Machine Learning-Based Early Warning Systems for Predicting Cost and Schedule Overruns: A Conceptual Framework

Abstract

Cost and schedule overruns remain a persistent and costly feature of large-scale infrastructure, construction, information technology, and public-sector projects, with deviations frequently exceeding half of the originally approved budget or timeline. Conventional monitoring practices, anchored in periodic manual reporting and static earned-value formulas, typically flag deviations only after corrective action has become difficult or expensive. This paper presents a conceptual framework for machine learning-based early warning systems (ML-EWS) capable of identifying overrun risk at considerably earlier stages of project execution than conventional methods allow. Drawing upon a structured synthesis of existing literature (2015–2026) and established theoretical foundations including Risk Management Theory, Systems Theory, and the Technology Acceptance Model (TAM) the study classifies the categories of predictive features relevant to overrun forecasting, compares the performance profiles of algorithms commonly applied to the problem (such as random forests, gradient boosting machines, support vector machines, artificial neural networks, long short-term memory networks, and Bayesian networks), and sets out an integrated architecture that fuses structured project management data with unstructured textual indicators processed through natural language techniques. A five-layer early warning framework is proposed, spanning data ingestion, feature engineering, model training and validation, risk scoring, and decision-support alerting. To address critical conceptual gaps regarding algorithmic validation, empirical lead-time claims, and performance metrics, this paper articulates a rigorous mathematical formalization of lead-time gain (ΔL), dynamic evaluation metrics (AUC-PR, Fβ-score), and synthetic portfolio simulation frameworks tailored to imbalanced overrun datasets. The paper further examines data governance implications, an illustrative simulation of risk scoring across a project portfolio, and the practical constraints limiting adoption within developing-country project environments such as Nigeria. Findings indicate that hybrid ensemble models drawing on financial, schedule, contractual, organizational, and textual features consistently outperform single-technique models, while explainability mechanisms such as SHAP remain indispensable to sustaining stakeholder confidence in model output. The paper concludes with testable conceptual propositions, a structured future research agenda, recommendations directed at project management offices seeking to institutionalize predictive analytics, and policy-level suggestions for national infrastructure oversight bodies.