In today's enterprises, planning and decision making is mostly based on forecasts of business figures such as cost of goods or expected demand levels. Almost all potentially important variables are measured and predicted on a regular basis considering collected historical data and further facts or expectations. As of today, business figures are mostly predicted bottom-up, starting for example with expected costs of goods for particular products, and then aggregating these values to product groups, etc.
However, there are dependencies with macro-economical factors such as commodity price developments, economic indices such as GDP or inflation rates, various sector-specific indicators, etc. Such macro-economic data are important variables for predicting business figures as there are usually causal relationships. For instance, cost of good curves depend on the price developments of raw materials. Such macro-economic data is also measured and predicted at regular frequencies. Theoretically, the historical data could be used for prediction, planning and optimization. Hence, an alternative approach to predict business figures are multivariate prediction models, with macro-data as driver. Unfortunately, the sheer data volume prohibits the identification of dynamics and interrelations of key variables and mathematical programs to not scale when parameterized with such high-dimensional, noisy, interrelated data. Dynamics of these key macro-economicial variables are often masked by other variables and are interrelated in various ways. For example, some variables move rather synchronously, some move together only in the long-run and some lead others in time with different time-varying lags.
This project is aimed at modelling these dynamic interrelationships in order to increase forecasting accuracy. We explore ways to formulate robust multivariate lagged prediction models. We derive a small set of predictors (or features) in terms of orthogonal macro-economical indicators and artificial factors that bundle the developments of related indicators to reduce uncertainty and noise. Regression models are developed considering the predictability of predictors themselves when selecting and combining features.