GPT: Mean Reversion strategy in Python makes 813%

Mean Reversion Trading Strategy Leveraging AI for Python Implementation

This video presents a meticulously crafted and backtested Mean Reversion trading strategy, implemented in Python with the assistance of an AI-powered Large Language Model (LLM). Designed to capitalize on price pullbacks towards an established mean, the strategy exhibits remarkable performance across multiple cryptocurrency assets over a 3.5-year backtesting period. The methodology underscores how even individuals with limited programming expertise can harness AI for sophisticated algorithmic trading research. 📈

The core of the strategy is Mean Reversion, diverging from trend-following approaches by aiming to profit from price reversals. Positions are taken contrary to the immediate price direction, anticipating a return to the mean. The strategy employs a multi-timeframe analysis, primarily utilizing a 1-hour (1H) timeframe for trade execution and a 4-hour (4H) timeframe for broader trend and volatility confirmation.

Entry Filters and Rules:

• Bollinger Bands (1H): This serves as the primary entry signal. For a long position, a buy order is placed at the lower Bollinger Band, signifying a significant deviation below the mean. Conversely, for a short position, an order is placed at the upper Bollinger Band. These bands are set at two standard deviations from a moving average, leveraging the statistical probability of price reverting to the mean after such excursions.
• Relative Strength Index (RSI) (4H): Unconventionally, the RSI is employed on the higher, 4-hour timeframe to identify strong underlying trends for "pullback" entries. For a long trade, the 4H RSI must be above 70, indicating an overbought condition on the longer timeframe, which the strategy interprets as a strong bullish trend. This allows the 1H Bollinger Band entry to act as a pullback within that established bullish context. The opposite applies for short positions (4H RSI below 30).
• Average Directional Index (ADX) (1H & 4H): ADX is used to confirm the presence of sufficient volatility and nascent trend strength, critical for mean reversion trades. A lower threshold of 20 is set for the 1H ADX to allow for earlier entries, while a higher threshold of 40 is applied to the 4H ADX, ensuring a more robust underlying volatility or trend strength on the longer timeframe. This dual ADX approach aims to filter out choppy, untradeable market conditions while still catching early momentum.
• SuperTrend (4H): This indicator acts as a final trend filter, also applied to the 4-hour timeframe, but with a nuanced application tailored for mean reversion. For a long entry, the 4H SuperTrend must indicate an uptrend (price above the SuperTrend line), meaning the 1H Bollinger Band lower band entry is a temporary dip or pullback within an overarching 4H uptrend. This ensures positions are taken against short-term movements but aligned with the dominant higher-timeframe trend direction. The SuperTrend value is programmatically represented as "1" for an uptrend and "-1" for a downtrend.

Exit Rules:

• Take Profit: The take-profit target is set at the opposite Bollinger Band. For long positions, profit is taken at the upper Bollinger Band, while for short positions, it is taken at the lower Bollinger Band. This aligns with the strategy's mean reversion core, where the price is expected to return to and potentially overshoot the mean.
• Stop Loss: A dynamic stop-loss mechanism is employed using the Average True Range (ATR). The stop loss is set at the entry price minus six times the current ATR. This relatively wide stop loss is a characteristic feature of mean reversion strategies, which often tolerate larger temporary drawdowns in exchange for a higher win rate, recognizing that winning trades typically yield smaller profits while losing trades, though less frequent, can be larger. Initial risk was set at 3% of the account per trade, though later adjusted for backtesting visualization.

Implementation and Backtesting Workflow:

The strategy's rules were communicated to an LLM, specifically Claude 3.7 Sonnet via JesseGPT, which then generated the initial Python backtesting code. This process highlights the evolving role of AI in accelerating algorithmic trading development, enabling rapid prototyping and iteration. The generated code underwent review and minor modifications, notably correcting the SuperTrend logic for distinguishing uptrends ("1") from downtrends ("-1") and removing an unintended update_position function that was continuously adjusting the take-profit level, which was not part of the initial design. Position sizing, initially set to risk 3% per trade, was later multiplied by five during backtesting to better illustrate market-beating potential, with the caveat that this altered the initial risk profile.

Backtesting Results:

Initial backtests on Bitcoin (BTC) from early 2024 to 2025 showed a perfect 100% win rate and a Sharpe Ratio of 2.48, but with limited trades (9) and modest P&L (9.23%). Expanding the backtest to cover 3.5 years (since early 2022) revealed a more nuanced performance:

• BTC-Only (3.5 years, post-correction): Achieved a 253% P&L, a maximum drawdown of -22% (significantly lower than BTC's market drawdown), a 94.59% win rate, and a Sharpe Ratio of 1.57. The average holding time was approximately 33 hours. However, a key limitation was the low trade frequency (37 trades over 3.5 years), raising concerns about the statistical significance of the results.

To address the low trade frequency and enhance statistical robustness, the strategy was expanded to include additional cryptocurrencies: Ethereum (ETH) and Dogecoin (DOGE). This multi-asset backtest across BTC, ETH, and DOGE yielded substantially improved overall results:

• Multi-Asset (BTC, ETH, DOGE; 3.5 years):
  • P&L: An impressive 813% 💰
  • Maximum Drawdown: Maintained at -22% (demonstrating resilience across assets).
  • Win Rate: 92% (slightly lower but still exceptionally high).
  • Sharpe Ratio: 1.95 (indicating strong risk-adjusted returns).
  • Total Closed Trades: Increased to 75, significantly enhancing statistical significance.
  • Average Holding Time: Approximately 70 hours (nearly three days).

The higher average holding time (70 hours) presents a potential psychological challenge for traders, as positions might be held at a loss for extended periods before reverting to profit. Despite this, the strategy demonstrates a robust equity curve and strong overall performance metrics.

Final Takeaway:

This Mean Reversion strategy, developed and refined with AI assistance, presents a compelling case for its efficacy in cryptocurrency markets. Its high win rate, exceptional Sharpe Ratio, and remarkable P&L of 813% over 3.5 years, coupled with a contained maximum drawdown of -22%, make it a standout performer. The comprehensive multi-timeframe filtering using Bollinger Bands, RSI, ADX, and SuperTrend creates a robust framework for identifying high-probability pullback entries. While the extended average holding time (70 hours) could be a psychological hurdle, the strategy's statistical robustness, bolstered by multi-asset trading, offers strong evidence of its potential. This work exemplifies the powerful synergy between sophisticated trading logic and advanced LLMs in quantitative finance research. 🛠️📊🧠