This study develops a decision support system for the selection of early-stage startups within the AC75 accelerator, in response to the transition from a vertical mandate focused on the Silver Economy to a generalist approach, which resulted in an increase in deal flow that could no longer be managed manually. Based on a historical dataset of 550 applications, of which only 27 were accepted, a feature selection procedure was conducted using a Decision Tree within stratified K-Fold Cross-Validation, with SMOTE applied exclusively within each fold to prevent data leakage. This process reduced the feature space from 21 to 15 variables, while deliberately removing sector-specific features in order to neutralize historical bias. On this basis, five classifiers (Logistic Regression, Random Forest, SVM, Decision Tree, and Gradient Boosting) were compared, revealing the systematic failure of the standard 0.5 decision threshold to identify successful cases, a phenomenon masked by deceptively high overall accuracy. A dynamic Threshold Moving procedure based on an asymmetric Cost Matrix was therefore introduced and explored through a grid search over eight cost configurations and SMOTE parameters. This approach made it possible to identify two distinct decision-making regimes, conservative and aggressive, separated by a discrete transition. The final validation, conducted using the cost matrix deemed consistent with the accelerator’s business analysis, shows that Gradient Boosting and SVM, despite achieving the same level of economic savings, adopt opposite predictive strategies: the former eliminates False Positives but fails to identify any successful cases, whereas the latter captures 40% of historical successes while tolerating a higher number of errors. The results indicate that the available dataset, due to the small size of the minority class and the absence of qualitative variables, imposes a structural limitation on the system’s predictive capability. The system should therefore be understood not as a substitute for human judgment, but rather as an objective preliminary screening tool to support the investment committee.
Il presente lavoro sviluppa un decision support system per la selezione delle startup early-stage all'interno dell'acceleratore AC75, in risposta alla transizione da un mandato verticale, concentrato sulla Silver Economy, a un approccio generalista che ha determinato un incremento del deal flow non più gestibile manualmente. Sulla base di un dataset storico di 550 candidature, di cui solo 27 accettate, è stata condotta una procedura di feature selection tramite Decision Tree in K-Fold Cross-Validation stratificata, con applicazione di SMOTE esclusivamente all'interno di ciascun fold per evitare data leakage, riducendo lo spazio vettoriale da 21 a 15 variabili e rimuovendo deliberatamente le feature settoriali per neutralizzare il bias storico. Su questa base sono stati confrontati cinque classificatori (Logistic Regression, Random Forest, SVM, Decision Tree, Gradient Boosting), rivelando il fallimento sistematico della soglia decisionale standard di 0.5 nell'identificare i casi di successo, mascherato da un'accuratezza globale ingannevolmente elevata. È stata quindi introdotta una procedura di Threshold Moving dinamico basata su una Matrice dei Costi asimmetrica, esplorata attraverso una grid search su otto configurazioni di costo e parametri SMOTE, che ha permesso di mappare due regimi decisionali distinti, conservativo e aggressivo, separati da una transizione discreta. La validazione finale, condotta con la matrice di costo ritenuta coerente con l'analisi di business dell'acceleratore, mostra che Gradient Boosting e SVM, a parità di risparmio economico, adottano strategie predittive opposte: il primo azzera i Falsi Positivi senza intercettare alcun successo, la seconda cattura il 40% dei successi storici tollerando un maggior numero di errori. I risultati indicano che il dataset disponibile, per numerosità della classe minoritaria e assenza di variabili qualitative, pone un limite strutturale alla capacità predittiva del sistema, che va inteso non come un sostituto del giudizio umano ma come un filtro preliminare oggettivo a supporto del comitato d'investimento.
Ottimizzazione delle decisioni in contesti di classificazione sbilanciata: un approccio Cost-Sensitive applicato al settore delle Startup
FIORITI, DANIELE
2025/2026
Abstract
This study develops a decision support system for the selection of early-stage startups within the AC75 accelerator, in response to the transition from a vertical mandate focused on the Silver Economy to a generalist approach, which resulted in an increase in deal flow that could no longer be managed manually. Based on a historical dataset of 550 applications, of which only 27 were accepted, a feature selection procedure was conducted using a Decision Tree within stratified K-Fold Cross-Validation, with SMOTE applied exclusively within each fold to prevent data leakage. This process reduced the feature space from 21 to 15 variables, while deliberately removing sector-specific features in order to neutralize historical bias. On this basis, five classifiers (Logistic Regression, Random Forest, SVM, Decision Tree, and Gradient Boosting) were compared, revealing the systematic failure of the standard 0.5 decision threshold to identify successful cases, a phenomenon masked by deceptively high overall accuracy. A dynamic Threshold Moving procedure based on an asymmetric Cost Matrix was therefore introduced and explored through a grid search over eight cost configurations and SMOTE parameters. This approach made it possible to identify two distinct decision-making regimes, conservative and aggressive, separated by a discrete transition. The final validation, conducted using the cost matrix deemed consistent with the accelerator’s business analysis, shows that Gradient Boosting and SVM, despite achieving the same level of economic savings, adopt opposite predictive strategies: the former eliminates False Positives but fails to identify any successful cases, whereas the latter captures 40% of historical successes while tolerating a higher number of errors. The results indicate that the available dataset, due to the small size of the minority class and the absence of qualitative variables, imposes a structural limitation on the system’s predictive capability. The system should therefore be understood not as a substitute for human judgment, but rather as an objective preliminary screening tool to support the investment committee.| File | Dimensione | Formato | |
|---|---|---|---|
|
Tesi_Daniele_Fioriti_Laurea_Magistrale (1).pdf
non disponibili
Descrizione: Ottimizzazione delle decisioni in contesti di classificazione sbilanciata: un approccio Cost-Sensitive applicato al settore delle Startup
Dimensione
1.4 MB
Formato
Adobe PDF
|
1.4 MB | Adobe PDF |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.12075/28365