Read on Twitter
Some notes on practical auctions vs auction theory
The year's 1959. The FCC allocates radio spectrum basically by soliciting applications and evaluating. Very discretionary process, also very time and resource intensive
The year's 1959. The FCC allocates radio spectrum basically by soliciting applications and evaluating. Very discretionary process, also very time and resource intensive
An economist named Ronald Coase comes up with an interesting idea: let's auction the spectrum! Whoever wants to pay the most for the spectrum should get it. The government would save time processing applicaitons, and would also raise a bunch of revenue from spectrum sales
Coase testifies in front of the FCC in 1959.
The first question: "is this all a big joke?"
The FCC thinks that, with all the work they put in to making sure spectrum ends up in the right hands, simply selling stuff could not possibly work!
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=286932
The first question: "is this all a big joke?"
The FCC thinks that, with all the work they put in to making sure spectrum ends up in the right hands, simply selling stuff could not possibly work!
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=286932
It takes a whole 35 years -- until 1959 -- for the FCC to even bother trying this. IIRC the first big spectrum auction is the PCS auction, one team working on the design is McAfee and McMillan, the other is of course Milgrom and Wilson
https://pubs.aeaweb.org/doi/pdfplus/10.1257/jep.10.1.159
https://pubs.aeaweb.org/doi/pdfplus/10.1257/jep.10.1.159
(A funny intervening episode: at some point before the auctions, the FCC doesn't have bandwidth (pun) to process all the applications so it just assigns by lottery. Coase theorem means the market will work its way to efficiency, right? well it worked terribly)
So the FCC realizes this stuff actually works! I don't know the intervening period as well, but lots of spectrum auctions are tried, in various parts of the world, debate happens about optimal designs and such, and nowadays most spectrum is sold by auctions
Some notes. Spectrum auctions nowadays run on pretty specialized machinery quite different from what is taught in grad theory classes. Stuff like simultaneous ascending auctions, combinatorial clock auctions, clock-proxy auctions, and so on, e.g.
https://www.journals.uchicago.edu/doi/abs/10.1086/262118?casa_token=xoFAj0tTXl0AAAAA:omvBhjZ18rgPSS46xXPvh-hYVGDEBW9SzD9yYqTXR1KthfiYRT8q7vHa6t7u2-yZyt_vLYmmROdP9Q
https://www.journals.uchicago.edu/doi/abs/10.1086/262118?casa_token=xoFAj0tTXl0AAAAA:omvBhjZ18rgPSS46xXPvh-hYVGDEBW9SzD9yYqTXR1KthfiYRT8q7vHa6t7u2-yZyt_vLYmmROdP9Q
Why isn't this taught much? Well it's somewhat specialized to the spectrum auction setting. The 1-unit setting is cleaner for teaching grad students single-parameter mech design.
The institutional constraints around spectrum allocation are really messy and detail-sensitive
The institutional constraints around spectrum allocation are really messy and detail-sensitive
Lots of licenses at once. Overlap between different parcels. Weird combinatorial stuff. The incentive auction involved lots of work from CS folks, like Kevin Leyton-Brown, working out how to combine an auction with basically a knapsack problem solver (iirc)
This is hard stuff, and there is theory behind it -- Paul Milgrom has a line of work on "gross substitutes" and the SAA, matroids, stuff like that
It also doesn't generalize as neatly as 1D mech design which probably is why it isn't taught much.
But it's really important!
It also doesn't generalize as neatly as 1D mech design which probably is why it isn't taught much.
But it's really important!
Point is that the stuff that is actually used in practice is a lot less clean, often, than what is taught in grad school classes. At some point, literatures diverge. Lots of papers which receive attention in the theory lit don't affect much in practice
lots of practically important stuff in specific settings doesn't always receive much attention in the theory lit
Another example: online ad auctions. Companies like Goog and FB have come up with clever hacks like "bid pacing". I worked in FB's ads team in 2013-ish and was fascinated by this thing. It seemed to make a lot of sense but didn't really look like anything I had seen in class
Some CS/AGT folks have started looking at pacing a bit, see e.g. this and related
https://arxiv.org/pdf/1811.07166.pdf
https://arxiv.org/pdf/1811.07166.pdf
Anyways there is no real moral, just some stories here and there. Maybe the moral is, the applied and theory lines of work on auctions seem to have diverged a bit. There is a lot of cool applied stuff that affects billions of dollars that economists pay little attention to
Which is potentially an area for doing some work!
More details available on some of the above, feel free to DM me if you want references etc
More details available on some of the above, feel free to DM me if you want references etc
OK maybe another piece of moralizing
Market design is messy, complicated, detail sensitive.
But so is the real world!
If you want to do work that actually makes a difference in the world, you just gotta deal with that
Market design is messy, complicated, detail sensitive.
But so is the real world!
If you want to do work that actually makes a difference in the world, you just gotta deal with that
One of the many things that inspired me about Paul Milgrom (my advisor!)'s work is how he simultaneously is an incredibly, amazingly good pure theorist -- just look at how incredibly many papers he has in so many literatures
And also is willing to roll his sleeves up and deal with the messiness of real problems to make a difference.
Really inspiring and really a goal that I'm striving towards in my research
Really inspiring and really a goal that I'm striving towards in my research
