Using Fractional Differencing as a Model Input

Summary

The discussion concerns fractional differencing as a way to balance stationarity with retention of information in time series used for supervised prediction. The question asks whether predictions from fractionally differenced data should be reintegrated into an absolute price target, and how to reverse the transformation. It gives no detailed inversion procedure.

The answer describes fractionally differenced series as model inputs, often drawn from data other than the underlying price, that can help generate a trading signal or return forecast. If the model predicts a return and a price-level target is needed, the answer suggests calculating the corresponding price from the prior price. It questions the motivation for reconstructing a level, but does not explain a general reintegration method or establish that reintegration is standard practice. The discussion is brief and leaves the choice of prediction target and transformation details to the researcher.

Key ideas

• Fractional differencing can be used to prepare time series as inputs to a supervised model.
• The transformation aims to balance stationarity with retaining information.
• Fractionally differenced features may feed a model that predicts a signal or return.
• A predicted return can be translated into a price using the preceding price.
• The discussion does not provide a general inverse procedure or evidence about standard practice.

Tags

# Reintegrating Fractionally Differentiated Time Series Prediction


# Reintegrating Fractionally Differentiated Time Series Prediction












I am working on a supervised learning approach to Time Series Regression, and am currently investigating fractionall differentiation (optimizing the stationarity/information tradeoff) discussed chapter 5 in Dr. Lopez de Parado's Advances in Financial Machine learning.

When I previously worked with differentiated time-series predictions of order 1, reintegrating the prediction to get an absolute price target was chosen as technique.

Now the book, nor anywhere else on the internet, makes mention of reintegrating a fractionally differentiated time series prediction in the context of Finance.

- Am I missing something?

- Is reintegration of predictions (returns to absolute prices) a commonly accepted approach?

- How would one go about undoing the operation applied in https://github.com/hudson-and-thames/mlfinlab/blob/master/mlfinlab/features/fracdiff.py

Thanks!

## Answer by Quantoisseur (score 1)

https://quant.stackexchange.com/a/57115

Fractionally differentiated features (often time series other than the underlying's price) are generally used as inputs into a model to then generate a trading signal/return prediction.

If you have a predicted return and want to convert it to the price level you could just do the actual return calculation from the previous price but I'm not sure what your motivation is.

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

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