Judgmental adjustments to statistically generated forecasts have
become a standard practice in demand forecasting, especially at a
stock keeping units level. However, due to the subjective nature of
judgmental interventions this approach cannot guarantee optimal use
of available information and can lead to substantial cognitive biases.
It is therefore important to monitor the accuracy of adjustments and
estimate persistent systematic errors in order to correct final forecast.
This paper presents an appropriate methodology for such analysis
and focuses on specific features of source data including time series
heterogeneity, skewed distributions of errors, and generally nonlinear
patterns of biases. Enhanced modelling and evaluation techniques
are suggested to overcome some imperfections of well-known standard
methods in the given context.
Empirical analysis showed that a considerable proportion of final
forecast error is formed by a systematic component which can be pre-
dicted. Proposed bias correction procedures allowed to substantially
improve the accuracy of final forecasts. In particular, one-factor mod-
els of the relationship between forecast error and adjustment were
found to be a simple, robust and efficient tool for the given purpose.