Leveraging Visual Embeddings from Instagram for Credit Scoring of Informal Microbusinesses

ICLR 2026 Conference Submission21875 Authors

Keywords: Alternative Credit Scoring; Financial Inclusion; Informal Microbusinesses; Instagram Data; Computer Vision; Visual Embeddings.

TL;DR: We leverage Instagram images and videos to enhance credit scoring models for informal microbusinesses, improving predictive accuracy and supporting financial inclusion.

Abstract: Access to formal credit remains limited for informal microbusinesses in Latin America, forcing entrepreneurs to adopt predatory lending practices characterized by exorbitant interest rates. Several fintech startups have sought to address this challenge by developing alternative credit scoring methodologies leveraging artificial intelligence and non-traditional data sources. This paper presents a novel approach that applies computer vision techniques to Instagram images and videos from microbusiness accounts, extracting visual features to improve predictive models of creditworthiness. The proposed method utilizes pre-trained vision language models, such as CLIP and X-CLIP, to obtain visual embeddings. Subsequently, dimensionality reduction (UMAP) and clustering (KMeans) techniques are applied to derive discriminative features. In addition, two distinct architectures are introduced: a Fully Connected Neural Network (FCNN) processing CLIP embeddings, and a separate Convolutional Neural Network (CNN) directly analyzing image data, each generating predictive visual scores. Preliminary results show that visual embeddings improved AUC by 2.16 points and F1-score by 9.86 points, with visual features contributing 25.52\% of predictive power—underscoring the potential impact of computer vision-based methodologies on financial inclusion among underserved communities.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 21875