Adapting image processing and clustering methods to productive efficiency analysis and benchmarking: A cross disciplinary approach

This dissertation explores the interdisciplinary applications of computational methods in quantitative economics. Particularly, this thesis focuses on problems in productive efficiency analysis and benchmarking that are hardly approachable or solvable using conventional methods. In productive efficiency analysis, null or zero values are often produced due to the wrong skewness or low kurtosis of the inefficiency distribution as against the distributional assumption on the inefficiency term. This thesis uses the deconvolution technique, which is traditionally used in image processing for noise removal, to develop a fully non-parametric method for efficiency estimation. Publication I and Publication II are devoted to this topic, with focus being laid on the cross-sectional case and panel case, respectively. Through Monte-Carlo simulations and empirical applications to Finnish electricity distribution network data and Finnish banking data, the results show that the Richardson-Lucy blind deconvolution method is insensitive to the distributional assumptions, robust to the data noise levels and heteroscedasticity on efficiency estimation. In benchmarking, which could be the next step of productive efficiency analysis, the ‘best practice’ target may not perform under the same operational environment with the DMU under study. This would render the benchmarks impractical to follow and, consequently, adversely affects the managers to make the correct decisions on performance improvement of a DMU. This dissertation proposes a clustering-based benchmarking framework in Publication III. In this framework, we group the DMUs into segments using clustering methods based on certain metrics under interest, and estimate the efficiencies afterwards to pin down the segment-specific benchmark for DMUs within each cluster. The empirical study on Finnish electricity distribution network reveals that the proposed framework novels not only in its efficient consideration on the differences of the operational environment among DMUs, but also its extreme flexibility, e.g., the clustering and efficiency estimation techniques are user-decided according to their specific needs and preference. We conducted a comparison analysis on the different combinations of the clustering and efficiency estimation techniques using computational simulations and empirical applications to Finnish electricity distribution network data. Based on the results, Publication IV proposes the combined use of ‘the normal mixture model based clustering’ and ‘the stochastic semi-nonparametric envelopment of data (StoNED)’. This is because that such a combination could produce more accurate 1 DOCTORAL DISSERTATIONS Aalto University publication series DOCTORAL DISSERTATIONS 134/2016 Adapting image processing and clustering methods to productive efficiency analysis and benchmarking: A cross disciplinary approach