# 5

A logarithmic scoring rule to elicit a probability distribution on a random variable is . Something that always seemed clear to me but I haven’t seen explicitly written anywhere is that the parameter is just the price of information on .

Firstly: for an agent with true belief , the expected score from making a report is where is cross-entropy. This is maximized when .

Well, this is just the standard proof that logarithmic scoring is proper. This max score itself is i.e. the entropy in . So your expected earning is exactly proportional to the information you have on (the negative of the entropy in your probability distribution for it), and the proportionality constant, the price of a bit of information on , is .

This can be made even clearer by considering the value of some other piece of information . If and you learn this fact, you will bet which would give you an expected score of . Taking the expectation over , your expected score if you acquire is which is the conditional entropy . Thus the expected profit from acquiring is .

So the value of is precisely multiplied by its mutual information with , i.e. is the price of one bit of information on .

I assume this is widely known. But I think it’s still pedagogically useful to actually think in these terms because it sheds light on things like:

• Choosing a good scoring rule — it’s not just that we want the scoring rule to be proper (incentivize honesty), we also want it to incentivize the optimal amount of effort in acquiring information. should be a measure of how important a question is.
• But also: like the price of any good, the price of information can vary (e.g. you might want to reduce after some key information becomes public, since you’re getting it for free)! And like any good it would have diminishing returns. This motivates things like [1].
• It makes clear the fact that prediction markets have huge positive externalities — the market-maker is paying for the information, but it becomes public. This is bad (see also: [2]) — in general, IP rights remain an unsolved problem: [3]. I have a very clever idea to solve it, which I will elaborate in another post.

[1] “Market Making with Decreasing Utility for Information” by Miroslav Dudik et al. https://arxiv.org/abs/1407.8161v1

[2] “Transaction costs: are they just costs?” by Yoram Barzel. http://www.jstor.org/stable/40750776

[3] “IP+ like barbed wire?” by Robin Hanson. https://www.overcomingbias.com/p/ip-like-barbed-wirehtml)

New Comment