Search results

The accurate valuation of second-hand vessels has become a prominent subject of interest among investors, necessitating regular impairment tests. Previous literature has predominantly concentrated on inferring a vessel's price through parameter estimation but has overlooked the prediction accuracy. With the increasing adoption of machine learning for pricing physical assets, this paper aims to quantify potential factors in a non-parametric manner. Furthermore, it seeks to evaluate whether the devised method can serve as an efficient means of valuation.

This paper proposes a stacking ensemble approach with add-on feedforward neural networks, taking four tree-driven models as base learners. The proposed method is applied to a training dataset collected from public sources. Then, the performance is assessed on the test dataset and compared with a benchmark model, commonly used in previous studies.

The results on the test dataset indicate that the designed method not only outperforms base learners under statistical metrics but also surpasses the benchmark GAM in terms of accuracy. Notably, 73% of the testing points fall within the less-than-10% error range. The designed method can leverage the predictive power of base learners by incrementally adding a small amount of target value through residuals and harnessing feature engineering capability from neural networks.

This paper marks the pioneering use of the stacking ensemble in vessel pricing within the literature. The impressive performance positions it as an efficient desktop valuation tool for market users.