A procedure for modular forecasting at scale with constraints for business time series

Nowadays artificial intelligence algorithms are capable to achieve impressive results with a reduced amount of physical and time resources. They cover many different topics with a discrete success, but one of the most challenging subject to model is the prediction of future trends in complex and mutable environments, such as market sales. From a deterministic point of view, the knowledge of the exact state and the rules of a system in a certain period intrinsically brings the faculty to forecast any future state. This perspective yields an exact prediction, but it lays its foundation on the assumption that its possible to model every aspect of the system, a premise that is usually satisfied only in simple cases. The difficulty of predicting time-series is amenable to many factors, one of the most important is the drastically instable and mutable domain subjected to the competitiveness of its constituents, vendors and buyers, as described by the principles of the game theory; in such system, the rules are constantly changing, hardening the predictions of future states. Furthermore, financial studies produced a variety of economic models which help to understand the market behaviour. By applying specific constraints to the prediction, it is possible to exploit these models to reach better and more explainable results and relate their components to the relative sources. The aim of this work is to propose a procedure capable of inject domain constraints in the prediction in a declarative fashion, addressing different economic models. This procedure helps the analysts to better express their domain expertise while keeping a completely explainable approach to describe their outcomes