Rare financial-distress events create a distinctive decision-support problem in e-commerce logistics ecosystems: missing a distressed partner can be costly, yet excessive alerts can overwhelm limited human review capacity. This study investigates whether large language models (LLMs), when provided with mechanism-structured financial information, can support rare-event screening under extreme class imbalance.
Using Chinese A-share transportation and logistics firms as an empirical proxy for logistics-partner risk monitoring, we construct a two-year-ahead financial-distress prediction task containing 880 firm-year observations from feature years 2013–2022, including 12 ST/*ST cases. All conventional machine-learning (ML) models and eight zero-shot LLMs are evaluated on the same chronological holdout sample of 176 observations, including four distress cases. We develop Financial Narrative Serialization (FNS) to organize 21 financial and textual indicators into interpretable financial contexts and combine it with a Temporal Context-Aware Chain-of-Thought (TCA-CoT) protocol that structures the assessment around liquidity stress, inefficient expansion, and fundamental deterioration. ML benchmarks include unweighted learning, cost-sensitive learning, enhanced features, mechanism-feature ablation, and threshold sensitivity.
Extreme class imbalance produces markedly different screening behaviors across models. At a 0.5 reference threshold, the most selective cost-sensitive ML ensemble identifies two of four distress cases while generating three false positives; lower thresholds can recover additional distress cases but may eventually approach an all-alert rule. TCA-CoT increases recall for five of the eight LLMs and leaves recall unchanged for three. For Qwen-Turbo, recall rises from 0.50 to 1.00 as the number of flagged observations increases from 33 to 104. The additional false-positive observations exhibit, on average, weaker liquidity, higher leverage, poorer operating cash-flow quality, and lower profitability than true negatives. These results indicate that structured prompting can shift the LLM screening boundary toward financially vulnerable observations, although broader risk coverage increases the burden of secondary review.
The findings support a layered screening workflow in which AI prioritizes potentially vulnerable firms for subsequent human assessment. Model selection should therefore reflect the organization’s tolerance for missed risks and its available review capacity rather than rely on a single predictive metric.
This study develops a mechanism-structured financial representation for zero-shot LLM screening and evaluates ML and LLM approaches in a common rare-event environment. It further introduces a decision-support perspective that jointly considers distress coverage, false alarms, and human review volume.
