Imputation av Saknad Finansiell Fundamentaldata : Fallstudie om Imputering av Finansiell Fundamentaldata för Systematisk Kapitalförvaltning

Missing data is a pervasive challenge across many data-driven domains, often included as one of multiple steps in pre-processing pipelines. In finance, the issue becomes particularly pronounced: domains like systematic investing rely heavily on Financial Fundamental Data (FFD) to extract predictive signals, but FFD is frequently incomplete, often in systematic and structurally complex ways. This thesis addresses the problem of handling missing FFD through imputation. A core contribution of this thesis lies in designing an evaluation strategy for imputation methods, a non-trivial task given the absence of labeled ground truth. To simulate realistic conditions, the thesis constructs a labeled dataset from fully observed entries and reintroduces missingness based on observed patterns. Two types of missing patterns are evaluated through the masking, enabling comparison of different evaluation design choices. This enables a consistent and realistic evaluation framework while avoiding look-ahead bias via strict time-based train-test splits. Five imputation techniques have been implemented and evaluated: Mean Imputation (MI), Last Observation Carried Forward Imputation (LOCFI), K-Nearest Neighbors Imputation (KNNI), Multiple Imputation by Chained Equations (MICE), and Miceforest (MF). Their performance was assessed using Mean Relative Error (MRE) across a multi-year, multi-sector dataset of over 13,000 publicly listed companies. Several modeling choices were examined, including whether to train imputation models globally or within sectors, and how the amount of historical data influences performance. Results show that LOCFI achieved the best overall performance, with MICE, MI, and KNNI performing well for specific variables. Sector-specialized models outperformed global models in several cases, highlighting the value of domain-informed modeling. Imputation accuracy improved with more historical data, though gains leveled off after a certain point. The implemented imputation methods yielded lower MREs when block masking had been applied compared to regular masking. These findings emphasize that both method selection and evaluation design choices are critical for effective imputation of missing financial fundamental data.