Measuring the impact of enterprise risk management on performance, value, and risk indicators of Borsa Istanbul XBANK companies with data mining prediction models

The study examines whether and how enterprise risk management (ERM) adoption—captured by the firm’s ERM implementation level and the sophistication of ERM’s organizational structure—affects firm performance, firm value, and risk in the banking sector. Motivated by sustainability concerns and the shift from silo-based risk management to holistic ERM, the paper focuses on Turkish banks listed in Borsa Istanbul’s XBANK index (2019–2022). It posits that higher ERM maturity and more sophisticated ERM governance should improve accounting and market performance and reduce insolvency risk. The paper also evaluates predictive modeling approaches by comparing panel data regression models (pooled, fixed, random, with robust variants) against a data-mining alternative, partial least squares regression (PLSR). The research aims to provide empirical evidence for ERM’s effects in an emerging market banking context and to propose a hybrid analytics framework for predicting performance, value, and risk indicators.

Prior research links ERM to performance, value, and risk with mixed evidence. Gordon et al. (2009) found a positive association between ERM and performance using an ERM index; McShane et al. (2011) noted positive correlations of ERM ratings with firm value in insurance; Hoyt and Liebenberg (2011) reported favorable ERM–value relationships. Ellul and Yerramilli (2013) showed robust and autonomous risk management reduces banks’ tail risk. Baxter et al. (2013) associated ERM quality positively with performance and value but observed lower ERM quality in high-risk firms. Farrell and Gallagher (2015, 2019) linked ERM maturity to higher value, moderated by stakeholders and complexity. Other studies, e.g., Florio and Leoni (2017), found that advanced ERM systems improve profitability and market assessments; Callahan and Soileau (2017) associated higher ERM levels with superior operations. Conversely, Lundqvist (2015) questioned the ERM value proposition, and Otero González et al. (2020) found no significant ERM impact on performance or bankruptcy risk in Spanish non-financial firms, though appointing a CRO reduced volatility of financial distress measures. Malik et al. (2020) found ERM effectiveness, supported by risk committees, improves performance (Q ratio) in UK firms. Several works propose ERM measurement frameworks (e.g., Adam et al., 2023; CAMELS integration; COSO and ISO 31000 references) and highlight methodological gaps in measuring ERM. Data mining has been applied in risk contexts (bankruptcy prediction, process mining, SVMs), but PLSR has not previously been used in ERM studies, indicating a research opportunity this paper addresses.

Sample and data: Ten banks listed on BIST’s XBANK index were analyzed over 2019–2022 (40 bank-year observations). Annual data for performance/value indicators and firm characteristics were sourced from BIST and Turkey’s Public Disclosure Platform; monthly market data were used to compute Beta and volatility. The study period includes the COVID-19 pandemic. Variables: Dependent variables (DVs) are accounting performance (ROA, ROE), market value (Tobin’s Q), and risk indicators: systematic risk (Beta), stock return volatility (Vol), and standardized z-score (Sz_score) for insolvency risk. Independent variables (IVs) include: total assets (TA, log), number of branches (logBranch), number of employees (logEmp), leverage ratio (LR), current ratio (CR), equity-to-asset ratio (EAR), market-to-book (MtB), earnings per share (EpS), cost-to-income ratio (CtI), ERM implementation level (ERM; ordinal 1–4: basic, standard, modern, advanced), and ERM organizational structure sophistication (ORG; ordinal 1–3: low, moderate, high). Modeling framework: Panel data (PD) regression models—pooled regression (PRM), fixed effects (FEM), random effects (REM)—were estimated with specification tests guiding model choice. Tests included Hausman (FEM vs REM), Modified Wald (groupwise heteroscedasticity in FEM), Levene/Brown-Forsythe (heteroscedasticity in REM), Pesaran CD (cross-sectional dependence), and Bhargava Durbin–Watson and Baltagi–Wu LBI (autocorrelation). Robust regression models (RRM; iteratively reweighted least squares with bisquare weights) were used when heteroscedasticity or cross-sectional dependence warranted robustness. Stepwise reduction was applied to retain significant predictors for final predictive equations. PLSR alternative: Partial least squares regression (SIMPLS algorithm) was employed to address multicollinearity and high predictor dimensionality, constructing latent components that maximize covariance with the DV. Models were tuned by cross-validation and the number of components was chosen considering variance explained and mean squared prediction error. Model selection outcomes: Specification tests suggested FEM for ROA, ROE, Q; REM for Sz_score, Vol, Beta. Due to detected heteroscedasticity and/or cross-sectional dependence, RRM was used for ROA, ROE, Q, Vol, and Beta, while Sz_score used REM. Software: Stata 18 for PD regressions and tests; MATLAB R2022b for PLSR.

