A logarithmic scoring rule to elicit a probability distribution r on a random variable X∈{1…n} is s(r)=blog(rX). Something that always seemed clear to me but I haven’t seen explicitly written anywhere is that the parameter b is just the price of information on X.
Firstly: for an agent with true belief p, the expected score from making a report r is Ep[s(r)]=∑x∈{1…n}blog(rx)px=−bH(p,r) where H is cross-entropy. This is maximized when r=p.
Well, this is just the standard proof that logarithmic scoring is proper. This max score itself is Ep[s(p)]=−bH(p) i.e. the entropy in p. So your expected earning is exactly proportional to the information you have on X (the negative of the entropy in your probability distribution for it), and the proportionality constant, the price of a bit of information on X, is b.
This can be made even clearer by considering the value of some other piece of information Y. If Y=y and you learn this fact, you will bet P(X|Y=y) which would give you an expected score of EP(X∣Y=y)[P(X∣Y=y)]=−bH(P(X∣Y=y)). Taking the expectation over Y, your expected score if you acquire Y is −bEP(Y)[H(P(X∣Y=y))] which is the conditional entropy −bH(X∣Y). Thus the expected profit from acquiring Y is −b(H(X∣Y)−H(X))=bI(X;Y).
So the value of Y is precisely b multiplied by its mutual information with X, i.e.b is the price of one bit of information on X.
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.b 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 b 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.
LMSR subsidy parameter is the price of information
A logarithmic scoring rule to elicit a probability distribution r on a random variable X∈{1…n} is s(r)=blog(rX). Something that always seemed clear to me but I haven’t seen explicitly written anywhere is that the parameter b is just the price of information on X.
Firstly: for an agent with true belief p, the expected score from making a report r is Ep[s(r)]=∑x∈{1…n}blog(rx)px=−bH(p,r) where H is cross-entropy. This is maximized when r=p.
Well, this is just the standard proof that logarithmic scoring is proper. This max score itself is Ep[s(p)]=−bH(p) i.e. the entropy in p. So your expected earning is exactly proportional to the information you have on X (the negative of the entropy in your probability distribution for it), and the proportionality constant, the price of a bit of information on X, is b.
This can be made even clearer by considering the value of some other piece of information Y. If Y=y and you learn this fact, you will bet P(X|Y=y) which would give you an expected score of EP(X∣Y=y)[P(X∣Y=y)]=−bH(P(X∣Y=y)). Taking the expectation over Y, your expected score if you acquire Y is −bEP(Y)[H(P(X∣Y=y))] which is the conditional entropy −bH(X∣Y). Thus the expected profit from acquiring Y is −b(H(X∣Y)−H(X))=bI(X;Y).
So the value of Y is precisely b multiplied by its mutual information with X, i.e.b is the price of one bit of information on X.
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.b 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 b 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)