In the Distributed Generation and recurring billing sector, customer default (both corporate and residential) is one of the main factors affecting operational cash flow. Identifying which customers are most likely to delay their invoice payments even before the due date allows for the automation of friendly reminders and preemptive renegotiations.

In this article, I detail the implementation of a predictive model based on the RandomForestClassifier algorithm, integrated directly with tables from the SAP Business One ERP on the SAP Hana database, executed on a schedule within a robust Oracle Solaris UNIX environment.

Choice of Environment

Oracle Solaris is renowned for its unmatched stability and advanced virtualization features (like Solaris Zones). We kept the Python script running natively in an isolated zone directly connected to the SAP Hana data subnet to ensure security and sub-millisecond network latency.

The Predictive Pipeline

The Python script performs the following key operations:

1. Extraction of Open Invoices: Queries in the SAP Hana database looking for customer payment history (table OINV).
2. Feature Engineering: Grouping by variables of historical days late, current invoice value, and past punctuality.
3. Classification via Machine Learning:

4.    from sklearn.ensemble import RandomForestClassifier

5. # Periodically trained model
6. clf = RandomForestClassifier(n_estimators=100, random_state=42)
7. clf.fit(X_train, y_train)
8. ``
9. Probability Calculator (Risk Score): For each open invoice, the probability of default (predict_proba) is generated.

score_invoices.py writes only CardCode and a Risk_Score` above 0.85. The ERP sends the notice. The model does not need a route to the internet.

Business Integration

When the risk score exceeds the configured threshold (typically 85%), the system automatically triggers the ERP API to send a preventive notification via email or WhatsApp, optimizing accounts receivable by up to 35%.