Filetype
This study examines whether organic content strategy Granger-causes paid advertising decisions, using a weekly panel dataset of organic and paid content on LinkedIn across five companies in the physical security industry over 11 months. We model the binary decision to run paid advertising as a Linear Probability Model estimated via OLS with company fixed effects. Three sequential models test: whet
This thesis investigates the application of Deep Reinforcement Learning (DRL) to continuous time portfolio optimization under proportional transaction costs and jump-diffusion market dynamics. Traditional analytical approaches for solving this frictional allocation problem require solving intractable Hamilton-Jacobi-Bellman (HJB) equations, which suffer from the severe curse of dimensionality as t
Excessive alcohol consumption poses a substantial public health burden, with binge drinking ranking as one of its most common and costly patterns. While existing literature has relied on classical machine learning methods to predict this behaviour, deep learning approaches remain unexplored in this specific context. Using a large population survey, this study bridges this gap by pre- dicting binge
This thesis develops a robust forecasting framework for electricity prices in the Spanish wholesale day-ahead market (OMIE), integrating meteorological, macroeconomic and energy-specific variables through advanced feature engineering techniques. The work opens with a structural analysis of the Iberian market — covering its historical evolution, the marginal price market structure, the growing rene
The chemical regulatory domain depends on large volumes of chemical data which originates from scientific literature, regulatory inventories, safety documents, patents, trade records, public chemical repositories and many more. However, these datasets are prone to contain errors while identifying and declaring their chemicals which can lead to compliance issues, wrongful obligation-mapping, hinder
Reliable socioeconomic data are essential for policymaking and resource allocation, yet traditional collection methods are costly, infrequent, and often lag behind rapidly changing local conditions. Satellite imagery offers a scalable and timely alternative, but most existing work has been conducted in data-scarce environments where model performance cannot be rigorously separated from measurement
This thesis examines whether Swedish municipalities can be grouped into political-socioeconomic regions using unsupervised machine learning methods. Identifying underlying political-socioeconomic structures can contribute to a better understanding of how political preferences and socioeconomic conditions are distributed across municipalities. Political data from the 2022 parliamentary election we
Accurate counting of industrial spare parts is a critical yet error-prone manual process that directly affects supply chain reliability, inventory costs, and a counting system designed for automated spare-parts counting in a warehouse environment with over 200,000 unique part categories. To overcome data scarcity, a copy-paste synthetic data generation pipeline built on SAM2 segmentation was devel
Fraud-related classification was explored with a focus on multimodal learning for combining transaction data and identity document images. Using synthetic transaction data from PaySim and document images from MIDV-2020, FMIDV and FantasyID, the thesis compares tabular-only, image-only and multimodal fusion models. Standard classification metrics and weighted misclassification cost were used for ev
Social engineering fraud has become a growing concern in European payment systems because, unlike other fraud types, it directly targets human psychology rather than technical vulnerabilities, convincing victims to authorize payments themselves. Previous research on fraud detection using machine learning methods has primarily focused on credit card payments. In cases where behavioral features have
Independent validation of forecasting models in the energy markets is increasingly being conducted under the black-box condition, in which only model outputs are observed and internal processes are not. In such an environment, the validity of validation depends entirely on the extent to which observed forecast behaviors provide enough credible evidence for justified reliance. This is difficult bec
This thesis explores how healthcare wholesale distributors can uncover hidden operational inefficiencies by analyzing historical purchasing data. Business-to-business customer databases often track administrative contracts rather than how physical clinics actually order. To address this, the study involved the processing of over 26 million historical invoices to merge highly fragmented administrat
This thesis aims to improve predictive models of marathon finishing times by analysing in-race pacing characteristics. The data collection includes the previous five editions of the Amsterdam Marathon, resulting in a dataset of approximately 75,000 marathon race results. Using unsupervised learning methods, 4 distinct pacing strategies are identified. Pacing and performance differences are discuss
This thesis applies machine learning methods to forecast yearly bilateral migration inflows into 21 European Union countries across 2,638 country-pair corridors, covering the period 2003– 2023. Using data from the OECD, CEPII, and the World Bank — totalling 782 features after imputation — five models are trained and compared: Random Forest, XGBoost, a feedforward neural network, a dropout-regulari
This thesis uses ESG and firm level characteristics to predict GRI compliance, as a binary target, among Swedish listed firms, and examines whether a tree based method improves prediction over a linear baseline. Using 2,124 firm year observations from the Nordic Compass database covering 424 unique firms between 2014 and 2024, we compare logistic regression and random forest. We split the data 80/
This Master thesis attempts to model the Portfolio Management (PM) problem as a Reinforcement Learning (RL) model with a representative environment and relevant agent. The model used was a Proximal Policy Approximation model that estimates a stochastic policy function using a fully connected Multilayer Perceptron. The model performance was compared to the classical Markowitz PM method in addition
The COVID-19 pandemic showed how quickly infectious diseases can spread through international travel and place healthcare systems under severe pressure, increasing the need for coordinated surveillance and preparedness in the European Union. This thesis develops a machine learning early warning system (EWS) for predicting healthcare capacity stress across 19 EU countries using weekly panel data fr
Image data are increasingly used in industrial quality inspection, but their spatial complexity makes them difficult to analyze with traditional statistical methods alone. This thesis extends the CNN architecture proposed by Sabahno (2026) by introducing a context-refinement block based on dilated convolutions, allowing neighboring patch information to be incorporated before classification while p
The Terra/Luna collapse of May 2022 confronted retail cryptocurrency investors with a strategy decision: sell to fiat, buy into declining prices, or hold their existing position. This thesis examines how portfolio size, age, and gender influence crisis response strategy selection using individual-level transaction data from Safello AB, covering 187,291 investors across the crisis window. Three dir
Hypertension remains a primary global driver of cardiovascular disease, yet traditional linear risk estimators struggle to capture non-linear physiological interactions, while complex machine learning ensembles often operate as opaque "black boxes." This thesis addresses this challenge by developing an interpretable machine learning framework and interactive what-if clinical dashboard su
