Fractional Differentiation for Stationary Financial Features
Summary
The article explains why financial time series are often made stationary for statistical inference and supervised machine learning, then presents fractional differentiation as a way to reduce nonstationarity while retaining more of a price series’ memory than ordinary differencing. It links this approach to Hosking’s work on long-range dependence and to a chapter on financial machine learning.
Its example applies different fractional differencing amounts to e-mini S&P 500 futures log prices. The article reports that a low differencing amount reaches the stated augmented Dickey-Fuller critical threshold while preserving high correlation with the original series, illustrating the trade-off between stationarity and retained information. That is an example rather than broad evidence of predictive performance: the article does not show out-of-sample trading results or establish that retained correlation improves forecasts. It points readers to software and a notebook for implementation.
Key ideas
- Fractional differencing aims to make a time series stationary while retaining more of its dependence structure than integer differencing.
- Stationarity supports statistical inference and supervised learning by making data characteristics more stable over time.
- The example compares stationarity and correlation across differencing amounts for futures log prices.
- High correlation with the original series indicates retained information, but does not by itself demonstrate predictive or trading value.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.