Adaptive Trend Following Strategy: Kalman Filter & Meta-Labeling

📖 Overview

This project presents a quantitative trading strategy for ETH/USDT designed to address the Non-Stationary (Non-IID) nature of financial markets. Unlike traditional Moving Averages which suffer from lag, this system utilizes a Dual Kalman Filter approach for adaptive trend detection.

To mitigate drawdown risks inherent in trend-following systems, the strategy incorporates a Machine Learning (Random Forest) layer for Meta-Labeling. This allows for Dynamic Leverage adjustment based on the predicted probability of signal success.

Key Performance (Out-of-Sample 2023-Present):

• Sharpe Ratio: 1.03
• Max Drawdown: Reduced from -73% (Fixed) to -52% (ML Optimized).

🛠️ Key Features

1. Adaptive Signal Generation (Dual Kalman Filter)

• Treats the "true price" as a hidden state in a noisy environment.
• Advantage: Significantly reduces lag compared to SMA/EMA, allowing for earlier entries and faster exits during trend reversals.
• Logic: Long Entry when Fast_Kalman > Slow_Kalman.

2. Machine Learning Risk Management (Meta-Labeling)

• Implements Marcos Lopez de Prado’s Meta-Labeling technique.
• A Random Forest Classifier is trained to predict whether a Kalman signal will result in a profit.
• Features Used: Kalman Spread, Volatility, RSI, and Volume Trend.
• Dynamic Leverage:
  • High Confidence (>55%): 2.0x Leverage
  • Neutral (>50%): 1.0x Leverage
  • Low Confidence: 0.5x Leverage (Defensive Mode)

3. Strict Backtesting Standards

• Look-Ahead Bias Free: All ML predictions are shifted (t+1) to ensure signals are available before the candle opens.
• Out-of-Sample Testing: The model is trained on 2019-2022 data and tested on unseen 2023-Present data.

📊 Performance Analysis

The integration of Machine Learning transformed the strategy from a high-risk system into a robust investment vehicle.

Metric	Fixed Leverage (1.5x)	ML Dynamic Leverage	Improvement
Annualized Return	56.41%	33.98%	Risk-Adjusted
Sharpe Ratio	1.02	1.03	+1%
Max Drawdown	-73.32%	-52.56%	+28% Safety
Avg Win/Loss	3.67	3.67	Maintained

Equity Curve Comparison

Figure 1: Comparison between Strategy (Red) and Buy & Hold (Gray).

Impact of ML Optimization

Figure 2: The ML Dynamic model (Purple) significantly reduces drawdown during market corrections compared to the Fixed Leverage model (Gray).

🚀 Getting Started

Prerequisites

• Python 3.8+
• Jupyter Notebook

Installation

1. Clone the repository:

   git clone [https://github.com/yourusername/Adaptive-Kalman-ML-Trading.git](https://github.com/yourusername/Adaptive-Kalman-ML-Trading.git)

2. Install dependencies:

   pip install -r requirements.txt

📜 License Distributed under the MIT License. See LICENSE for more information.