite. Universal rules miss this |
| 6 -- Structural | Audit bottom-5 accounts | Consistent drags while top 5 drive 62% of profit -- reallocate or retrain |
Trader-Sentiment-Analysis/
|
+-- Sentiment_Trader_Analysis.ipynb <- Full EDA, merging, analysis, visualisations
+-- app.py <- Streamlit interactive dashboard
+-- dashboard.html <- Static dashboard (GitHub Pages, no install)
+-- dashboard_data.js <- Pre-aggregated JSON data for the HTML dashboard
+-- data/
| +-- historical_data.csv.gz <- 211,224 Hyperliquid trades, 32 accounts (gzipped)
| +-- fear_greed_index.csv <- Bitcoin Fear & Greed Index, daily (2018-2025)
+-- images/ <- Dashboard screenshots
+-- requirements.txt
+-- README.md
| Layer | Tool | Detail |
|---|---|---|
| Data Merging | Python / Pandas | 211K trade rows merged with daily F&G index on date key |
| Analysis | NumPy / SciPy | Regime aggregation, correlation, win rate calculation |
| Notebook Viz | Matplotlib / Seaborn / Plotly | EDA charts, regime breakdowns, account analysis |
| Interactive App | Streamlit | Full Python dashboard with live filters |
| Static Dashboard | Chart.js 4.4 + HTML/CSS | Dark-theme GitHub Pages dashboard, zero dependencies |
Open dashboard.html directly in any browser -- no install needed. Or view the live GitHub Pages version.
pip install -r requirements.txt
streamlit run app.pypip install -r requirements.txt
jupyter notebook Sentiment_Trader_Analysis.ipynbRun all cells top to bottom. The notebook contains the full merge logic, regime analysis, account-level breakdowns, and per-coin regime cross-tab.
| File | Description |
|---|---|
data/historical_data.csv.gz |
211,224 Hyperliquid trades, 32 accounts (2023-2025). Columns: account, coin, side, size, price, closed_pnl, fee, timestamp |
data/fear_greed_index.csv |
Bitcoin Fear & Greed Index daily values + categorical label (2018-2025). Source: alternative.me |
The two datasets are merged on date (trade date = F&G date) to produce a regime-labelled trade table.
- Data loading & validation -- shape, dtypes, nulls, date range alignment
- Regime merge -- join trades to daily F&G index; assign 5-bucket and 3-bucket regime labels
- Univariate regime analysis -- trade count, PnL, win rate, avg size, fee per regime
- Correlation analysis -- Pearson r between daily PnL and F&G raw value
- Directional bias -- buy/sell mix by regime
- Per-coin regime breakdown -- avg PnL/trade for top 10 coins across Fear/Neutral/Greed
- Account-level analysis -- total PnL, win rate, trade count per account; top/bottom 10
- Concentration analysis -- PnL breakdown by account rank
- Recommendations -- regime-specific sizing rules, fee reduction, per-coin playbooks
- Merged and analysed 211,224 crypto trades across 32 accounts with an external sentiment index -- end-to-end in Python
- Quantified regime-level trading edge: Extreme Greed produces 46.5% win rate and $67.9 avg PnL/trade vs 37.1% and $34.5 in Extreme Fear
- Identified the inverted sizing problem -- traders use largest position sizes in the worst-performing regime (Fear, $7.8K avg)
- Proved raw sentiment score is a weak predictor (r = -0.08); categorical regime label is the actionable signal
- Built both a Streamlit app and a static GitHub Pages dashboard from the same dataset
- Surfaced concentration risk: top 5 of 32 accounts drive 62% of total PnL -- critical for portfolio allocation decisions
Built to demonstrate quantitative analytics and data storytelling for trading, fintech, and AI model risk roles.