0xroba/piggyback

Token Price Growth Prediction Model

Model Description

This model predicts the maximum price growth potential of crypto tokens over a 15-minute window. It uses XGBoost (converted to ONNX format) trained on historical token metrics data to predict potential short-term price movements.

Use Cases

• Price movement anticipation for trading strategies
• Risk assessment for token investments
• Liquidity analysis in token markets

Model Details

• Model Type: XGBoost (converted to ONNX for deployment)
• Training Algorithm: Gradient boosting with Huber loss (reg:pseudohubererror)
• Target: Maximum price growth percentage within a 15-minute window
• Input Features: Various token metrics including price, volume, transaction counts, and velocity indicators
• Preprocessing: StandardScaler normalization and log-transformation of target values

Limitations

• The model predicts potential maximum growth, not guaranteed outcomes
• Performance depends on market conditions similar to those in the training data
• Not suited for long-term price predictions
• Does not account for external market events or news

Usage Example

import numpy as np
import onnxruntime as ort
import json
import pandas as pd

# Load model and scaler
model = ort.InferenceSession("models/xgb_model.onnx")
with open("models/scaler_stats.json", "r") as f:
    scaler_data = json.load(f)

# Function to preprocess input data
def preprocess(data, scaler_data):
    # Convert to numpy array
    data_np = np.array(data, dtype=np.float32)

    # Apply scaling using the saved scaler parameters
    scaled_data = (data_np - scaler_data["mean"]) / scaler_data["scale"]

    return scaled_data.astype(np.float32)

# Function to make predictions
def predict(features, model, scaler_data):
    # Preprocess
    processed_features = preprocess(features, scaler_data)

    # Get input name
    input_name = model.get_inputs()[0].name

    # Run inference
    pred_log = model.run(None, {input_name: processed_features.reshape(1, -1)})[0]

    # Convert log prediction back to percentage
    pred_pct = np.expm1(pred_log) * 100.0

    return pred_pct[0][0]  # Return as scalar

# Example usage
features = [...]  # Array of features matching the model's expected input
prediction = predict(features, model, scaler_data)
print(f"Predicted max price growth: {prediction:.2f}%")