The fact is there will always be exceptions, but in my experience cold, hard data - such as SATs data to generate GCSE target grades and GCSE results to generate A Level target grades is far more reliable than teacher-led predicted grades. I remember an inset day a few years ago where, as a school, our A Level predictions were SHOCKING - something like an 11% level of accuracy.
Obviously some teachers/departments will be better at predictions than others. Some subjects easier to predict too. But my experience was at an otherwise good school, with more experienced colleagues than average, and solid heads of department - yet we still totally missed the mark with predictions.
And, yes, the tendency is to overpredict - whether that is a misguided attempt to fit with Uni requirements or just being ambitious for our pupils. So my view is the more cold data required in the process that teachers need to follow, the better.