Late Payment prediction in Fabric instead of BC
Building a late payment prediction model in Microsoft Fabric using AutoML instead of the native Business Central feature. The post demonstrates how to export BC customer ledger data, engineer features from payment history, train an automated machine learning model, and score open invoices to forecast payment dates for cash flow planning.
post Discover Microsoft Business Central - The wonderful world of Business Central
"Only closed invoices (where you know exactly how long it took to pay) are used to train the model."
Key points
- AutoML in Fabric automates model selection, hyperparameter tuning and feature engineering, opening ML to non data scientists
- Business Central customer ledger data reaches Fabric via API or bc2adls; features include payment term days, posting month and weekday, and per-customer payment history
- Only closed invoices train the model; open invoices get predicted days open plus a customer-average baseline and an expected close date
- The trained model is registered in MLflow, and scored results go to a lakehouse table usable in Power BI for cash flow
- The author calls it an experiment; training can take up to 30 minutes
Quotes
- “Only closed invoices (where you know exactly how long it took to pay) are used to train the model.” (Critical principle for ensuring model accuracy by using only reliable historical data without future information leakage)
Context
- Features: AutoML, Late Payment Prediction, MLflow Model Management, Feature Engineering, Data Export via API, bc2adls Integration, Lakehouse Storage
- Versions: October 2018 release
Source: Discover MS BC - Bert Verbeek, community blog. Summary, key points and quotes are derived (CONTENT-NOTICE.md); read the original for the full text.