Advances in Financial Machine Learning

CHAPTER 4 Sample Weights

4.1 Motivation

Chapter 3 presented several new methods for labeling financial observations. We introduced two novel concepts, the triple-barrier method and meta-labeling, and explained how they are useful in financial applications, including quantamental investment strategies. In this chapter you will learn how to use sample weights to address another problem ubiquitous in financial applications, namely that observations are not generated by independent and identically distributed (IID) processes. Most of the ML literature is based on the IID assumption, and one reason many ML applications fail in finance is because those assumptions are unrealistic in the case of financial time series.

4.2 Overlapping Outcomes

In Chapter 3 we assigned a label y_i to an observed feature X_i, where y_i was a function of price bars that occurred over an interval [t_{i, 0}, t_{i, 1}]. When t_{i, 1} > t_{j, 0} and i < j, then y_i and y_j will both depend on a common return , that is, the return over the interval [t_{j, 0}, min{t_{i, 1}, t_{j, 1}}]. The implication is that the series of labels, {y_i}_{i = 1, …, I}, are not IID whenever there is an overlap between any two consecutive outcomes, ∃i|t_{i, 1} > t_{i + 1, 0}..

Suppose that we circumvent this problem by restricting the bet horizon to t_{i, 1} ≤ t_{i + 1, 0}. In this case there is no overlap, because every feature outcome is determined before or at the ...