Knowledge-based adjustment on the forecasts of time series models

Abstract

The existing popular statistical forecasting models, like time series analysis, concentrate on the numerical analysis under strict assumptions such as static state and normality. However, forecasting environments in organizations are usually dynamic and almost always violate the assumptions. To accommodate irregular factors that are idiosyncratic for each organizations, the integration of statistical forecasting model and knowledge-based support for judgmental adjustment is attempted. In this approach, the qualitative irregular factors are organized in knowledge base. To fulfill this approach, qualitative factors are extracted out from the historical data, and the filtered data set is used for time series models. The proposed architecture, namely KAST (Knowledge-based Adjustment Support System for Time Series Forecasting Models), has the inference system to adjust the statistical forecast. In this thesis, we describe what qualitative factors are included in time series forecasting model for petroleum industry, and show how to learn the effect of qualitative factors, and how to compose normal forecast with judgmental knowledge. A prototype KAST is implemented using Common Lisp and Pascal in the microcomputer environment. Some experimental evaluation has shown the performance of KAST.