News for hackers. Stuff that matters.

au_gambler

There are very high regulatory costs if you're going to run a book/tote and pay out people's bets. But yes, your data could be worth more than your bookmaking margins. If it's worth buying, it's worth using yourself. I like the idea of a private prediction market, something between kaggle and that investment fund that uses homomorphic encryption, the name escapes me. 'Invest' on betting markets, return a share of profits to contributors.
There needs to be some degree of risk aversion given the uncertainties in estimating your edge. I am thinking more along the lines of a CRRA utility function - something a little less aggressive than kelly.
My reasoning is that the 'invisible hand' of many less informed bettors leads to market efficiency in aggregate, particularly in highly liquid markets. Greater liquidity in parimutuel markets makes for odds that are less of a moving target, but it leaves fewer mispriced prospects to capitalise on. So not only do you have to beat a fairly efficient market, but you have to beat it on average by a margin equal to the parimutuel operator's take. In less liquid markets the unknown odds are more of a problem. That's not to say it profits can't be made, just that it's harder imho. With fixed pricing, even though the numbers don't 'add up to one' there either, you can lock in a positive expectation long before price consensus is reached.
A couple of suggestions to improve on this method.

The authors' regression left an intercept or 'adjustment term' of 3.4% - 5.7%. For a perfect bookmaker, this intercept term would be equal to the overround. The number calculated unfortunately averages that overround between different bookmakers and at different times (overrounds often decrease over time). It might be more effective to adjust for the actual overround of each market sampled, i.e. divide each price by the sum of the inverse of the prospects.

They appear to use a flat betting strategy, and the threshold to bet or not was selected based on profitibility. I was simplifying in another comment when I said profitibility should be the goal. In reality it's utility you should be optimizing for. Nobody wants a ultimately profitable system that reads like an EKG, they want a high sharpe ratio. The paper's results are actually very good here, but the trend could be lifted and stabilized further by betting proportionally to expectation, or by explicitly optimizing for such.

In paramutuel betting the prices are not fixed, so you cannot leverage differences between the odds on offer and your estimate of the prospect's chances. The final odds you receive are equal to the total stakes placed on that prospect divided by (sum of all stakes less the operator's take). You can't formulate an optimal staking plan with unknown odds, and you're forced to compete on win rates rather than price ineffeciencies.
> A strategy intended to beat the bookmakers at predicting the outcome of sports games requires a more accurate model than the ones bookmakers have developed over many years of data collection and analysis.

I disagree with this assumption and I think they have painted themselves into a corner because of it. To illustrate, imagine charting win rates against bins of price-implied-chances. $3 horses win roughly 33% of the time, $4 horses 25% for example. It resembles a noisy 1:1 linear relationship. Do the same for your selections and your line will be noisier, but crucially you're not taking bets where the price is worse than your estimate. This can leave a window of profitibility when you subtract the two, even when you are less 'accurate' as measured by win rate or KLD or other measures.

The goal is profitibility, not accuracy. The problem with including the odds you are betting against as a feature for your ensemble is that it dampens that window. If you're right about your selections, you'll bet less and win less. * If you're concerned about the volitility that comes with being less accurate, there are better ways to address that.

I've been doing this for a couple of years and in many ways it's a dream side-project. Location independent, no customers, automatable, and in some jurisdictions tax-free. It can be a little lonely at times though. I would love to chat with anyone else applying tech/math to beat the bookies. Sorry for the throwaway, I'll put a contact in my profile.