ST-CRAAN: A Spatio-Temporal Hybrid Deep Learning Architecture for Large- Scale Longitudinal Default Prediction via Multi-Modal Data Fusion
Year: 2026
Type: preprint
Abstract: Preprint. This is the author-prepared preprint of a manuscript on spatio-temporal deep learning for agricultural credit-risk prediction. It has not been peer reviewed. Any subsequently published version may differ from this record. This preprint presents the Spatio-Temporal Credit Risk Agricultural Assessment Network (ST-CRAAN), a novel hybrid deep learning architecture designed for large-scale agricultural credit scoring. By treating default prediction as a spatio-temporal sequence-modeling problem, ST-CRAAN addresses the limitations of traditional static tabular models. Multi-Modal Data Fusion: The architecture successfully fuses two structurally incommensurate data modalities at a nationwide scale: a longitudinal financial panel of 1,496,721 borrowers (over 4.49 million records) and a petabyte-scale, multi-sensor satellite raster archive reduced via Google Earth Engine. Hierarchical Fallback Reconstruction: Introduces a reconstruction algorithm combining temporal interpolation with a three-level spatial back-off to guarantee gap-free monthly satellite input sequences, effectively resolving the issue of persistent monsoon cloud contamination. Economically Calibrated Asymmetric Loss: Handles severe class imbalance natively by employing a cost-sensitive loss function derived from the lender's minimum retail rate. This penalizes misclassified defaults based on realized economic damage rather than relying on synthetic resampling techniques. Superior Predictive Performance: Under a strict Out-of-Time evaluation on an unseen test set, ST-CRAAN achieved a test recall of 81.11%, exceeding the strongest gradient-boosted baseline (XGBoost) by 13.23 percentage points. This performance translates to a 41.2% reduction in false negatives and the avoidance of 7.01 billion baht in misclassified exposure. Keywords / Index Terms: Cost-sensitive learning, credit risk prediction, data engineering, deep learning, explainable artificial intelligence, long short-term memory, multi-modal data fusion, remote sensing, spatio-temporal data. Data availability: the borrower-level financial panel is restricted administrative data provided by the Bank for Agriculture and Agricultural Cooperatives (BAAC), Thailand, and is not publicly available. Earth observation inputs derive from public MODIS, CHIRPS, SMAP, and VIIRS products. Analysis code is available from the corresponding author on reasonable request. Ethics: ethical review was waived; the study uses strictly secondary, fully anonymized administrative data containing no personally identifiable information, consistent with Thailand's Personal Data Protection Act (PDPA).
Institution: University of the Thai Chamber of Commerce
Language: English
Subfield: Accounting
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7204207209