README.md

metadata

title: TradeVerse AI Trading Experiment
emoji: 📈
colorFrom: green
colorTo: blue
sdk: gradio
sdk_version: 4.44.0
app_file: app.py
pinned: false
license: mit

TradeVerse AI Trading Experiment

A psychology research tool for studying human-AI interaction in financial decision-making.

Overview

This application combines:

• Trading Game: Scenario-based trading decisions in a fictional financial universe
• AI Chatbot: RAG-based advisor with tunable parameters
• Experiment Tracking: Comprehensive data collection for research analysis

Features

For Participants

• Interactive trading scenarios with fictional companies
• AI advisor that provides recommendations and answers questions
• Adjustable AI settings (explanation depth, communication style)
• Real-time portfolio tracking

For Researchers

• Pre-defined experimental conditions for A/B testing
• Hidden parameters (accuracy rate, confidence framing, risk bias)
• Comprehensive tracking of:
  • Trading decisions and outcomes
  • AI consultation patterns
  • Trust metrics and confidence changes
  • Response timing
• Data export utilities (CSV, Excel reports)

Experimental Conditions

Condition	Accuracy	Confidence	Description
high_accuracy_high_confidence	85%	High	Correct AI, assertive
high_accuracy_low_confidence	85%	Low	Correct AI, hedging
low_accuracy_high_confidence	40%	High	Often wrong, assertive
low_accuracy_low_confidence	40%	Low	Often wrong, hedging
conservative_advisor	70%	Medium	Risk-averse recommendations
aggressive_advisor	70%	Medium	Risk-seeking recommendations
passive_advisor	70%	Low proactivity	Rarely initiates
active_advisor	70%	High proactivity	Frequently initiates

Setup

LLM Provider Options

The chatbot supports multiple LLM providers (auto-detected based on available API keys):

Provider	Env Variable	Cost	Notes
HuggingFace	HF_TOKEN	Free tier available	Recommended for HF Spaces
DeepSeek	DEEPSEEK_API_KEY	Pay-per-use	Higher quality
Fallback	None needed	Free	Rule-based, limited

Recommended HuggingFace Models (free tier)

# In chatbot.py, change DEFAULT_HF_MODEL to any of these:
"Qwen/Qwen2-1.5B-Instruct"             # Smallest, fastest
"microsoft/Phi-3-mini-4k-instruct"     # Small, good quality
"HuggingFaceH4/zephyr-7b-beta"         # Default, best quality
"mistralai/Mistral-7B-Instruct-v0.2"   # Popular alternative

Environment Variables

Option 1: HuggingFace (Recommended for HF Spaces)

export HF_TOKEN="hf_your_token_here"

Get your token at: https://huggingface.co/settings/tokens

Option 2: DeepSeek

export DEEPSEEK_API_KEY="your-api-key-here"

Local Development

pip install -r requirements.txt
python app.py

Hugging Face Spaces

1. Create a new Space with Gradio SDK
2. Upload all files
3. Add HF_TOKEN as a secret in Space settings (Settings → Repository secrets)
   • Your HF token works automatically in Spaces

Data Export

# Quick statistics
python export_data.py stats

# Export all data to CSV
python export_data.py export

# Generate full Excel report
python export_data.py report

Project Structure

IntegratedGame/
├── app.py                 # Main Gradio application
├── trading.py             # Trading engine and portfolio management
├── chatbot.py             # RAG chatbot with parameter-aware responses
├── tracking.py            # Database and experiment tracking
├── config.py              # Scenarios, conditions, parameters
├── export_data.py         # Data analysis and export
├── requirements.txt
├── knowledge_base/
│   ├── companies.txt      # Fictional company profiles
│   ├── market_context.txt # TradeVerse economic backdrop
│   ├── trading_basics.txt # Trading fundamentals
│   └── ai_persona.txt     # AI behavior guidelines
└── db/
    └── experiment.db      # SQLite database (auto-created)

Database Schema

sessions

• Participant sessions with portfolio tracking
• AI reliance metrics

decisions

• Individual trading decisions
• AI parameters at decision time
• Timing and confidence data

chat_interactions

• All AI-participant conversations
• Proactive vs reactive classification
• Engagement metrics

trust_metrics

• Pre/post advice confidence
• Advice adherence
• Outcome tracking

Research Applications

This tool supports research into:

• Trust calibration in AI advisors
• Effect of AI confidence framing on decision-making
• Proactive vs reactive AI assistance preferences
• Risk perception under AI guidance
• Learning and adaptation to AI accuracy

License

MIT License