The Book of Alternative Data

CHAPTER 7Missing Data: Background

7.1. INTRODUCTION

As we discussed in Section 3.3.2, dealing with missing data – a ubiquitous problem – is one of the crucial steps in making data useful at all. In this chapter we will describe the problem of missing data imputation in more general terms. We will present a specific case study that focuses on filling gaps in multivariate financial time series in the next chapter.

Providing a general recipe for tackling missing data is not possible, given that the problem arises in many different-in-nature practical applications. For example, filling gaps in financial time series can be quite different from filling gaps in satellite images or text. Nevertheless, some techniques can be widely reused over different domains, as we will show in this chapter and the next. Techniques to fill missing data are applicable regardless of whether or not a dataset is alternative, so in what follows we will not make such distinction. We only remark that, in general, we expect to have more missing data and data quality problems in the alternative data space. This is due to the increased variety, velocity, and variability of alternative data compared to more standardized traditional datasets.

Treating missing data is something that must be performed before any further analysis is attempted. A predictive model (e.g. an investment strategy) can then be calibrated on the treated dataset as a second step. We must be careful, though, to understand whether the missing ...