crypto market analysis tool: hmm market regime detection

overview

a trading model for newly listed crypto assets using hidden markov models (hmm) with baum-welch algorithm. detects market regimes early and trades based on state transitions.

quick start

git clone [email protected]:0xDVC/vela-crypto.git
cd vela-crypto
python3 -m poetry shell # need poetry? python3 -m pip install poetry
python3 -m poetry install
streamlit run app.py

dependencies

python = "^3.11"
pandas = "2.1.4"  # very important
scikit-learn = "^1.6.0"
python-binance = "^1.0.27"
hmmlearn = "^0.3.3"
streamlit = "^1.41.1"
tqdm = "^4.67.1"
joblib = "^1.4.2"
matplotlib = "^3.10.0"
seaborn = "^0.13.2"
requests = "^2.32.3"
ipykernel = "^6.29.5"
plotly = "^5.24.1"
numpy = "1.26.4"    # very important, otherwise you'll run into convergence issues with the model

how it works

data pipeline

• fetches 15m candles from binance api (rate-limited)
• 97 days of historical data (90 training, 7 testing)
• smart caching to avoid api abuse
• market cap data scraped from binance web (api limitation)

feature engineering

@dataclass
class FeatureConfig():
    """feature configuration for short-term HMM on small caps"""
    FEATURE_COLUMNS: list = ["returns", "rsi", "bb_position", "volatility", "momentum"]
    RSI_PERIOD: int = 10
    BB_PERIOD: int = 15
    BB_STD: float = 1.8
    MOMENTUM_SPAN: int = 3

hmm implementation

• 3 states (bull/bear/neutral) based on market regimes
• state classification using:
  • returns (2x weight)
  • momentum (1.5x weight)
  • rsi (normalized)
  • volatility (bear detection)

trading strategy

@dataclass
class TradingConfig:
    initial_balance: float = 10000.0
    trading_fee: float = 0.001
    position_size: float = 0.95
    stop_loss_pct: float = 0.05  # 5% stop loss
    trailing_stop_pct: float = 0.07  # 7% trailing stop

how to use

1. run the streamlit app:

streamlit run app.py

1. what you'll get:

• feature correlation matrix
• market state classifications
• trading signals
• performance metrics with charts

current performance

• sharpe: -0.20 (work in progress)
• max drawdown: -6.66%
• daily vol: 0.20

model catches regime changes but needs work on neutral states and stop-loss implementation

known issues

1. model behavior:

   • regime changes detected with lag
   • parameter sensitivity (work in progress)
   • state classification needs work

2. trading logic:

   • basic position sizing

3. data limitations:

   • no market cap in binance api
   • 15m minimum interval
   • rate limiting challenges

roadmap

• model improvements:

  • cross-validation for state selection
  • dynamic state classification
  • parameter optimization

• risk management:

  • implement stop-loss
  • smart position sizing

project structure

.
├── README.md
├── app.py
├── core
│   ├── __init__.py
│   ├── collect.py  # data collection
│   ├── features.py # feature engineering
│   ├── model.py    # hmm implementation
│   ├── report.py   # metrics and charts
│   └── trading.py  # trading logic
├── data
├── image.png
├── poetry.lock
├── pyproject.toml
├── requirements.txt
└── vela_crypto_model.ipynb

why these choices?

data decisions

• 15m intervals: closest to 10m target while working within binance api limitations
• 90/7 day split: enough data to catch market regimes while keeping test set relevant
• small cap focus: more predictable regime changes in lower liquidity markets

feature selection

@dataclass
class FeatureConfig():
    """feature configuration for short-term HMM on small caps"""
    FEATURE_COLUMNS: list = ["returns", "rsi", "bb_position", "volatility", "momentum"]
    RSI_PERIOD: int = 10  # shorter for small caps
    BB_PERIOD: int = 15   # adapted for volatility
    BB_STD: float = 1.8   # tighter bands for small caps
    MOMENTUM_SPAN: int = 3 # quick momentum detection

why these features?

• returns: primary state indicator
• volatility: crucial for bear market detection
• rsi: proven momentum indicator, especially in crypto
• bb_position: catch mean reversion opportunities
• momentum: early trend detection

hmm architecture

• 3 states: empirically better than 2 (too simple) or 4+ (overfitting)
• feature weights:

  # Calculate combined score with more weight on returns and momentum
  state_scores = (
      state_means[:, returns_idx] * 2.0 +  # Double weight on returns
      state_means[:, momentum_idx] * 1.5 +  # 1.5x weight on momentum
      (state_means[:, rsi_idx] - 50) / 50   # Normalized RSI contribution
  )

• baum-welch: optimal for unsupervised regime detection
• multiple random starts: avoid local optima

trading parameters

@dataclass
class TradingConfig:
    initial_balance: float = 10000.0
    trading_fee: float = 0.001    # binance standard fee
    position_size: float = 0.65   # high conviction trades
    stop_loss_pct: float = 0.05   # tight stops for small caps
    trailing_stop_pct: float = 0.07  # balance between profits/whipsaws

optimization choices

• aic/bic metrics: prevent overfitting in state selection
• standardscaler: handle different feature scales
• cross-validation: robust parameter selection

future work

• sentiment analysis for better regime detection
• on-chain data for better risk management
• portfolio management for better performance
• strategy self-adaptation for dynamic market conditions (highly volatile markets like the crypto market and especially low-cap assets require more adaptive strategies, and more so, self-adaptive strategies)