Machine Learning-Based Cost Variance Prediction for Early Detection of Budget Overruns in Procurement-Intensive Operations
DOI:
https://doi.org/10.63125/7g92wd03Keywords:
Machine Learning, Cost Variance, Procurement, Budget Overruns, Early DetectionAbstract
This study developed and evaluated a machine learning–based framework for predicting cost variance and detecting budget overruns early in procurement-intensive operations. A retrospective longitudinal design integrated financial, contractual, supplier, purchasing, inventory, schedule, and market data from 42 organizations across eight sectors. The final dataset contained 1,286 projects, 4,800 contracts, 3,648 suppliers, 76,420 purchase orders, 186,735 invoices, and 407,925 procurement transactions, with 99.7% completeness and 99.8% financial reconciliation. Mean approved budget was $8.42 million, whereas mean actual cost reached $9.06 million; average cost variance was $0.64 million, or 7.8%. Overall, 57.5% of projects exceeded budget and 25.1% recorded major or critical overruns. The weighted ensemble achieved the strongest regression performance, with an explained variance of .915, mean absolute error of $0.42 million, root mean square error of $0.84 million, and mean absolute percentage error of 5.6%. For classification, the ensemble produced 91.7% accuracy, .942 recall, a .929 F1 score, and a .975 receiver operating characteristic area. Cost-sensitive training increased severe-overrun recall from .711 to .911 and reduced estimated missed financial loss from $118.4 million to $32.6 million. The first stable warning occurred an average of 95.2 days before budget breach, when 38.7% of approved funding remained uncommitted. Schedule slippage, supplier delay, purchase-price escalation, and amendment density were the strongest predictors. Robustness testing-maintained classification accuracy above .887 and receiver operating characteristic area above .949 across sector, organization, and stress conditions. These findings demonstrated that integrated machine learning consistently generated accurate, calibrated, interpretable, and financially meaningful early warnings of procurement-related budget overruns.


