Predicting the success of Ukraine's restoration projects: a machine learning analysis using DREAM ecosystem data

Abstract

Ukraine’s full‑scale war has generated 157 billion USD in infrastructure losses and an urgent 524 billion USD reconstruction bill. While prior scholarship isolates single drivers of project delivery, it rarely combines finance, governance, digital capacity and societal sentiment in one empirical frame. This study merges five open datasets (DREAM project register, DREAM‑Completeness audit, Transparent Cities scores, Digital‑Index metrics and IRI opinion surveys), yielding a moderate‑N panel of 190 fully documented wartime restoration projects. A six‑pillar theoretical model is operationalised through logistic regression and three ensemble learners. Results show that fragmenting procurements into additional contract lots multiplies completion odds by ≈9.8; publishing real‑time finance schedules raises success probability by 12 percentage points independent of budget size; and a one‑SD increase in regional digital maturity adds five points, but only where e‑services accompany raw openness. Findings nuance ‘transparency backlash’ theory and suggest an integrated policy bundle: mandatory micro‑lotting, conditional disbursement upon schedule disclosure, and targeted e‑government investment.