prediction markets

what are prediction markets?

prediction markets are markets that let you bet on e.g. what'll be the most streamed spotify song this week or whether the knicks or cavs will win tonight. they're somewhat sophisticated—technically you trade event-based contracts whose prices indicate belief in how probable certain outcomes are, etc. etc.—but they fundamentally come down to betting what you think will happen. more info.

are they gambling?

i get this question a lot, and my answer is that i don't think it matters. because what counts as gambling? what counts as investing? most people seem to implicitly subscribe to the notion that while gambling has a negative expected value, investing has a positive one; although in the short term it can be difficult to distinguish between the two, the difference is clear in the long term. and i mean like, sure, i could buy that, but blackjack card-counting and being decent at poker have +EV. how do we reckon with that?

probably a better answer is that gambling is a zero-sum game (and, in the case of prediction markets, negative-sum, considering fees) whereas investing is positive-sum, since the market is a flywheel. but the layman doesn't engage with that, so i think these semantics are a waste of time. the long and short of it is that you should not use these sites; you will almost definitely lose money in the long run.

(note: this page will be more technical from here on out)

documentation

at some point i'll set up an automatically updating table here showing pnl for the strats i run. also, i'll post most of my code and models. but, for now, this'll be blank. i work with a bunch of different groups and i'm not sure who appreciates transparency and to what degree.

what i do now

i currently work on market making low-liquidity instruments.

in particular, i outsource the theo calculation and then focus my efforts on the execution end. i do this because running the nth linear regression over a low signal-to-noise ratio data set just to have an overly backfit model is realllllllly mindnumbing, whereas managing adverse selection, tail exposure, slippage, risk profiles, inventory, etc. poses many interesting problems.

to give just one example, consider a market maker who works with real-world securities. how do they manage their inventory? they can't hold too much inventory because that overexposes them; if the security crashes they're cooked. but since they sell the security, their inventory has to be nonzero. how do they calculate the optimal range? it's a good question. however, in prediction markets, selling a yes contract is equivalent to buying a no contract, so we can sell even with no inventory. so the question shifts to: how do we calculate an upper limit? and then how can we hedge it, thus increasing our limit? which side should we err on? and so on and so forth.

my strat is currently quite profitable (up 34.3% ytd) so i fear that i can't go into more detail until its underlying edge dries up. that being said, my algorithms aren't particularly impressive (after all, i outsource the hard part). i do think i have some pretty nifty techniques w.r.t. execution—but i am very much just picking low-hanging fruit (read: retail). however, i don't consider myself above that. on some level, it's all low-hanging fruit.

what i did prev.

for a time, i ran correlation arb on NFL markets.

the thesis was that although game moneylines were highly accurate, spreads markets were less accurate due to (a) natural lag and (b) the lumpy nature of football scores, which make the probability that e.g. the pats win by at least 2.5 but not more than 6.5 much more difficult to calculate (hence scorigami) than e.g. hockey where score differentials are definitely unimodal and probably normal.

so, i dredged through past historical data to model these expected probabilities and then had a program constantly checking for mispricings and then arbing when one had meaningful +EV (>4 cents). i also worked with my friend on a dynamic model which could consider game-states (i.e. factor in the current score differential, possession, time of game, etc.). both models had high predictive power at lower score differentials but degenerated at higher score differentials. problematically, this meant that they rarely spotted good bets; at lower differentials EV is almost always close to 0 since the deviation from the moneyline is lesser (i.e. it's priced in), and at higher differentials the models fell apart. occasionally, the model would have a great find—it once detected (chiefs lose) + (chiefs win by 2.5) for <70c—but it wasn't worth its strain on api capacity.

i also tried correlation arbing between binary macro events and fomc mention markets

this did not work well! i maintain that the general strat was +EV and powell just had a grudge against me. the thesis was that the fed (or i suppose powell in particular) has a terribly limited lexicon—take a shot each time he says 'dual mandate'—and thus we can predictably hedge binary macroeconomic markets in which the two options are a norm and a deviation (e.g. will a fed governer dissent?). then, with the expectation that a deviation would be mentioned and a norm would not, we can hedge a bet on the norm with a bet on the associated mention.

it wasn't all too rigorous, but i do believe mention markets have crazy latent alpha. i want to return to this idea when trump appoints a new director so that i'll have a fresh slate to analyse.