This paper investigates whether firms’ overtime intensity affects the cost of debt financing. Using satellite-based night-time light data for Chinese listed firms between 2013 and 2022, we construct an objective measure of weekday overtime that captures firms’ operational effort and capacity utilization. We find that higher overtime intensity is associated with significantly lower bond offering yield spreads. The effect is stronger among smaller, less-followed, less-profitable, and non-AAA-rated issuers, consistent with an information-asymmetry channel where investors rely more on observable operational behavior when hard information is weaker. The findings suggest that overtime functions as a priced form of soft information in debt markets, offering new evidence that real-time operational signals influence credit risk assessment.

This paper creates default-probability-(PD)-implied credit ratings for Chinese firms following the S&P global rating standard. The domestic credit rating agency (DCRA) ratings are higher than the PD-implied ratings by ten notches on average for the identical level of default risk, implying that the domestic ratings are significantly inflated. The PD-implied ratings outperform the DCRA ratings in detecting defaults and complement the latter in predicting yield spreads. The PD-implied ratings draw information from operating efficiency-related variables; in contrast, the DCRA ratings pay attention to scale-based firm characteristics in credit risk assessment.

Over the last decade, China’s P2P lending industry has been seen as an important credit source but it has recently suffered from a wave of bankruptcies. Using 126,090 P2P loan deals from RenRen Dai, one of the biggest online P2P websites in China, this paper attempts to predict credit default probabilities for P2P lending by implementing machine-learning techniques. More specifically, thisstudy proposes a stacking ensemble machine-learning model to assess credit default risk for P2P lending platforms. A Max-Relevance and Min-Redundancy (MRMR) method is used for feature selection and then irrelevant features are eliminated by using k-means clustering method. Finally, the stacking ensemble model is performed to produce accurate and stable predictions in the feature subset. Experimental results show that stacking ensemble model yields high performance, not only in prediction accuracy but also in precision and recall. In comparison to single classifiers, the stacking ensemble machine-learning model has a minimum error rate and provides more accurate credit default risk prediction. The results also confirm the efficiency of the proposed stacking ensemble model through the area under the ROC curve.