Auto-ARIMA Forecasting with ADF Differencing and Rolling Retraining

Summary

This module outlines an out-of-sample forecasting workflow built around Auto-ARIMA. It first applies an Augmented Dickey-Fuller test at a five percent significance level, repeatedly differencing the training series until the test indicates stationarity or a configured maximum differencing order is reached. It then asks pmdarima to select autoregressive and moving-average orders using an AIC-based search, with configurable starting and maximum orders.

For prediction, the class iterates through a supplied holdout series and can periodically refit the selected model. Refit data may include all prior training observations or a trailing training window, along with earlier holdout observations. The code avoids including the current observation in its refit data, which is intended to prevent look-ahead bias. It provides implementation mechanics rather than empirical results: there is no trading strategy, forecast evaluation, benchmark, or evidence of profitability. Forecast quality and suitability depend on the series, search settings, retraining schedule, and validation design.

Key ideas

• The workflow uses an ADF test to choose a differencing order before fitting Auto-ARIMA.
• The AR and MA search ranges are configurable, and the candidate model is selected using AIC.
• Out-of-sample forecasts can use periodic refitting with expanding or trailing training data.
• Refitting excludes the current holdout observation to reduce look-ahead bias.
• The module supplies no forecast-performance results or evidence of trading effectiveness.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.