Design and Evaluation of an Open-Source Continuous Anomaly Detection and Revenue Assurance System for Improving Transaction Record Integrity in Small and Mid-Sized Enterprises

Small and mid-sized enterprises (SMEs) depend on digital transaction records for billing, collections, forecasting, and automated decision support, but many lack continuous assurance controls that are affordable, interpretable, and operable without specialist data-science staff. This paper designs and evaluates an open-source continuous anomaly detection and revenue assurance system intended as the technical reference layer of an Open Revenue Data Assurance, Integrity and Reliability (R-DAIR) approach. The system combines deterministic integrity rules, robust statistical deviation measures, and machine-learning anomaly scores with a six-dimension Transaction Integrity Score (TIS), revenue-risk estimation, severity-based alerting, named-owner routing, remediation logging, and model feedback. The supplied evaluation dataset contains 120,000 transactions from 12 anonymized SMEs, US$28.74 million in gross transaction value, and 4,800 verified anomalies (4.0%). On a 36,000-record held-out test set, the hybrid detector achieved 0.940 precision, 0.930 recall, 0.935 F1-score, 0.978 ROC-AUC, and a 0.25% false-positive rate, exceeding Isolation Forest, One-Class SVM, Local Outlier Factor, Z-score detection, and static business rules. Median ingestion-to-alert latency was 3.8 s, sustained throughput was 420 transactions/s, and average memory use was 1.4 GB. After remediation, composite TIS increased from 93.8% to 98.2%, while estimated leakage exposure declined from US$645,000 to US$223,000, a 65.4% reduction. The results support a vendor-neutral, self-administered, low-resource assurance architecture for SMEs and provide a transferable basis for community financial institutions, subject to domain-specific validation.