Ανάλυση δημοσίων διαγωνισμών μακράς διάρκειας για την ανίχνευση του χρονικού κινδύνου με χρήση τεχνικών μηχανικής μάθησης και επιχειρησιακή προσομοίωση - Λειτουργική αρχιτεκτονική έγκαιρης προειδοποίησης με Streamlit και FastAPI

Αρσενούδη, Αλεξάνδρα

Master Thesis

Author

Αρσενούδη, Αλεξάνδρα

Date

2026-02

Abstract

Title: «Analyzing Long-Duration Public Procurement Procedures for Temporal Risk Detection through Machine Learning and Business Process Simulation: An Operational Early-Warning Archi-tecture with Streamlit and FastAPI» This Master’s Thesis investigates the temporal behavior of public procurement procedures during the pre-award stage, with a particular focus on the early identification and quantification of the risk of severe time delays. To this end, the study develops an integrated early-warning framework that combines Machine Learning techniques with BPMN-based business process simulation. The research is structured around four core research questions addressing: (i) the factors influencing the duration and delay risk of public procurement procedures at the pre-award stage, (ii) the po-tential of large-scale data to reliably distinguish between different temporal regimes under condi-tions of high heterogeneity, (iii) the operational superiority of a two-stage discriminative prediction architecture compared to single-stage regression approaches, and (iv) the transformation of pre-dictive outputs into operationally interpretable scenarios through process simulation. The empirical analysis is based on a multi-country dataset comprising more than 8.6 million public procurement procedures from the period 2009–2024. The results show that procedure durations exhibit pronounced right skewness, heavy-tailed distributions, and recurring spikes at institutional-ly significant temporal thresholds). These findings indicate that evaluating predictive models solely using mean-error metrics is insufficient, as such metrics obscure systematic failures near critical time thresholds that are directly relevant to decision-making. To address this limitation, a two-stage discriminative prediction architecture is proposed, combining a first-stage classifier that es-timates the probability of exceeding 720 days with two specialized regressors for short and long durations, respectively. This approach is evaluated against single-stage regression models and is shown to be operationally superior in terms of temporal risk calibration. The final model achieves a Mean Absolute Error (MAE) of 229.6 days on a fully unseen test set (2023) and near-zero deviation in predicting the proportion of long procedures (ΔLong≥720 ≈ +0.1 percentage points), effectively eliminating the structural bias observed in single-stage approaches. Finally, the predictive results are transformed into early-warning risk indicators, which are embed-ded in a BPMN-based simulation implemented in Bizagi, enabling the exploration of operationally interpretable scenarios. The analysis highlights significant systemic effects associated with high-risk temporal regimes, including a sharp decline in completion rates and a substantial increase in overall process cycle time. Overall, the thesis reframes duration prediction from a point-estimation mechanism into a functional early-warning tool, supporting evidence-based decision-making in public administration and business process management contexts.

Postgraduate Studies Programme

Πληροφοριακά Συστήματα και Υπηρεσίες

Department

Σχολή Τεχνολογιών Πληροφορικής και Επικοινωνιών. Τμήμα Ψηφιακών Συστημάτων

Number of pages

222

Language

Greek

URI

https://dione.lib.unipi.gr/xmlui/handle/unipi/18937

Collections

Τμήμα Ψηφιακών Συστημάτων

Show full item record

Except where otherwise noted, this item's license is described as
Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα