A machine learning pipeline for predicting Ethereum priority fees based on historical transaction data.
- Data Collection: Reads transaction data from JSON files containing mempool and historical fee statistics.
- Feature Extraction: Extracts relevant gas-related features like mempool fee distribution and historical gas spikes.
- Training the Model: Uses a
RandomForestRegressorwrapped inMultiOutputRegressorto predict different priority fee quantiles. - Inference: Given real-time network conditions, the model estimates optimal priority fees for different urgency levels.
Requires Python 3.7+
pip install -r requirements.txt
Run the training script to process data and train the model:
python gasflow_ml.py
This will output a trained model saved as gas_model.pkl.
To use the trained model for predictions:
import joblib
from gasflow_ml import predict_priority_fees
# Load trained model
model_artifacts = joblib.load("gas_model.pkl")
model = model_artifacts['model']
scaler = model_artifacts['scaler']
# Example input data
network_context = {
"mempool": {
"count": 112,
"p10": 0.0001,
"p30": 0.1,
"p50": 0.5,
"p70": 1,
"p90": 2
},
"history": {
"gas_ratio_5": 0.526,
"gas_spikes_25": 4,
"fee_ewma_10": 0.845,
"fee_ewma_25": 0.784
}
}
# Predict fees
predictions = predict_priority_fees(network_context, model, scaler)
print(predictions)The training data should be stored as JSON files in a directory (./data). Each file should follow this structure:
{
"network": {
"mempool": {
"count": 100,
"p10": 0.5,
"p30": 1.2,
"p50": 2.5,
"p70": 3.8,
"p90": 5.1
},
"history": {
"gas_ratio_5": 0.52,
"gas_spikes_25": 3,
"fee_ewma_10": 0.82,
"fee_ewma_25": 0.76
}
},
"txs": [
{"max_priority_fee": 2000000000},
{"max_priority_fee": 3000000000}
]
}Note: This is an initial version and may have areas for improvement. Contributions and feedback are welcome.
MIT License