A Novel approach to ship valuation prediction: An application to the supramax and ultramax secondhand markets.

Accurate ship valuations are very important in ship sales and purchase (S&P) transactions and for marine insurance purposes. It is equally important to select an appropriate valuation methodology. Today, one of the methods is Machine Learning (ML) algorithms stand out in generating better results than traditional methods. The aim of this study is to propose a highly accurate ship valuation model for the supramax/ultramax segment to interested parties using ML methods, with models established on the basis of linear regression. For this purpose, a four-stage path was followed. (i) The first data set, in which the significance of independent variables for supramax/ultramax ships was tested and linear regression models were created with statistically significant variables, covers the period from August 2005 to December 2022. At this stage, a model was first created that takes into account the values of the Baltic Exchange indices in the month of sale as an independent variable, and then another model that takes into account the values in the month of sale and the values in the months before the month of sale as independent variables. (ii) For the two linear regression models created; Price predictions were made with Linear Regression, Decision Tree, Random Forest and XGBoost ML algorithms. (iii) In the next stage; the two models created were simplified to include independent variables that can be easily obtained from the market in the model; and the obtained simplified models were re-predicted with ML algorithms. The fact that the first model and the simplified model are close in terms of prediction performance shows that the simplified model can be used in prediction. (iv) In order to show that the simplified model can produce reliable results when the data set is expanded, 2023 data was added and price predictions were made again with ML algorithms. As a result, the simplified model's predicting performance was further improved with the addition of new data; the model established with the Baltic Exchange indices in the months before the sales month provided a significant superiority over the other model. XGBoost stood out as the best method according to performance criteria.