Judgmental adjustment in time series forecasting

Abstract

Time series models are a highly useful forecasting method, but are deficient in the sense that they merely extrapolate past patterns in the data without taking into account the expected irregular future events. To overcome this limitation, forecasting experts in practice judgmentally adjust the statistical forecasts. Typical judgmental factors may be treated as outliers in statistical analysis. To automatic the judgmental adjustment process, neural network models are developed in this study. To collect the data for judgmental events, judgmental effects are filtered out of raw data. The main trend is captured by a neural network model using the filtered data, while judgmental effects are modeled by another neural network. Then the judgmental effects are additively adjusted. Performance of this architecture is tested in comparison with five other architectures: 1) A single neural network model using the filtered data, 2) A single neural network model using the raw data, 3) A single neural network model using the filtered data and judgmental factors as the inputs of the model, 4) Major trend in ARIMA model with a neural network based additive adjustment, 5) A single ARIMA model using the raw data According to the experiments, the architecture of neural network based additive judgmental adjustment significantly improves the forecasting performance. To support the implementation of the architecture, a prototype UNIK-FCST/NN is implemented on refinery case.