Utveckling av ett AI-baserat rekommendationssystem för aktiekorgar : En innehållsbaserad metod för att övervinna cold-start-problemet

This thesis describes the design and implementation of a content-based recommendation system for stock baskets on the Moosebit platform (an investment application under development) to overcome the cold-start problem (the difficulty of generating recommendations for new users without interaction history) for new users. User preferences collected through an onboarding questionnaire and basket metadata are both represented in a shared feature space of sector allocations, market-capitalization levels, historical volatility and geographic exposure. To simulate new users, we generate synthetic profiles based on those attributes. We evaluate two recommendation approaches: a K-Nearest Neighbors (KNN) method using Euclidean distance and cosine similarity, and a supervised eXtreme Gradient Boosting (XGBoost) regression model. Performance is assessed across accuracy, relevance, coverage and a self-consistency check in which a basket’s own features (with added noise) are used as a query to verify that the system can still retrieve that basket. The study shows that simple, transparent content-based methods provide effective personalization without requiring any historical interaction data. Future work includes testing with real user data, adding richer financial and sustainability features, and exploring hybrid or online learning extensions.