Deep Learning for Finance

Chapter 9. Deep Learning for Time Series Prediction II

This chapter presents a few techniques and methods to complement the forecasting task of machine and deep learning algorithms. It is composed of different topics that each discuss a way to improve and optimize the process. At this point, you should have a sound understanding of the basics of machine and deep learning models, and you know how to code a basic algorithm that predicts the returns of a financial time series (or any stationary time series). This chapter bridges the gap between the basic knowledge and the advanced knowledge required to elevate the algorithms to a functional level.

Fractional Differentiation

In his book Advances in Financial Machine Learning, Marcos López de Prado describes a technique to transform nonstationary data into stationary data. This is referred to as fractional differentiation.

Fractional differentiation is a mathematical technique used to transform a time series into a stationary series while preserving some of its memory. It extends the concept of differencing (or taking the returns), which is commonly used to remove trends and make time series stationary.

In traditional differencing, the data sequence is differenced by a whole number, typically 1, which involves subtracting the previous value from the current value. This helps eliminate trends and makes the series stationary. However, in some cases, the series may exhibit long-term dependencies or memory effects that are not effectively ...