Description
Résumé: La prédiction des entreprises qui connaîtront une croissance rapide, et les raisons pour lesquelles elles le font, fait l'objet de recherches depuis de nombreuses décennies. Les chercheurs explorent l'utilisation de techniques d'apprentissage machine supervisées (modèle de filet élastique, forêt d'arbres décisionnels et filet neuronal) pour trouver une population embryonnaire d'entreprises à forte croissance (EFC) dans les données administratives canadiennes des entreprises dérivées de dossiers fiscaux anonymisés. Le vaste ensemble de variables comprend des indicateurs de l'industrie, de la géographie et de la complexité (p. ex., de multiples activités industrielles), de même que la propriété étrangère. Même si environ une entreprise sur huit de l'ensemble de l’échantillon deviendra une EFC, la méthodologie fait ressortir un sous-échantillon d'environ 25 % à 30 % de cette dimension, c'est-à-dire qu'une sur quatre parviendra à ce statut et que trois résultats sur quatre seront faussement positifs. Les variables de l'industrie sont d'une importance manifeste sur le plan des prédictions, tout comme les variables indicatrices d'entreprises plus petites et (particulièrement) plus jeunes. Cependant, les chercheurs conseillent de ne pas interpréter l'analyse comme causale.
Abstract: Predicting which firms will grow quickly—and why—has been the subject of research studies for many decades. We explore the use of supervised machine-learning techniques (elastic net model, random forest, and neural net) to identify a population of nascent high-growth firms (HGFs) in Canadian administrative firm-level data derived from anonymized firm tax records. Our large set of variables include indicators of industry, geography, complexity (e.g., multiple industrial activities), and foreign ownership. While about one firm in eight of the overall sample will prove to be an HGF, the methods identify a subsample about 25–30 percent of the size, where one in four will achieve that status, meaning that there are still three out of four false positives. Industry variables are clearly important for prediction, as are variables that indicate smaller and (especially) younger firms. However, we caution against interpreting our analysis as causal.
Abstract: Predicting which firms will grow quickly—and why—has been the subject of research studies for many decades. We explore the use of supervised machine-learning techniques (elastic net model, random forest, and neural net) to identify a population of nascent high-growth firms (HGFs) in Canadian administrative firm-level data derived from anonymized firm tax records. Our large set of variables include indicators of industry, geography, complexity (e.g., multiple industrial activities), and foreign ownership. While about one firm in eight of the overall sample will prove to be an HGF, the methods identify a subsample about 25–30 percent of the size, where one in four will achieve that status, meaning that there are still three out of four false positives. Industry variables are clearly important for prediction, as are variables that indicate smaller and (especially) younger firms. However, we caution against interpreting our analysis as causal.