Retraining RightShip Qi

As with all predictive models, RightShip Qi is undergoing the first of its regular adjustments, known as a ‘retrain’, in September. Retraining is a ‘business as usual’ practise for a predictive model.

The practical outcome of a model retrain is that a minority of vessel star ratings may move (either up or down), however most will stay the same.

What a model retrain does is:

  • preserves the desirable features of the original model
  • remedies identified shortcomings
  • increases performance compared with original, and
  • adapts the model to changes in input data since the first training.

At RightShip, we anticipate a retrain of Qi every 4 – 6 months. We won’t do it too often (we want to keep Qi stable, rather than changing constantly), or too infrequently (we want to keep Qi current).

As this is our first retrain since launching Qi nine months ago, our learning curve is more acute, and there is also data associated with millions of different occurrences to take into account. As such – even though most vessel ratings will not change – there will be greater than usual ‘movement’ of risk ratings in this first retrain compared with subsequent retrains.

We have refined how we name some of our risk inputs to better reflect what they are measuring:

  • ‘PSC Risk’ is now called ‘PSC Performance’
  • ‘DOC PSC Risk’ is now ‘DOC PSC Performance’, and
  • ‘Size Risk’, ‘Age Risk’ and ‘Type Risk’ have been simplified to ‘Size’, ‘Age’ & ‘Type’.

As the ‘alert’ function has not yet been set up in Qi, you will not receive automated messaging if you vessel ratings move either up or down. The development of this alert function is a complex task but a priority for us, and we anticipate that it will be available for future retrains.

RightShip’s focus remains firmly on improving maritime safety standards, and Qi increasingly helps us target those vessels with a high likelihood of having an incident over the next 12 months. With this in mind we believe that it would be confusing and therefore counter-productive to run two systems concurrently, and as such there will be no ability to preview the retrain results before they come into effect.

Please contact if you have a specific query regarding the re-train process.