Influence analyses (panel data models): - ERM and performance/value/risk: - ROA: ERM level positively and significantly associated with ROA (high significance). - ROE: ORG sophistication positively associated with ROE (significant). - Q: ORG sophistication positively associated with Tobin’s Q (significant); ERM level itself not significant for Q. - Sz_score (insolvency risk): ORG sophistication positively associated with Sz_score (significant), implying lower insolvency risk with more sophisticated ERM structures. - Vol and Beta: No significant associations with ERM or ORG. - Significant controls (from PD models): - ROA: TA (negative), logEmp (negative), EpS (positive). - ROE: TA (positive), CR (positive), logBranch (negative), logEmp (negative), EAR (positive). - Q: logEmp (negative). - Sz_score: MtB (positive), CtI (negative). - Vol: TA (negative), CR (negative). - Beta: LR (positive), logEmp (negative), EAR (negative). Model choices per DV: RRM for ROA, ROE, Q, Vol, Beta; REM for Sz_score. Prediction performance (R2): - ROA: RRM 0.8540 vs PLSR10 0.8789 (PLSR better). - ROE: RRM 0.5240 vs PLSR10 0.7098 (PLSR better). - Q: RRM 0.0530 vs PLSR10 0.4951 (PLSR much better). - Sz_score: REM 0.3816 vs PLSR6 0.5822 (PLSR better). - Vol: RRM 0.2580 vs PLSR6 0.2766 (PLSR slightly better). - Beta: RRM 0.8990 vs PLSR7 0.3520 (RRM far better). Overall, ERM implementation improves accounting performance (ROA), while ERM organizational sophistication improves accounting performance (ROE), market value (Q), and reduces insolvency risk (higher Sz_score). Market-based risk measures (Vol, Beta) show no significant relation to ERM/ORG. PLSR yields superior predictive accuracy for most outcomes except Beta, where robust PD regression dominates.

The findings support the core hypothesis that adopting ERM—especially establishing a more sophisticated ERM organizational structure—enhances banks’ performance and value and reduces insolvency risk. The positive ERM–ROA association suggests that higher ERM maturity enhances decision quality and operational efficiency. ORG’s positive effects on ROE and Q imply that formalized governance (e.g., CEO/CRO involvement, risk committees, enterprise risk manager roles) translates into better capital efficiency and higher market valuation. ORG also increases Sz_score, indicating reduced bankruptcy likelihood. The lack of significant relations for Vol and Beta suggests that ERM maturity and structure do not directly affect market-based volatility and systematic risk over the study period, possibly reflecting broader market dynamics or the short panel length. Compared with mixed results in prior studies, differences may arise from how effectively ERM processes are operationalized and integrated within governance, not only from adherence to standards. On prediction, PLSR effectively handles multicollinearity and improves accuracy for ROA, ROE, Q, Sz_score, and Vol, while robust panel regression remains preferable for Beta. These results highlight the value of combining econometric PD models with data-mining approaches for forecasting performance, value, and risk in financial institutions.

The study provides novel evidence for BIST XBANK banks (2019–2022) that ERM implementation and especially the sophistication of ERM’s organizational structure are associated with improved accounting performance (ROA, ROE), higher firm value (Q), and reduced insolvency risk (higher Sz_score). No significant ERM/ORG effects were found for stock return volatility or market beta. For prediction, PLSR outperformed panel regressions for most DVs except Beta, supporting data-mining approaches under multicollinearity. Contributions include: (1) modeling ERM’s performance–value–risk effects for Turkish banks; (2) integrating PD regression with PLSR to compare predictive accuracy; (3) proposing a pathway toward hybrid prediction–optimization frameworks for ERM analytics. Future research should incorporate richer ERM components and disclosures, broader sectors and jurisdictions, longer pre- and post-pandemic windows, and metaheuristic optimization to tune PLSR components and coefficients, potentially yielding hybrid models to optimize firms’ performance, value, and risk indicators.

The analysis is limited to 10 Turkish banks over 2019–2022, a period influenced by COVID-19; findings may not generalize across time or sectors. ERM measures rely on disclosures and ordinal proxies (implementation level and organizational structure sophistication), which may be incomplete or coarse. Short panel length may constrain identification of effects on market-based risks (Vol, Beta). Data access limitations restricted inclusion of finer-grained ERM components. Future work should expand samples, industries, and geographies, enrich ERM measurement, and separate pre/post-pandemic periods.