Customer complaint analytics in digital financial services: temporal prediction of untimely responses and capacity-aware prioritization

Abstract Digital financial service providers often receive more customer complaints than can be examined through intensive review and individualized handling. Prior research has examined technology adoption, service quality, satisfaction, trust, loyalty, and artificial intelligence in finance, but less is known about how complaint records can support transparent prioritization when review capacity is limited. This study develops a customer complaint analytics framework for capacity-constrained digital financial services using public Consumer Financial Protection Bureau records. Because realized churn, customer lifetime value, intervention histories, and treatment effects are unavailable, the empirical target is not customer defection or a proxy constructed from complaint text. Instead, the study predicts an independently observed post-intake service process outcome: whether the company subsequently provides an untimely response. The final analytical sample contains 266,337 complaints received between January 2024 and April 2026. Model development is restricted to 2024–2025 through three blocked temporal validation folds, while January–April 2026 is reserved as an untouched final holdout. Structured attributes, sparse lexical features, and fixed dense semantic embeddings are compared under the same temporal design, with mean development average precision (AP) used for model selection. The selected term frequency–inverse document frequency (TF–IDF) logistic model achieves a final holdout area under the receiver operating characteristic curve (ROC AUC) of 0.7653 and AP of 0.04985. Within the highest ranked 10% of complaints, precision is 4.83% against a holdout prevalence of 1.24%, corresponding to a lift of 3.89 and capture of 38.94% of all untimely responses. The study therefore shows that intake time complaint information contains predictive signal available at intake for identifying complaints that later receive untimely company responses and can support capacity-aware early review. These results should not be interpreted as evidence of customer churn, retention effects, consumer harm, or causal service recovery benefit.