Use this URL to cite or link to this record in EThOS:
Title: Integration of judgmental and statistical approaches for demand forecasting : models and methods
Author: Davydenko, Andrey
Awarding Body: Lancaster University
Current Institution: Lancaster University
Date of Award: 2012
Availability of Full Text:
Access from EThOS:
The need for the composite use of judgmental and statistical approaches in forecasting is caused by the fact that each of these approaches itself cannot ensure the desired quality of forecasts. The topic of integrating forecasting methods has been long addressed in literature. However, due to the specific nature of demand data existing solutions in this area often cannot be efficiently applied in demand forecasting. The aim of the research is to develop efficient models and methods which would better correspond to realistic problem definitions in the context of demand forecasting. The first question that requires resolution is measuring the quality of demand forecasts. A critical analysis of existing error measures has shown that they are not always suitable for demand data due to their statistical properties. To provide a more robust and interpretable indication of forecasting performance the use of an enhanced statistic is proposed. One area of the research relates to the correction of judgmental forecasts. Since judgmental forecasts are inherently affected by cognitive biases, special means are required for producing an adequate probabilistic representation of future demand. Alongside the analysis of independent judgmental forecasts the research has examined the statistical properties of judgmental adjustments to statistically generated forecasts. Empirical analysis with real-world datasets shows that classical assumptions do not hold true and therefore standard procedures and tests cannot be correctly applied. Therefore more flexible methods have been designed to ensure more efficient and reliable analysis of judgmental forecasts. The results from the proposed techniques make it possible i) to reveal and eliminate systematic errors, and ii) to adequately evaluate the uncertainty associated with judgmental forecasts. Another area of research has focused on using prior judgmental information as an input " into statistical modelling, thereby obtaining consistent forecasts using both expert knowledge and latest observations. The proposed approach here is based on constructing a model with a combined dataset where available actual values and expert forecasts are described by means of corresponding regression equations. This allows the use of judgmental information in order to derive the prior characteristics of a data generation process. Model estimation is done using Bayesian inference and iterative algorithms, which make it possible to use sufficiently flexible model specifications. Analysis based on real data has shown that the approach and the proposed models can be easil?, and efficiently applied in practice. In summary, the contribution of the thesis is as follows. i) A number of previously unstudied effects are identified that can potentially lead to misinterpretation of measurement results obtained with the use of various well-known accuracy measures including MAPE, MdAPE, GMRAE, and MASE. ii) A new general error measure with improved statistical properties is proposed to overcome some imperfections of existing error measures. iii) New models and methods for efficient processing of point independent judgmental forecasts and judgmental adjustments to statistical forecasts are proposed based on Bayesian numerical analysis, iv) A new approach is proposed for the efficient incorporation of judgment into a statistical model of process dynamic.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available