> utility functions will never describe an individuals preferences perfectly (hell you couldn’t perfectly describe the utility of any of the transactions you ma

> utility functions will never describe an individuals preferences perfectly (hell you couldn’t perfectly describe the utility of any of the transactions you make) but they’re useful when describing and predicting aggregate behavior, especially when fed a lot of consumer data (higher confidence) I think you're describing some sort of regression there. I don't doubt that regression is a useful tool. But in that use case I think the notion of 'utility' is redundant: if I want to predict aggregate behaviour, I just collate some datapoints and stick a line (or logit, or softmax etc...) though them. Why do I need to concern myself with whether my regression approximates the agents' 'utility'? I think game theory shares a flaw with many other constructs in economics: in the real world, it's actually quite rare that the payoffs are given to you. Instead you have to hypothesize payoffs that hopefully kinda-sorta reflect your preferences. There's a lot of theory about optimal behaviour in games, but in reality the hard part is setting the problem up. As an example, I used to build trading strategies for a hedge fund. We'd usually optimise the parameters of a strategy to try to maximise its Sharpe ratio. But the strategy's Sharpe didn't actually reflect what we wanted to achieve. A 20% gain was nice and would get the team paid, but a 20% loss was an existential threat to the company. We were never able to find a 'utility function' that encoded that information without introducing all sorts of other weird artifacts. In the end, we just used a Sharpe ratio and accepted that it didn't actually encode our beliefs.