Isn't an estimate wrong practically by definition, since it is an 'approximation'. Or, the other side of the coin, without specified tolerances, all estimates are right. e.i. "This will take 3 weeks +/- 3 years".
Kahneman's work was in calibrated predictions. We can view time estimates as a special domain of calibrated predictions.
Of course, time estimates are uncalibrated. We can fix this by adding some implicit calibration -- I will use 90%, because this is the calibration I have trained myself to feel such that it is 90% accurate.
If I make 100 time estimates (in the form of "Project X is completed at or before date Y") with 90% confidence, and 50% of them are correct, I'm overconfident. If 100% of them are correct, I'm underconfident. Both are bad.
To solve this problem in the real world, I'd bet that most people make wildly underconfident predictions externally, and make wildly overconfident predictions internally.
But yes, you will nearly always be wrong if you make predictions of the form "Project X with be done at unix time Y," and you can nearly always make the prediction "Project X will be done by the year 3000, assuming the organization is still there and has not decided to abandon the project."
