A model for predicting the win propensity of new B2B opportunities across multiple companies

Abstract: For a sales representative, a complex part of their work is determining the outcome of a potential deal with a customer. Most (if not all) processes today are based on guesses and gut feelings, making it hard to track progress. Given great advancements in digitizing processes of companies, many sales companies use a software system to manage deals and relations with customers, known as a CRM system. This project aims to help determine the outcome of a sales opportunity using supervised learning and data from the CRM systems of hundreds of companies. A Random Forest classifier is trained on three types of data, company, financial, and data from interactions; the constructed classifier reaches an accuracy better than sales representatives using only a limited set of numerical features. The classifier is evaluated both by itself and in comparison with predictions done by sales representatives, using accuracy, precision, recall, and Brier score. The evaluation showed an accuracy of 69.3% for the random forest classifier, compared to an accuracy of 66.8% for sales representatives, a statistically significant difference. The study shows similarities in the predictions made by the classifier and sales representatives. When observing the results, the classifier shows better performance across all evaluated performance metrics. The only metric where the sales representatives managed to outperform the classifier was certainty, the results of which were just slightly in favor of the representatives. The certainty of predictions is measured using Brier score, presenting a score of 0.22 for the sales representatives and a score of 0.31 for the classifier. Unlike previous research, the proposed classifier is trained and predicts on the opportunities of multiple companies.