Skip to content

SKcoder6344/Trader_Sentiment_Analysis

Repository files navigation

Trader Performance vs Market Sentiment

Hyperliquid × Fear/Greed Index


Overview

This project analyzes how Bitcoin market sentiment (Fear/Greed) relates to trader behavior and performance on Hyperliquid. The goal is to uncover patterns that could inform smarter, sentiment-aware trading strategies.


Datasets

Dataset Source Size
Hyperliquid Trader Data Provided (.csv.gz) 211,224 trades, 32 accounts
Bitcoin Fear/Greed Index Alternative.me 2,644 daily readings

Date range after merge: May 2023 – May 2025 (479 days, 211,218 trades retained)


Project Structure

├── MarketSentiment.ipynb             # Main notebook
├── fear_greed_index.csv              # Sentiment dataset
├── README.md                         # This file
├── chart1_performance_fear_vs_greed.png
├── chart2_behavior_fear_vs_greed.png
├── chart3_segments.png
├── chart4_segment_behavior.png
├── chart5_insights.png
├── chart6_clusters.png
└── chart7_model.png

Setup & How to Run

  1. Clone this repository
  2. Upload compressed_data.csv.gz and fear_greed_index.csv to your Google Drive
  3. Open trader_sentiment_analysis.ipynb in Google Colab
  4. Update the file paths in Block 2 to match your Drive location:
TRADER_FILE    = '/content/drive/MyDrive/compressed_data.csv.gz'
SENTIMENT_FILE = '/content/drive/MyDrive/fear_greed_index.csv'
  1. Run all cells top to bottom (Runtime → Run All)

Required libraries: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn — all available by default in Google Colab.


Methodology

1. Data Preparation

  • Cleaned comma-formatted numeric columns in the trader dataset
  • Parsed IST timestamps into daily date keys for alignment
  • Merged both datasets on date using an inner join — 6 rows lost (<0.01%)
  • Simplified sentiment into Fear vs Greed by collapsing Extreme labels, retaining the full classification for reference

2. Feature Engineering

  • Built daily per-account metrics: total PnL, win rate, trade frequency, estimated leverage, long/short ratio, consistency score (% of profitable days), and a Sharpe-like ratio (mean return / std of returns)
  • Estimated leverage as Size USD / abs(Start Position), capped at 200x to remove data artifacts

3. Analysis

  • Compared performance and behavior across Fear vs Greed days
  • Used Mann-Whitney U test for statistical significance (non-parametric — PnL distributions are heavily skewed)
  • Segmented traders into three schemes: Leverage, Frequency, and Winner Type

4. Bonus

  • KMeans clustering with silhouette-score-based k selection to identify behavioral archetypes
  • Logistic regression to predict next-day trader profitability using today's behavior and sentiment features

Key Insights

Insight 1 — Greed days produce consistent profits; Fear days produce outliers

Median PnL on Greed days ($265) is more than double Fear days ($123), and losing days drop from 12% to 7.6%. However, mean PnL is actually higher on Fear days ($5,185 vs $4,144) — driven by a small number of traders making outsized gains during volatile conditions. The difference is borderline significant (Mann-Whitney p = 0.062), meaning the edge is real but not guaranteed.

Insight 2 — Traders get more aggressive during Fear, not less

On Fear days, traders place more trades (31 vs 28), use higher leverage (7.7x vs 6.4x), and generate higher daily volume ($83,640 vs $61,128). This is counter-intuitive — fear does not cause retreat, it triggers overactivity. Only per-trade size drops slightly ($1,854 vs $2,005), the single sign of caution.

Insight 3 — Win rate is a misleading performance metric

Unprofitable traders in this dataset have a 72.9% win rate but negative total PnL. Consistent Winners have an 83.4% win rate and strongly positive total PnL. The difference is risk management — unprofitable traders win small and lose large. Consistency score (% of profitable days) predicts total PnL far more reliably than win rate alone.


Trader Segments

Segment Traders Median PnL Win Rate Consistency
High Leverage 16 $120,598 73.3% 58.9%
Low Leverage 16 $114,961 67.2% 45.0%
Frequent 16 $129,522 73.3% 67.0%
Infrequent 16 $106,037 69.9% 47.6%
Consistent Winner 13 $168,658 83.4% 77.1%
Profitable but Inconsistent 16 $117,655 58.9% 47.6%
Unprofitable 3 -$70,436 72.9% 42.4%

Behavioral Archetypes (KMeans Clustering)

Silhouette score selected k=2 as optimal.

Archetype A — High-Output Specialists (2 traders) 457 trades/day, $1.27M median PnL, 86% win rate, 19x leverage. Frequency and consistency far exceed the rest of the dataset. Almost certainly algorithmic or semi-automated traders.

Archetype B — Passive Participants (30 traders) 58 trades/day, $107k median PnL, 71.5% win rate, 31x leverage. Active human traders who compensate for lower frequency with higher leverage.


Predictive Model

A logistic regression model predicts whether a trader will be profitable the next day using today's behavior and sentiment as features.

Metric Value
ROC-AUC 0.6913
Accuracy 71%
Training samples 1,023
Test samples 256

Top predictors:

  • Fear/Greed index value (+0.49) — strongest positive predictor
  • Sentiment direction / Greed label (-0.30) — negatively correlated, suggesting Greed days can precede pullbacks
  • Win rate (+0.24) and trade frequency (+0.23) both improve next-day profitability odds

The raw Fear/Greed index value is the strongest predictor of next-day profitability, while the binary sentiment direction is negatively correlated — indicating that the magnitude of sentiment matters more than its label alone.

Note: The model is conservative and biased toward predicting profitable outcomes (recall of 12% for the Not Profitable class), reflecting the class imbalance in the dataset (69% profitable days).


Strategy Recommendations

Rule 1 — On Fear Days: Trade more, size down Frequent traders earn $330 median PnL on Fear days vs just $1 for infrequent traders. The edge on Fear days comes from activity and adaptability, not from large bets. Increase number of trades, reduce per-trade USD exposure.

Rule 2 — On Greed Days: Scale up position size Low leverage traders see a $230 Greed-day lift vs only $70 for high leverage traders. Greed days are trending markets — even modest sizing captures the move effectively. Increase position size and prioritize trend-following setups.

Rule 3 — Manage risk/reward, not win rate Unprofitable traders win 72.9% of trades but still lose money. Consistent Winners win 83.4% and are robustly profitable. The differentiator is not entry skill — it is cutting losses faster and letting winners run. Implement hard stop-losses and asymmetric position sizing.


Releases

No releases published

Packages

 
 
 

Contributors