Prediktion av Utfall av Onlineauktioner med Maskininlärning

This study evaluated several machine learning models, including logistic regression, CatBoost and a feed-forward neural network, in predicting the outcome of online auctions—namely whether items would sell or remain unsold. Additionally, this study assessed the economic and environmental impacts of rejecting those items forecasted not to sell. Using data on listed items from the online auction market Auctionet and from the local auction house Kolonn, a Swedish auction house selected as representative of an average auction house, this study engineered categorical, numerical and temporal features alongside multilingual text embeddings derived from BERT. Each model was tuned to optimize the macro-average F1 score and assessed based on accuracy, macro-average F1 score and area under the curve (AUC). CatBoost achieved the highest overall accuracy and AUC while the neural network attained the greatest macro-average F1 score and significantly reduced unnecessary seller round trips resulting in notable potential CO 2 emission savings. A cost-benefit analysis revealed that CatBoost maximizes revenue for Auctionet, whereas the neural network delivered better economic and environmental benefits for the average auction house. Logistic regression provided a clear and interpretable baseline, but performed slightly worse than the other two models across all key metrics. These results highlighted trade-offs between predictive accuracy, profitability and sustainability, informing platforms’ decisions to select the model that best fits their strategic priorities.