Skip to content

README.md

Latest commit

 

History

History
150 lines (109 loc) · 4.15 KB

README.md

File metadata and controls

150 lines (109 loc) · 4.15 KB
Raw

CryptoMarket Regime Classifier

Adaptive Market Intelligence for Crypto Strategies

Most trading strategies fail not because the logic is wrong, but because they are applied in the wrong market regime.

A breakout strategy thrives in trends and bleeds in chop.
Mean-reversion works in ranges and dies in momentum.

CryptoMarket Regime Classifier is a machine learning pipeline that detects and predicts crypto market regimes using multi-timeframe OHLCV data, technical indicators, and a two-stage ML approach (HMM → LSTM).

It is built as a foundational intelligence layer for:

  • strategy selection
  • position sizing
  • risk management

and is intended to power the regime-awareness module in Dazai.

High-Level Pipeline

OHLCV (5m, 15m) ↓ Feature Engineering (momentum, trend, volatility) ↓ PCA Reduction ↓ Hidden Markov Model (Regime Discovery) ↓ LSTM (Regime Prediction) ↓ Current Regime (+ future probabilistic output)

Key Ideas (Why this is different)

  • Regime-aware, not signal-based
    The model does not predict price — it predicts market conditions.

  • Unsupervised → Supervised learning
    HMM discovers latent regimes first.
    LSTM then learns temporal structure to predict them.

  • Multi-timeframe context
    Combines short-term and slightly higher-timeframe behavior (5m, 15m).

  • Designed for integration
    Models and scalers are exported for downstream systems (bots, dashboards, APIs).

Key Features

  • Multi-timeframe OHLCV data (5m, 15m) from Binance
  • Technical indicators covering:
    • momentum
    • volatility
    • trend
  • Hidden Markov Models (HMM) for unsupervised regime discovery
  • LSTM trained on HMM-labeled sequences
  • 6 discovered regimes, including:
    • Strong Trend
    • Weak Trend
    • Range
    • Choppy High-Volatility
    • Volatility Spike
    • Squeeze
  • Evaluation metrics:
    • Precision / Recall / F1
    • Confusion Matrix

Project Structure

├── dashboard/ # Visualizations, regime plots
├── models/ # Trained models & scalers
├── src/ # Feature engineering + training scripts
├── main.py # End-to-end pipeline execution
├── requirements.txt # Dependencies
└── README.md

Workflow Details

1. Data Fetching

  • Periodically fetches OHLCV data from Binance
  • Currently optimized for 5m data, with support for higher TF context

2. Feature Engineering

  • Computes momentum, trend, and volatility indicators
  • Aligns and scales features for ML stability

3. Regime Discovery (HMM)

  • PCA-reduced feature space
  • 6-state HMM selected using lowest BIC
  • Produces regime labels without human bias

4. Regime Prediction (LSTM)

  • Sequence model trained on HMM labels
  • Captures temporal transitions between regimes
  • Hyperparameters tuned using Keras Tuner
  • Planned upgrade: probabilistic regime distributions

5. Model Export & Usage

  • Trained LSTM + scalers saved to /models
  • Designed for reuse in live systems

Results (High-Level)

  • Strong separation between trend vs non-trend regimes
  • Transitional regimes (range ↔ weak trend, spike ↔ chop) are naturally harder — and informative
  • Confusion matrix reflects realistic regime overlap instead of artificial sharp boundaries

Installation

git clone https://github.com/akash-kumar5/CryptoMarket_Regime_Classifier.git
cd CryptoMarket_Regime_Classifier
pip install -r requirements.txt

Usage

Run the full pipeline: streamlit run dashboard/app.py

Models & scalers will be saved in /models for reuse.

Notes

  • Data range: ~2 years (to prioritize recent regime behavior and avoid stale market patterns).

  • Designed as a research + foundational tool for live trading systems.

  • Future versions will connect directly into Dazai as a core regime intelligence component.

-----

Disclaimer

This project is for research and educational purposes only. It does not constitute financial advice.