📊 What I Learned Analyzing 100K+ Crypto Trades Using Market Sentiment

Udit Jain — Wed, 15 Apr 2026 12:19:26 +0000

Most traders think market sentiment (Fear & Greed) directly drives profits.

So I decided to test it.

I analyzed 100K+ trades from Hyperliquid and combined it with the Bitcoin Fear & Greed Index to uncover how sentiment actually impacts trading performance.

Here’s what I found 👇

🧠 The Question

Does market sentiment really affect:

Profitability?
Win rate?
Trader behavior?

Or is it just noise?

⚙️ What I Did
Cleaned and processed trading data
Filtered only closed trades (real profits/losses)
Merged trades with daily sentiment
Engineered features like:
Win rate
ROI
Segmented traders into:
Top performers
Worst performers
📈 Key Findings

Greed = Profits… but only for some Extreme Greed had the highest average PnL and win rate But here’s the twist:

👉 Top traders made the most money
👉 Worst traders LOST the most money

Same market. Different outcomes.

The Overconfidence Trap

Worst traders:

Had ~55% win rate in Greed (not terrible)
But still lost money

Why?

👉 They over-leveraged and mismanaged risk

This suggests:

Losses weren’t due to bad predictions, but bad execution.

Fear Isn’t Always Bad Traders performed surprisingly well in Fear phases Likely because: Only high-conviction trades were taken Risk exposure was lower

👉 Discipline > sentiment

Top Traders Are Consistent

Top performers:

Stayed profitable across all sentiment conditions
Didn’t rely on “market mood”
Focused on execution + risk management

Timing Beats Sentiment (Advanced Insight)

Instead of just sentiment, I analyzed sentiment transitions:

👉 Example:

Fear → Greed
Greed → Fear

Result:

Transitions often produced better outcomes than static sentiment.

🧠 What This Means

Market sentiment doesn’t determine success.

👉 Trader behavior does.

Weak traders react emotionally
Strong traders stay consistent
🚀 Practical Takeaways

If you trade (or build trading systems):

Don’t blindly follow bullish sentiment
Control position size during Greed
Avoid overtrading in “easy markets”
Focus on consistency, not prediction
🏁 Final Thought

The market doesn’t reward sentiment.
It rewards discipline.

🔗 Built Using
Python (Pandas, Matplotlib)
Real trading + sentiment data

If you're into data science, trading, or building in Web3—would love to hear your thoughts 👇

Building a Voice-Controlled AI Agent with Real-Time Intent Execution

Udit Jain — Sun, 12 Apr 2026 13:53:04 +0000

Building a Voice-Controlled AI Agent for Real-Time Intent Execution

🚀 Overview

I built a voice-controlled AI agent that can take audio input, understand user intent, execute local actions, and display results through a web interface.

The goal was to design an end-to-end system that connects speech processing with intelligent execution.

🧠 Architecture

This modular pipeline design allows each component (STT, LLM, execution) to be independently optimized and replaced, which is a common approach in production voice AI systems.

The system follows a simple pipeline:

Audio → Speech-to-Text → Intent Classification → Tool Execution → UI

Each component is modular and communicates sequentially, making the system easy to debug and extend.

🎤 Speech-to-Text

For converting audio to text, I used Groq’s Whisper-based API.

Although the assignment preferred local models, I initially attempted to run local Whisper models but faced RAM limitations. To ensure stable performance, I switched to an API-based solution, which provided fast and reliable transcription.

🤖 Intent Understanding

The transcribed text is processed using a language model to classify intent into:

Create file
Write code
Summarize text
General chat

I also added simple rule-based overrides to improve accuracy for code-related requests.

⚙️ Tool Execution

Based on the detected intent, the system performs actions such as:

Creating files (restricted to a safe output folder)
Generating executable code using an LLM
Summarizing text
Handling conversational queries

This layer connects AI decisions with real system operations.

🖥️ User Interface

The frontend is built using Streamlit and displays:

Transcription
Detected intent
Action details
Final output

This ensures full transparency of the pipeline.

🔥 Key Enhancements

Human-in-the-Loop: Confirmation before file operations
Session Memory: Tracks past interactions
Context-Aware Chat: Maintains conversational continuity
Error Handling: Graceful failure management

⚡ Challenges

Running local models under hardware constraints
Ensuring clean code generation without extra formatting
Designing reliable intent classification
Handling audio input and system safety

🎯 Conclusion

This project demonstrates how to design a practical AI agent by combining speech processing, language understanding, and real-world execution. It highlights the importance of modular architecture, system safety, and user interaction in building reliable AI systems.

🔗 Links

GitHub: github.com/uditjainofficial/assignment-voice-controlled-ai-agent
Demo Video: youtube.com/watch?v=6frrIILn5BQ&t=5s

Forem: Udit Jain