Predicting detention and deficiencies using random forests

Research output: Working paperAcademic

5 Downloads (Pure)

Abstract

The aim of this exploration study is to predict detention and twelve deficiency types which can be used to enhance port state control targeting as well as domain awareness for coastal administrations. A total of 234 combinations of random forest variants are explored evaluating over 400 covariates. The study uses a comprehensive and unique, global inspection dataset of over 200k inspections and 400k deficiencies (2014 to 2019) and out of sample data from 2020 to 2021 for evaluation. The results show that based on the used data, normal random forests outperform other variants and overall detention has the highest decile lift with 3 or higher compared to random selection. This is followed by the deficiency groups safety of navigation, certificates and qualification and the Maritime Labor Convention. Deficiencies related to newer areas such as MARPOL Annex VI, ballast water treatment and anti-fouling are more difficult to predict and are also more difficult to detect compared to other areas where detection often depend on the training and background of inspectors. Future work will evaluate further model variants and evaluate inspection policies by filtering out high risk vessels that were missed.
Original languageEnglish
PublisherEconometric Institute, EUR
Publication statusPublished - 2022

Fingerprint

Dive into the research topics of 'Predicting detention and deficiencies using random forests'. Together they form a unique fingerprint.

Cite this