UNIK-FCST: Knowledge-assisted adjustment of statistical forecasts

Abstract

Time series models have served as a highly useful forecasting method, but are deficient in that they merely extrapolate past patterns in data without taking into account expected future events and other qualitative factors. To overcome this limitation, forecasting experts in practice judgmentally adjust the statistical forecasts. To partially replace the role of forecasting expert's judgement, we have developed an expert system UNIK-FCST (UNIfied Knowledge-ForeCaST). The UNIK-FCST learns from historical judgmental adjustments through generalization and analogy; reasons based on similar cases; and composes and decomposes the impacts of simultaneous judgmental events non-monotonically. Here, the UNIK-FCST is applied to the demand forecasting of oil products, for which five types of judgmental factors exist.