Predicting inspection outcomes and evaluating port state control targeting using random forests

This study uses global inspection data of 790k inspections and 1.5 million deficiencies (2013 to 2021) which is complemented by 50k incidents and ship particulars of 132k unique vessels. The results show that over 70% of ships that had very serious and serious incidents (2020 to 2021) were not inspected and only 2.5% were detained. The global averages of percentage of inspections without deficiencies is around 50% with high variability across the port state control (PSC) regimes (2013 to 2021). Since there is ample room for improvement to target risky vessels for inspection, it is not recommended to continue with the status quo of the industry by using detention alone as proxy to target future risk. Instead, the study develops 13 prediction models for detention and deficiency types using ML methods by evaluating over 400 risk factors. The results vary across the endpoint of interest but overall, the normal random forests variants outperform the other variants. The top 5 most influential covariates towards prediction are found to be the size of the vessel (GRT), age, previous number of deficiencies within 365 days prior to the inspection, the year of existence of the beneficial owner and safety manager company. These prediction models can be combined with incident type models to enhance targeting of risky vessels and reduce future incidents compared to the current status quo of 70% false negative events.