Beware of overfitting. A more accurate model of gathered data from a system may not be a better predictor of future behavior of a system.
The above image shows two models of some data.
The linear line is somewhat accurate on the training data (the points on the graph), and (one would expect) it will be somewhat accurate on the testing data (where the points are likely to be for x < 5 and x > -5).
By contrast, the polynomial is 100% accurate for the training data, but (unless you have any reason to believe the 9th degree polynomial is reasonable for some physical reason), you would assume this will be an extremely poor predictor for x > 5 and x < -5.
The linear model is 'less accurate', based on any comparison of errors with the data we have gathered. But it is more generalisable.
Additionally, Engineers have to worry less about their model, and more about what people will do with the model.
If I tell you that we're going on a walk on a hot day and it's expected to last 426 minutes. You are likely to bring less water than if I tell you the walk will last 7 hours, and even less than if I say the walk will last 4-8 hours. This is because you are responding to my implied level of confidence in my forecast, rather than the mid point of my stated times.
If you give people an accurate model, people will reduce their margin of error. This leads to bigger risks.
Taking the walk on a hot day example, if I know the walk will take 4-8 hours in 95% of cases, with some uncertainty around navigation and walking speed. Perfectly knowing our walking speed will decrease the uncertainty of the 4-8 figure, but it won't significantly effect the 'chance of us taking so long that water becomes an issue', because that is driven almost entirely by the uncertain navigation, not the uncertain walking speed.