Linguistic knowledge about temporal data in Bayesian linear regression model to support forecasting of time series

Experts are able to predict sales based on approximate reasoning and subjective beliefs related to market trends in general but also to imprecise linguistic concepts about time series evolution. Linguistic concepts are linked with demand and supply, but their dependencies are difficult to be captured via traditional methods for crisp data analysis. There are data mining techniques that provide linguistic and easily interpretable knowledge about time series datasets and there is a wealth of mathematical models for forecasting. Nonetheless, the industry is still lacking tools that enable an intelligent combination of those two methodologies for predictive purposes. Within this paper we incorporate the imprecise linguistic knowledge in the forecasting process by means of linear regression. Bayesian inference is performed to estimate its parameters and generate posterior distributions. The approach is illustrated by experiments for real-life sales time series from the pharmaceutical market.