Claude AI vs ChatGPT: Which AI Trades the Most Profitable?
Testing AI Trading Strategies with ChatGPT and Claude
Overview of the Tests
- The video outlines three tests to evaluate the trading strategies generated by ChatGPT and Claude, focusing on BTC trading on a 1-hour timeframe.
- The first test involves generating a profitable trading strategy using only the AI's thinking model without any external tools.
Test 1: Basic Strategy Generation
- Both ChatGPT and Claude are prompted to create a simple trading strategy in Pine Script for BTC/USDT without backtesting capabilities.
- The results from both models are expected to be quick as they rely solely on their internal logic.
Test 2: Enhanced Strategy with Tools
- In the second test, both AIs are given access to TradingView for backtesting and optimizing their strategies based on performance metrics.
- The goal is to improve upon initial strategies by incorporating better indicators and refining them through backtesting.
Test 3: Automated Trade Execution
- The final task involves scanning all coins on Bybit for high-probability trades, where both AIs will execute trades autonomously without human intervention.
- Each AI is instructed to check account balances and identify potential trade setups, including stop-losses and take profits.
Insights into AI Performance
Comparison of Models
- David shares his experience with both models, noting that Claude has been favored until recent updates made ChatGPT more competitive in coding tasks.
- He highlights differences in user interface preferences between the two platforms, indicating that Claude offers a more integrated workspace compared to ChatGPT's setup.
Initial Strategy Results
- After executing the first round of strategies, neither model produced favorable results; both showed negative returns after testing their scripts in TradingView. ChatGPT resulted in -3% while Claude yielded -2%.
Final Evaluation of Strategies
Backtested Results
- After allowing both models access to TradingView for optimization:
- Claude improved its win rate from 32% to 55%, reducing total trades significantly while achieving a profit factor of 1.18 despite still being low overall.
- Max drawdown decreased from 81% to 17%.
- ChatGPT achieved a net profit of 48% but had only a win rate of 23%, with similar reductions in closed trades compared to its initial attempt.
- Its max drawdown was recorded at 23%.
Conclusion on Trading Efficacy
- Despite improvements through backtesting, both models ultimately led to losses when real-time trades were executed.
- Combined losses amounted approximately $65 across various positions taken during testing sessions.
- This suggests limitations when relying solely on large language models for financial decision-making without additional strategic frameworks or data inputs.