Hacker News Reader

Comment by soVeryTired | Hacker News Reader

Regarding points 1 and 2, I think any argument that relies on the concept of utility in some essential way is either false or trivial. Take for example the theorem that "utility-maximizing individuals equalize the marginal utilities of goods consumed divided by their prices".

Utility can't be measured directly, and must be inferred from a consumer's behaviour. It's begging the question to infer an agent's utility function from their behaviour and to simultaneously conclude that the behaviour maximises the utility function: that's true by construction.

On the other hand, we can assume a functional form for an agent's utiilty for a given basket of goods. I guarantee that real-world preferences are more complicated than any function we can write down, so any prediction we make about an agent's preferences will almost certainly be wrong.

All in all, utilities are a nice toy, but aren't really any use outside of the world of theory.

View on HN

in econ 101 utility is measured in “utils”

eg 1 banana = 10 utils, 1 apple = 5 utils, 1 car = 100 utils, etc

utils are just a proxy for money and/or time which are interchangeable through wages paid for labor

money is our real world way of valuing the utility of goods, labor, and more generally time (the scarcest resource of them all)

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)

all of game theory also uses the concept of utility to weigh payoffs and there are many many real world applications of game theory

its not perfect but no models are perfect, that’s why we have statistics and probability theory to guide us as well

Except you can't compare the utility of things from different classes like bananas, hammers, cars, overcoats and paintings. It just doesn't make any sense. Finally "utility" is a different way to formulate the same thing as "use value" as opposed to "exchange value".

In Marginalism, you compare utility from agent to agent but you can never attribute an absolute value.

say i pay you $1000 in wages this month

how do you decide how to allocate it between rent, food, savings, entertainment, etc.

most rational people would use some type of internal or subconscious utility calculation to decide what how to budget their income

i dont know what marginalism is or why its useful when you can simply approximate a value by looking at the price of the good/service and use that to compare with other goods and services

There's nothing wrong with the idea of an internal utility calculation. The problem comes in when people try to compare or sum utility across multiple individuals. I might have a personal utility function which says an apple has higher utility to me than an orange, while your personal utility function says just the opposite. It makes no sense to try to mix these together in an attempt to estimate just how much apples or oranges are worth to the two of us collectively.

Money is a reasonable approximation of "aggregate utility" most of the time, but that holds only so long as interactions between individuals remain voluntary and there is enough trade in similar items to establish a "customary price". It doesn't work well for goods which are not customarily traded for money, or as a basis for orchestrating involuntary interference in the market.

> 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.

Reply on news.ycombinator.com