Hope I got everything.
Quote:
Originally Posted by Tall Paul
I'm pretty sure I've made my position clear, I'm not arguing "against analytics". I understand they have value, but they aren't a substitute for watching the games and I believe that the xG model in particular has plenty of flaws.
On the bit in bold, I'm not a stats expert (when I did my maths A-Level it was with mechanics, not statistics) so I'm not entirely sure how the closing lines can be regressed like that. Continuing the poker analogy, we know the chances of a flush draw coming in with one card to come because there are a finite number of cards in the deck and even if we didn't know that, the exact situation can be replicated a large number of times.
In football, every game is a unique event. For example, that Arsenal team will never play that Burnley team again under the same conditions, so how can we ever know how well the lines predict the result?
Maybe I'm being thick...
Quote:
Originally Posted by Tall Paul
I've no idea how weather forecasting works either, I agree that it's a better comparison, but no I'm not looking for other examples of imperfect prediction models.
I'm looking for someone to explain Baba's statement that it's possible to regress the closing lines of match results to show how accurately they predict them.
I'm not sure how that's a strawman. The statement was that closing lines accurately reflect the strength of a team and this can be proven by regression, but I don't see how that's the case.
As I said, I don't have any education in statistics, but my (possibly erroneous) logic tells me that you can't perform regression on a set of results when there are so many variables which are constantly changing.
Quote:
Originally Posted by Tall Paul
I know how it works, I've read the explanations, and I'm not dismissing them as useless. I've lost count of how many times I've said that I accept they have value.
We were told that it's possible to prove by regression that the market is very (spookily) accurate, I've asked for that proof and nobody has yet provided it, or explained how it works.
We also established that the market sometimes gets it wrong and I asked why the default position when a team over/under performs the market is rungood/bad and not perhaps the market is wrong. Still waiting for the answer to that as well.
Obviously you can never know how accurate the market was for a single game (if/when we develop superintelligence - and assuming we survive it - and if it turns out complete neuroimaging of every atom in the brain can be used to perfectly predict future behavior I guess in theory you could have a perfect model ... but for now, nope). Unless a certain game outcome is predicted to be 100 %, by definition the result will always deviate from EV.
Looking at the aggregate level, e.g.
Franck, Verbeek and Nüesch (2010) look at football. Their focus is on showing that a non-restricted market (Betfair) predicts match outcomes better than soft bookies, but you can still look at the summary statistics or goodness-of-fit models to get an idea of BF's standalone predictive power. One thing to note is that they assume the vig (=overround) is equally distributed over the outcomes. This is a very good approximation but I would say suboptimal (or to put it another way, the market prediction could be thought of as any probaility derived from the vig range and there are reasons to skew it a bit based on market movements but I won't say more on that). Anyway, the effect on overall results is very marginal. I'd also say it's better to use Pinny's historical CLs and if you want you can freely get all that data from football-data.co.uk and run your own regressions.
That said, what can you say about individual games then? You have to realize that by flipping 20 different coins 38 times it would be far more unlikely for all coins to flip very close to their EV than one/a few to run significantly hot/cold. Thus pointing out teams that severely over or under perform their market predictions as an example of market failure is misguided. Now, if over a large sample the amount of teams over/under performing is significantly higher than you'd expect from pure variance, you might be onto something.
The way to do it is by constructing your own model and showing it's profitable over a statistically significant sample. Let's say the key distinction between the TPmodel© and what Starlizard et al. use is that your model better utilizes shot blocking data. Turns out you can even beat the CL for 101 % ROT and given that close to game time you can bet 6 figs on PL without moving the line, your bankroll is soon large enough to correct mispricings -> TPmodel© IS now the CL. Of course, you still won't know how your model performs for any given one game, but at least it's reasonable to assume your edge comes from what you consider your innovation.
FWIW I think it would be naive to consider the PL market perfectly accurate as there's always room for improvement, though in general the (big) sports betting markets are
better predictors than the stock market given almost all relevant information is publicly available, transaction costs are lower and it's much, much easier to model. Perhaps looking at the importance of momentum is the next big breakthrough...
Last edited by Babalatexi; 01-25-2017 at 02:37 PM.
