2011 Round 2 Results: A Round of Learning

Well that's settled a few things. The Gold Coast weren't a 1,000 MARS-Rated team at the start of the weekend (and they certainly aren't now), games where the line is under 6.5 points do seem as though they'll be tricky propositions this seasons, Investors aren't going to enjoy a season of unremitting joy, and if it's truly the mean we've just regressed to, I don't want any part of it.

Last week we lost just 1 wager of 13, and this week we won just 1 wager of 8, the net result of which has been to drag the Head-to-Head Fund into loss and the Line Fund back to a far more modest profit.

I was comforting myself with the thought that we'd probably seen in these first two weeks the best and the worst wagering results of the home-and-away season until I worked through the probabilities and calculated that the chances of this being true were about 1 in 310,224,200,866,620,000,000,000 - the likelihood of selecting at random two pre-specified objects from amongst 24 of them.

That's about one forty-thousandth the likelihood of selecting a single grain of sand marked "winner" from amongst all the grains of sand on the planet. (Be a tough job writing "Sorry, try again" on all the other grains wouldn't it?) By any standard, that's an unlikely event - although surely an order of magnitude or two higher than the Gold Coast's flag chances - so I'm resigned to the fact that we'll see at least one worse result sometime this season. Of course we might encounter a better result instead or as well, but I'm struggling with a glass half-full attitude to MAFL at the moment.

Retaining a sense of perspective for just a moment though, in the same way that we humans learn most when we encounter something that we don't expect, statistical models thrive on variability and "surprise" in the probabilistic sense, so this week's results were a veritable learning bonanza for the Head-to-Head and Line Funds. I just hope they're both quick studies.

Our wagering struggles were mirrored in the week's tipping performances with only Combo_NN_2 doing better than 5.5 from 8. The cumulative results for Rounds 1 and 2 are shown in the table below.

The three leading head-to-head tipsters are on 11 from 16 (about 69%). Two of these tipsters are algorithms built with Eureqa - Bookie_9, which uses the TAB Bookmaker's probability assessments for input, and H2H_Adj_7, which uses the (adjusted) probabilities from the Head-to-Head Fund. The third tipster from the leading bunch, Combo_NN_2, is a neural network built using Tiberius and takes as inputs the competing teams' MARS Ratings and bookmaker prices, as well as information about whether or not a game is an interstate clash.

Almost of all of our traditional heuristics are back on just 8 from 16, BKB and HSH being the only exceptions, both on 10.

A bookie-based predictor, Bookie_3, leads all Margin Predictors on the MAPE metric with a score 25.89 points per game, and is one of eight Predictors currently with a sub-30 MAPE. As we've often seen in the past, however, exceptionally strong performance on one metric rarely implies outstanding performance on another. Here, Bookie_3's impressive MAPE hasn't translated into an ability to predict the winners on line betting. In fact, if we used each Predictor's projected margin as the basis for making line bets, Bookie_3 would be ranked third-last.

ProPred_7 and ProPred_3 lead all tipsters on this metric, having correctly selected 12 of the 16 line winners so far this season, even though their MAPEs are over 3 points a game worse than Bookie_3's. As at the end of Round 2, the correlation between MAPE and Line Betting performance across all Margin Predictors is +0.02 signifiying a complete absence of a linear relationship.

Finally, on probability scoring we find that, for the head-to-head probability tipsters, last week's ordering has been completely reversed, leaving the TAB Sportsbet bookmaker atop the pile of four and H2H in last with a score just below what you'd expect from random guessing.

The Line Fund's performance in assigning probabilities to line outcomes has also dipped below chance level.