1. The risk associated with the buyer, including their probability of default.
2. The supplier’s risk profile and their probability of default.
3. The risk of a specific transaction, focusing on the likelihood of non-payment for an individual invoice.
How Machine Learning Can Help
Machine learning models can solve various tasks like classifying data or identifying patterns based on historical trends. For instance, a traditional underwriter might analyze financial statements and look for red flags when assessing a borrower’s risk. Machine learning models, however, take this process a step further by analyzing thousands of data points simultaneously, identifying patterns that even the most skilled human might miss.
With enough data, these models can uncover hidden correlations between seemingly unrelated factors—such as whether a company’s social media activity might predict future cash flow stability.
When trained on millions of records, ML models can identify patterns across a vast number of features. A feature is a measurable property of a data object, and the more features provided to a model, the better its predictive power. Advanced models, such as those built using LightGBM, often start with thousands of features but can optimize down to the hundred most relevant ones.
The Importance of High-Quality Data
While machine learning models are powerful, their success is contingent on the quality of the data used for training. Building this dataset requires expertise across risk assessment, data science, and analytics. Every feature must be measurable, reliable, and relevant.
For example, one useful feature could be the probability of a company having a negative account balance in a given month based on its transaction history. Other features might derive from external sources, such as AI-powered sentiment analysis on a company’s Google reviews, providing insight into customer satisfaction.
Example Credit Score Factors
To calculate a credit score for SME underwriting, we would consider various score factors:
- Bank account balance (high or low).
- Sparse payment history.
- Insufficient cash flow.
- Recent negative bank activity.
- Income levels relative to the credit amount requested.
- Recent account openings (e.g., new credit lines or loans).
- Delinquent credit obligations (past or present).
These score factors are designed by risk specialists and may differ depending on the type of credit or loan being offered. For machine learning models, each score factor is represented as a feature, which helps explain why a model might recommend rejecting or approving a loan application.
Data Sources for Training Models
Integrating with third-party platforms like accounting or ERP systems, open banking providers, and eCommerce platforms (e.g., Shopify, Amazon) opens access to vast datasets. Platforms such as Invoys simplify this process, allowing businesses to authorize access to their data in just seconds.
However, the raw data—such as transaction histories, accounts receivable, and accounts payable—needs to be processed into useful features. Here are examples of some potential features:
• Payment terms (e.g., 60 days).
• Company income over the last year or last quarter.
• Net cash from operating activities.
• Merchant Cash Advance History score (e.g., repayment history across various cash advance providers).
Additional data may be sourced from brokers like Dun & Bradstreet, though these reports may have limitations, such as missing regional data or outdated information.
Model Outputs
One possible output of an underwriting ML model is a credit score that reflects the likelihood of default. Another output could be the weighted impact of various features on the final score, helping explain the reasoning behind the model’s recommendation.
Unlike consumer lending, which benefits from detailed credit reports and standardized credit scores (e.g., FICO), SME lending often suffers from limited credit report data. This challenge is particularly acute in invoice financing, where current financial information is crucial for accurately assessing risk.
To overcome these challenges, Invoys provides an AI-driven broker platform that matches businesses with relevant invoice financing offers. By pre-assessing risk through machine learning, we ensure that companies and lenders have access to the data they need to make informed decisions.
