Forecasting Production Volume with Multivariate Time-series models : Comparative evaluation of time-series models

Abstract: One challenge for people working with data science at large industrial companies is that they have to deal with data sets that have a short history or a small dataset size. Unfortunately that causes many machine learning algorithms to lose accuracy and thus not produce any meaningful results. Modern machine learning algorithms are well suited to handling data with a long time series or large data sets with many features. Data scientists working with these large data sets have access to numerous algorithms and tools that help them to work with these data sets more efficiently. Through the years these tools and algorithms have been continuously improving and are meticulously tested by industrial practitioners, increasing their experience with them. When it comes to short data sets, several new challenges occur (such as overfitting and outliers), making these data sets trickier to handle and requiring different approaches. This is compounded by the fact that there are not so many algorithms able to generate accurate forecasts from small datasets and not so many practitioners who have experience with them. To overcome these problems, it is necessary to predict the target series by building a model that is based on short lengths time series. The target series being, in this thesis, a forecast of production volume at the company in question. The thesis also aims to increase the accuracy of this forecast over today’s methods, which requires an examination of multiple different techniques.