AutoSNARIMAX

Using the River and Optuna packages to provide an automatically optimized ARIMA-derived model with the possibility of online updates.

Overview

AutoSNARIMAX is a Python class for automatic hyperparameter optimization of SNARIMAX models using Optuna and the River library. It supports:

• Automatic hyperparameter tuning for SNARIMAX models.
• Incorporation of additional features, including holidays.
• Online updates: the model can be updated incrementally with new observations.
• Forecasting with uncertainty intervals.

Installation

pip install autosarimax

Usage

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from autosarimax import AutoSNARIMAX, add_holiday_feature

# 1️⃣ Create example DataFrame
dates = pd.date_range(start='2025-01-01', periods=30, freq='D')
np.random.seed(42)
y = np.random.randint(50, 150, size=len(dates))
X = pd.DataFrame({'date': dates, 'feature1': np.random.randn(len(dates))})

# 2️⃣ Add holiday feature
X = add_holiday_feature(X, country='BR', date_col='date')
X = X.drop(columns=['date'])

# 3️⃣ Initialize and fit model
model = AutoSNARIMAX(n_trials=5, horizon=1)
model.fit(X, y)

# 4️⃣ Make incremental predictions
pred_df = model.predict(X, force=False)

# 5️⃣ Plot real vs predicted values
plt.figure(figsize=(12, 6))
plt.plot(dates, y, label='Real', marker='o')
plt.plot(dates, pred_df['pred'], label='Predicted', marker='x')
plt.fill_between(dates, pred_df['lower'], pred_df['upper'], color='gray', alpha=0.3, label='Uncertainty')
plt.xlabel('Date')
plt.ylabel('Value')
plt.title('AutoSNARIMAX Forecast vs Real')
plt.legend()
plt.grid(True)
plt.show()

Features

• add_holiday_feature(df, country, date_col): Adds a binary column is_holiday indicating whether a date is a holiday.

• fit(X, y): Optimizes hyperparameters and trains the SNARIMAX model.

• update(X, y): Updates the model incrementally with new observations.

• predict(X, horizon, force): Makes forecasts.

  • force=True: forecasts the full horizon without updating the model.
  • force=False: forecasts incrementally and updates the model using the predicted values.

License

MIT License

GitHub