Comment by chollida1 | Hacker News Reader

Comment by chollida1 | Hacker News Reader

Parent Full thread

Modelling for the purpose of "betting" is sort of the definition of my day job. Over the past few years I've started applying some of the same techniques we use in the equity and bond markets to the fantasy sports arena.

The biggest issues I've come across are:

1) small sample size, football teams only play 16 games which means you really don't have alot of data to use.

2) counter party risk, lots of fantasy betting sites spring up and then fold, just like bitcoin. Ideally there would be a new form of prediction market that hosts bets but has a reliable third party escrow the funds.

3) Not enough volume, this is getting better but only due to Draft Kings huge advertising push this past year.

4) Lack of apis, most of these sites seem to want to cater to people who make 1 or two bets. Ideally I'd be able to make 1000's of bets each week.

If anyone else out there is working on models for fantasy sports betting please feel free to drop me a line if you'd like to discuss techniques for modelling( I believe the cool kids now call this machine learning).

I've got a friend who used to work in quant hedge funds. He's moved to a company that exclusively trades sports. There's a lot of interesting alpha in there, and you don't get the same portfolio effects (market up/down) that you get in finance.

One thing that juices returns is access. If you know about the new markets opening you run your arbs on them and make good money.

Hours are weird.

I can't help but be reminded of that Taibbi quote about Wall Street being a vampire squid plunging its blood funnel into anything that smells like money. It's a living, I suppose, but does "running your arbs" create anything of value?

>It's a living, I suppose, but does "running your arbs" create anything of value?

Something similar could be said of the people chasing the ball around the field.

Entertainment has value.

Arbitrage definitely adds value to society. Without arbitrage, traded assets could exist with arbitrary values.

The process of arbitrage involves observing a mis-priced asset and then making a profit from that observation. By doing so, you push the asset towards its correct value. This kind of activity maintains the integrity of a market.

It does up to a point. I think we're well beyond that point, when people pay for microwave relays to get a few milliseconds of arbitrage, because light is faster in air than in fiber. If someone ever drills a hole through the earth's core, it will be traders looking to save latency between New York and Hong Kong (kidding, I hope).

Yes I absolutely believe that there are diminishing returns associated with arbitrage. I would personally think that exchanges should put some lower limit on the 'time resolution' of their markets to prevent HFT investment beyond a certain point.

"Running arbs" is fine and definitely creates value. The problem with vampire squids is their parasitic relationship with govt, i.e. the people with the guns.

1) definitely true, but it's interesting that it doesn't seem to affect the betting so much. There seems to be rapid consensus on what the 'correct' odds are for any given match, with only minor odds fluctuations. Compare this with horse racing, where the odds before the start can swing wildly, even though the same horses may have run many times before. There are different unknowns in the two sports, of course, but it is strange how bettors seem to quantify the football unknowns similarly.

2) All down to your country's silly gambling laws, unfortunately. Europe, and the UK in particular, have got the right idea here. It's overwhelmingly the shady offshore bookies that disappear without paying out.

3) Probably the same reason as 2) really... American football doesn't seem to attract significant betting volumes outside the US. American horseracing is more popular, strangely.

4) There's very few companies with APIs here. Basically only the exchange-based bookies offer APIs. All other bookies want nothing to do with sophisticated players, and if your customers are at the level of automating bets, you probably don't want to be the one taking them on :)

(Edit): I forgot about Pinnacle - they are (I think) the only traditional bookie who offers an API. Mind you, they also are happy to take on profitable customers.

To your point on #2, Augur (augur.net) should overcome a lot of that risk. Probably not all of it, since the oracles can rule against your win, but at least you know no one can make off with your money.

I tried this for a while in MLB but came to the conclusion that rake + irreducible variance made it pretty unattractive. Not sure about other sports. Apparently the real key was knowing all the other players lineups in MTT formats, according to recent stories.

>> small sample size

Here is where a classically trained statistician can really shine. Classical statistics is really all about extracting insight from data when it is hard to come by...

I'm dubious of this claim.

The average player's career is 3 years, the average coaches tenure is 3 years. Teams implement new offensive and defensive schemes even more frequently than that.

Couple those facts with a short season( 16 games) and injuries limiting most players to less than that and it means you really can't use historical stats with any certainty.

Essentially you start from scratch each season.

I mean there is a reason why the phrase "statistically significant" is often used.

if you can figure out how a "classically" trained statistician can create an acurate model with:

1) almost no data

2) inaccurate data

3) seasons in which the next one often bears no resemblance to the previous one

....then please feel free to contact me!

>> I mean there is a reason why the phrase "statistically significant" is often used.

Not sure if this is meant to be satire / sarcasm?

By "classically" trained statistician, i meant someone who has taken a bunch of grad level courses in statistics (as opposed to CS flavoured ML).

A lot of Fisher's and Gosset's work was done in an era of "small" data.

>>if you can figure out how a "classically" trained statistician can create an acurate model with:

>>1) almost no data

>>2) inaccurate data

>>3) seasons in which the next one often bears no resemblance to the previous one

Stats / ML can't help much with 1) and 2). But wrt 3), dealing with issues of non-stationarity is very much a subject of statistical literature.

> just like bitcoin

You mean like many bitcoin companies right?

Yeah I presume he means 'bitcoin exchanges', since bitcoin itself is obviously still here.

Reply on news.ycombinator.com