Abstract
Small and Medium-sized Enterprises (SMEs) play a significant role in most economies by contributing to job creation and economic growth. A majority of such merchants rely on business financing, and thus, financial institutions and investors need to assess their performance before making decisions on business loans. However, current methods of predicting merchants' future performance involve their private internal information, such as revenue and customer base, which cannot be shared without potentially exposing critical information. To address this problem, we first propose a novel approach to predicting merchants' future performance using credit card transaction data. Specifically, we construct a merchant network, regarding customers as bridges between merchants, and extract features from the constructed network structure for prediction purposes. Our study results demonstrate that the performance of machine learning models with features extracted from our proposed network is comparable to those with conventional revenue- and customer-based features, while maintaining higher privacy levels when shared with third-party organizations. Our approach offers a practical solution to privacy concerns over data and information required for merchants' performance prediction, enabling safe data-sharing among financial institutions and investors, helping them make more informed decisions on allocating their financial resources while ensuring that merchants' sensitive information is kept confidential.
Introduction
Small and Medium-sized Enterprises (SMEs) are the backbone of the global economy, yet they face a staggering 33% failure rate within their first two years. Financial institutions need to predict which merchants will thrive to make informed lending decisions. However, there is a "data deficiency" gap: unlike public corporations, SMEs aren't required to publish financial reports. Traditional prediction models rely on highly sensitive internal data—like exact revenue and customer lists—which merchants are reluctant to share for fear of competitive leaks. The challenge is finding a way to assess a business's health using "privacy-safe" proxies that can be shared with third-party analysts without exposing a merchant's trade secrets.
The core innovation of this research is the realization that consumer behavior patterns are a powerful proxy for business health. If a merchant's customers also frequent high-performing or diverse businesses, that structural position carries a predictive signal. By shifting from looking at a merchant in isolation to looking at their position within a "Merchant Network," we can extract valuable insights while maintaining privacy.
Methodology
The researchers proposed a novel framework for constructing a merchant network and extracting predictive features.
Network Construction
The team built an undirected graph where each node represents a merchant. An edge exists between two merchants if the same customer shopped at both. The "weight" of the edge is determined by the number of shared customers. This approach treats customers as "social bridges" between businesses.
Feature Extraction
Instead of raw financial metrics, the study extracted several sets of features to compare their predictive power. These included traditional merchant, customer, and revenue features, alongside novel network-based features.
The network features included:
- Centrality Metrics: Degree, betweenness, closeness, and eigenvector centrality to measure a merchant's "importance" and "connectedness" in the local economy.
- Diversity Metrics: Using Shannon entropy to calculate the geographical and business-category diversity of a merchant’s "neighbors" (the other shops their customers visit).
- Node2vec: This algorithm was used to create 128-dimensional vector embeddings of the merchants, providing a layer of abstraction that is harder to reverse-engineer than standard network metrics.
Performance Labeling
A new three-tier performance metric (well-performing, medium, and poorly-performing) was defined based on the relative rate of change in revenue, transaction count, and unique customers over two six-month periods. These were benchmarked against peers in the same industry category (Merchant Category Code, or MCC) to account for seasonality and industry-wide trends.
Data
The study utilized three primary data sources:
- Credit Card Transactions: A large-scale dataset from a major bank in an OECD country covering July 2014 to July 2015 (2.5 million transactions, ~62,000 customers, and ~4,500 merchants).
- Points of Interest (POI): Data from Here.com to identify surrounding amenities (parks, transport hubs, etc.) within a 200-meter radius of each merchant.
- Demographics: District-level data on population and average household income.
Key Results
The study yielded several significant findings:
- Comparable Performance: Models using only network-based features (AU-ROC ~0.58) performed nearly as well as those using private revenue-based features (~0.59).
- The Power of Combination: The best results (AU-ROC ~0.61) came from combining revenue, customer, and network features, proving that network position captures unique signals that raw financial data misses.
- Top Predictors: Feature importance analysis showed that while transaction count remains the strongest predictor, network features like degree centrality and betweenness ranked in the top 10 most influential variables for the Random Forest classifier.
- Validation: The researchers confirmed that their performance labels weren't biased by the wealth of a merchant's district or the income level of their customers, suggesting the model measures business "hustle" rather than just "location luck."
Conclusion
This research demonstrates that network features are a viable, privacy-safe alternative to sensitive financial data for predicting merchant success. Knowing a merchant has high "betweenness centrality" tells a bank they are a vital hub in the community without revealing their exact daily take-home pay.
However, limitations remain. The study relies entirely on credit card data, which may be incomplete in cash-heavy economies. Additionally, while network embeddings are safer than raw data, they are not perfectly anonymous and could theoretically be subject to inference attacks. Future work should focus on incorporating defense mechanisms like matrix perturbation to further harden privacy and tailoring models specifically to SMEs as more metadata becomes available.
Summary updated