ARFIMA Models for Preserving Long Memory in Financial Series

Summary

The article introduces ARFIMA models, which extend ARIMA by allowing the integration parameter to be fractional. This lets the model represent persistent dependence, or long memory, that may be diminished when prices are converted to returns through ordinary differencing. It explains the fractional differencing idea and describes estimating ARFIMA models, including model selection with the Bayesian Information Criterion and using R tools to obtain residuals.

For Python workflows, the article describes passing data to an R script and returning the estimated residuals. It also presents a fractional differencing procedure and illustrates it with Apple prices: the selected fractional parameter is reported as 0.3145, the residuals’ ADF p-value as 2.6%, and their correlation with prices as 0.84. These are examples from the article, not evidence of profitable trading. It suggests using the residuals as features in machine-learning models and notes that ARFIMA use in retail trading is limited. Model fit, stationarity choices, data windows, and the possibility of changing market behavior constrain how these results should be interpreted.

Key ideas

• ARFIMA generalizes ARIMA by allowing fractional integration values between negative one and one.
• Fractional differencing aims to model persistent dependence while retaining more price-series memory than ordinary differencing.
• ARFIMA parameters can be estimated and candidate models compared using information criteria such as BIC.
• The article demonstrates an R estimation workflow that can be connected to Python to produce model residuals.
• Residuals may serve as machine-learning features, but the examples do not demonstrate trading profitability.

Tags

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