Purpose Financial distress prediction is important in the construction industry, where long project cycles, high capital intensity and sensitivity to macroeconomic conditions increase bankruptcy risk. Existing financial distress prediction models primarily rely on structured financial indicators or require high-quality external textual disclosures that are often unavailable for many construction enterprises. To address this limitation, this study develops an embedding-enhanced hybrid prediction framework that leverages the semantic representation capability of pre-trained language models without requiring external textual data, with the aim of improving early-warning performance for construction firm financial distress prediction. Design/methodology/approach A hybrid predictive framework is proposed that combines structured financial ratios with high-dimensional semantic embeddings generated from purpose-designed textual prompts using a pre-trained sentence embedding model. The embedding vectors are concatenated with the numerical features and used to train a gradient boosting decision tree (GBDT) classifier. The model is evaluated using a panel dataset of publicly listed USA construction firms and compared with seven benchmark machine learning (ML) models and a standalone GBDT across one-year and two-year prediction horizons. Findings The proposed semantic embedding-enhanced embed-GBDT framework demonstrates superior predictive performance. Across ten independent experiments, the model achieves average AUC values of 0.9203 and 0.8224 for the one-year and two-year prediction horizons, respectively, outperforming seven benchmark machine learning models and a standalone GBDT model. The framework also achieves stronger precision-recall balance and substantially higher F1-scores. Statistical tests and threshold stability analysis further suggest that embedding-enhanced features provide more robust representations of financial information and contribute to improved long-term predictive performance. Originality/value This study proposes an indicator-to-semantic embedding mechanism that transforms structured financial, market and macroeconomic variables into standardized firm-year textual profiles before semantic encoding. Rather than extracting embeddings from existing disclosures, the framework generates semantic representations from the economic meaning and relational structure embedded in the indicators themselves, enabling embedding-enhanced distress prediction under limited-text conditions.
Wang et al. (Tue,) studied this question.