Finance Automation
FastAccounting
Party-group classification at 95.9% accuracy across 5 companies' books.
About the Project
A proof of concept on 5 companies' Tally exports: Random Forest models with text features classify transactions by party and expense group (95.9% average party-group accuracy), XGBoost and Prophet forecast recurring expenses 3 months ahead, and Isolation Forest flags unusual entries for review.
Technical Overview
Data Sources
- Tally daybook exports
- Voucher and ledger entries
- Transaction narrations
- 60 months of history
ML / NLP Stack
- Random Forest classification
- TF-IDF text features
- XGBoost and Prophet forecasting
- Isolation Forest anomaly detection
- MLflow and Optuna experiment tracking
Key Challenges
- Inconsistent narrations and vendor names
- Imbalanced expense categories
- Irregular and sparse expense series
- Anomalies without labels
Interested in this solution?
Let's discuss how we can adapt this approach for your business.
Get in Touch