A comparative study of machine learning approaches in data envelopment analysis: bank data
Quality and Quantity, 2026 (Scopus)
- Yayın Türü: Makale / Tam Makale
- Basım Tarihi: 2026
- Doi Numarası: 10.1007/s11135-026-02968-8
- Dergi Adı: Quality and Quantity
- Derginin Tarandığı İndeksler: Scopus, IBZ Online, ABI/INFORM, Index Islamicus, Political Science Complete, Psycinfo, Political Science Abstract (IPSA), Social Science Premium Collection (ProQuest), Health Research Premium Collection (ProQuest), Sociology Database (ProQuest), Sociology Source Ultimate (EBSCO)
- Anahtar Kelimeler: Adaptive LASSO, Data envelopment analysis, LASSO, Principal component analysis, Variable selection, Weighted LASSO
- Bilecik Şeyh Edebali Üniversitesi Adresli: Evet
Özet
Data Envelopment Analysis (DEA) is a nonparametric method widely used to evaluate the relative performance of institutions and organizations operating under multiple input–output structures. However, when the empirical setting involves a relatively large number of candidate variables and correlated financial indicators, the reliability, discriminatory power, and comparability of DEA efficiency estimates may be affected. This study provides a comparative analysis of machine-learning-based variable selection approaches within a DEA framework for evaluating bank financial performance. Using a dataset composed of financial ratios obtained from bank financial statements, the study examines the contribution of variable selection before DEA implementation. In this context, a two-stage framework is adopted. In the first stage, Least Absolute Shrinkage and Selection Operator (LASSO)-based methods are used as screening tools to obtain reduced subsets of candidate input variables. In the second stage, DEA efficiency scores are estimated using the selected variables. The proposed approach extends the standard LASSO framework by assigning covariate-specific penalty weights derived from Principal Component Analysis (PCA) loadings, resulting in a PCA-informed Weighted LASSO (WLASSO) structure. These weights guide the variable selection stage before DEA estimation and allow the analysis to examine how efficiency scores respond to alternative input specifications. For comparison, standard LASSO and Adaptive LASSO (ALASSO) are implemented within the same two-stage DEA framework. The empirical findings indicate that different variable selection strategies may lead to variations in DEA efficiency scores, while the overall ranking structure across decision-making units is often preserved. The results therefore provide comparative empirical evidence on the role of machine-learning-based input selection in DEA applications. These findings should be interpreted as application-oriented comparative evidence rather than as formal superiority claims.