Sustainability proxies and PCA-based financial resilience in romanian real estate firms: fixed-effects evidence and leakage-safe ML prediction

Corporate resilience has become a central concern for real estate businesses facing financing volatility, cost shocks, and rising sustainability expectations. This study considers examines whether accounting-based sustainability proxies are associated with economic-financial resilience and whether nonlinear models improve out-of-sample classification. Using a balanced panel of Romanian real estate firms for 2014–2023 (331 firms; 3,310 firm-year observations), we build a financial-social-environmental sustainability score, from winsorized, within-year standardised financial-statement indicators and construct a multidimensional resilience score using PCA on eight standardised ratios. We combine two-way fixed effects with a leakage-safe temporal prediction design, training models on 2014–2022 and testing them on 2023. The fixed-effects results show weak and selective sustainability-resilience associations: sustainability and resilience classes are independent in 2023 (chi-square = 2.302, p = 0.316), and only the asset-structure/provision-based environmental accounting proxy is significantly associated with resilience; the financial and social pillars are not. This should not be read as evidence from externally validated ESG performance. In prediction, Gradient Boosting provides the strongest benchmark performance on the 2023 hold-out sample (accuracy = 0.9154; balanced accuracy = 0.9106; macro-F1 = 0.9086), while Random Forest remains competitive (accuracy = 0.8550; balanced accuracy = 0.8193; macro-F1 = 0.8295; Cohen’s kappa = 0.7642) and supports SHAP/PD/ALE interpretation. The framework can support comparative monitoring and general screening when ESG disclosures are unavailable, but weaker low-resilience detection limits its use as a definitive early-warning tool.