Italy's productive base depends heavily on small and medium-sized enterprises (SMEs), and demographic pressure together with the decline of intra-family succession is making the continuity of many such firms dependent on external ownership transfer. Mergers and acquisitions (M&A) offer a viable continuity channel, but post-acquisition outcomes depend on the quality of the match between target and acquirer, and the upstream identification of suitable acquirers remains a costly bottleneck that the AI-oriented M&A literature does not directly address. As a proof of concept, this study develops an AI-assisted recommendation prototype that supports the sell-side phase of external succession by ranking and screening candidate acquirers for Italian SMEs from observable pre-deal information. The analysis draws on completed controlling acquisitions of Italian targets from Orbis M&A, combined with year-industry benchmarks from BACH. Acquisition success is captured by three benchmark-adjusted accounting outcomes, return on assets (ROA), return on sales (ROS), and revenue turnover (TURN), each modeled separately. For each outcome, two target-specific CatBoost models are trained on leakage-controlled predictors: a ranking head that orders candidate acquirers and a screening head that shortlists the most promising ones. Models are evaluated under acquirer-grouped cross-validation and interpreted with SHAP. On a base sample of 2,615 deals, the ranking head attains Spearman correlations between predicted and realized outcomes of 0.67, 0.63, and 0.64 for ROA, ROS, and TURN, while the screening head identifies top-quartile matches with an area under the ROC curve between 0.81 and 0.85. The top predicted decile outperforms the bottom by 71 to 76 percentile points, and this performance holds out-of-time and on acquirers never observed during training. The recommendations rest on an interpretable, match-based mechanism dominated by the target's own pre-deal starting position and the fit gaps between the target's and acquirer's financial profiles. The framework is deployed as a publicly accessible web prototype, paving the way for a more advanced acquirer-recommendation tool.
La base produttiva italiana dipende in larga misura dalle piccole e medie imprese (PMI), e la pressione demografica, insieme al declino della successione intra-familiare, rende la continuità di molte di queste imprese dipendente dal trasferimento della proprietà a soggetti esterni. Le fusioni e acquisizioni (M&A) sono un canale di continuità praticabile, ma gli esiti post-acquisizione dipendono dalla qualità dell'abbinamento tra target e acquirente, e l'identificazione a monte di acquirenti adeguati resta un collo di bottiglia costoso che la letteratura sull'AI nell'M&A non affronta direttamente. Come proof of concept, questo studio sviluppa un prototipo di raccomandazione assistito dall'AI che supporta la fase sell-side della successione esterna, ordinando e selezionando potenziali acquirenti per PMI italiane sulla base di informazioni osservabili prima dell'operazione. L'analisi si basa su acquisizioni di controllo completate di target italiani da Orbis M&A, con benchmark anno-settore da BACH. Il successo è misurato da tre outcome contabili corretti per benchmark, return on assets (ROA), return on sales (ROS) e revenue turnover (TURN), modellati separatamente. Per ogni outcome, due modelli CatBoost per target sono addestrati su predittori controllati per leakage: una ranking head e una screening head, che ne seleziona una shortlist. I modelli sono valutati con cross-validation raggruppata per acquirente e interpretati con SHAP. Sul campione base di 2.615 operazioni, la ranking head raggiunge correlazioni di Spearman di 0,67, 0,63 e 0,64 per ROA, ROS e TURN, mentre la screening head identifica gli abbinamenti nel quartile superiore con area sotto la curva ROC tra 0,81 e 0,85. Il decile con previsione più alta supera quello più basso di 71-76 punti percentile, e tale performance si mantiene fuori campione temporale e su acquirenti mai osservati. Le raccomandazioni si fondano su un meccanismo interpretabile basato sul match: la posizione pre-deal del target e i gap di fit tra i profili finanziari di target e acquirente. Il framework è reso operativo come prototipo web pubblicamente accessibile, aprendo la strada a uno strumento di raccomandazione più avanzato.
Supporting italian SMEs facing succession: an AI-assisted acquirer recommendation approach
de GROOT, DORIAN SILVIO
2025/2026
Abstract
Italy's productive base depends heavily on small and medium-sized enterprises (SMEs), and demographic pressure together with the decline of intra-family succession is making the continuity of many such firms dependent on external ownership transfer. Mergers and acquisitions (M&A) offer a viable continuity channel, but post-acquisition outcomes depend on the quality of the match between target and acquirer, and the upstream identification of suitable acquirers remains a costly bottleneck that the AI-oriented M&A literature does not directly address. As a proof of concept, this study develops an AI-assisted recommendation prototype that supports the sell-side phase of external succession by ranking and screening candidate acquirers for Italian SMEs from observable pre-deal information. The analysis draws on completed controlling acquisitions of Italian targets from Orbis M&A, combined with year-industry benchmarks from BACH. Acquisition success is captured by three benchmark-adjusted accounting outcomes, return on assets (ROA), return on sales (ROS), and revenue turnover (TURN), each modeled separately. For each outcome, two target-specific CatBoost models are trained on leakage-controlled predictors: a ranking head that orders candidate acquirers and a screening head that shortlists the most promising ones. Models are evaluated under acquirer-grouped cross-validation and interpreted with SHAP. On a base sample of 2,615 deals, the ranking head attains Spearman correlations between predicted and realized outcomes of 0.67, 0.63, and 0.64 for ROA, ROS, and TURN, while the screening head identifies top-quartile matches with an area under the ROC curve between 0.81 and 0.85. The top predicted decile outperforms the bottom by 71 to 76 percentile points, and this performance holds out-of-time and on acquirers never observed during training. The recommendations rest on an interpretable, match-based mechanism dominated by the target's own pre-deal starting position and the fit gaps between the target's and acquirer's financial profiles. The framework is deployed as a publicly accessible web prototype, paving the way for a more advanced acquirer-recommendation tool.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_07_deGroot_Executive Summary_02.pdf
non accessibile
Descrizione: Executive Summary
Dimensione
403.79 kB
Formato
Adobe PDF
|
403.79 kB | Adobe PDF | Visualizza/Apri |
|
2026_07_deGroot_Thesis_01.pdf
non accessibile
Descrizione: Thesis
Dimensione
2.06 MB
Formato
Adobe PDF
|
2.06 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/260907