"So in general, the difference is between data fitting--so you have the data, and you fit model on that. And many studies--also in economics--stop there, and report a great fit or R-square.
The proof is in predictions. Whether that model that fits so wonderfully, actually predicts. And there is statistical theories like the bias-variance dilemma where one can understand that, in prediction, you're better to make it simpler.
And these are usually what's heuristics are. Heuristics have a bias. So, they are simple. You can't have many free parameters to fit them. But they reduce the error, what's called by variance. That means over-fitting. They're not fine tuning on the past. And, fine tuning on the past only pays if the future is like the past. That's in a situation of risk but not under uncertainty."