It feels like, well, only yesterday that I was writing up the results for Round 5 and here we are, talking about the wagers for Round 6.

This week, the AFL's only managed to drag the round's seven games out over four days, but we start again on Thursday night so, come Sunday, that means there'll have been an AFL game on somewhere in the country on nine of the preceding eleven days. On current form, I'm not sure Investors can afford that rate of activity.

Anyway, this week we've wagers in all seven contests, most of them line wagers on underdogs. The two head-to-head wagers come at round's end and comprise a modest wager on the Dons, short-priced and at home to the Suns, and a large wager on the Pies, not quite as short-priced but also at home to the Dogs.

Here's the detail:

The Line Fund is particularly confident about the Dees' chances with 7.5 start away to the Eagles. Its 9.1% wager on them is its 3rd-largest for the season, behind only its 12.3% losing wager on the Saints in Round 2 and its 9.3% winning wager on West Coast in the same round.

This Fund also has two "I'm quite confident" sized wagers on Carlton with 3.5 start facing the Swans and on the Lions with 18.5 start taking on the Tigers, an "icing on top" wager on the Pies that'll be an additional collect to the head-to-head bet should the Pies cover the 34.5 point spread, and a couple of "mildly interested" wagers on Port getting 15.5 start and the Saints getting 4.5 start.

When line wagering predominates, the upside and downside in each game tend to be matched, which is what we see in this week's Ready Reckoner for all but the last two games of the Round.

All seven contests have Home team favourites this Round, but in four of them the Away team is preferred by the majority of Head-to-Head Tipsters. The underdog Melbourne are preferred to the Eagles, as are Carlton (playing the Swans), Port (playing the Roos), and the Saints (playing the Crows). 

The Margin Predictors aren't anywhere near as willing to abandon the favourites. Collectively they're backing six of the seven favourites - five of them unanimously or nearly so - and tipping the upset only in the Sydney v Carlton game. Even for that game there's solid support for the favourites, the Swans, and the average margin prediction has the Blues victorious by less than a point.

Currently, Bookie_3's MAPE is 0.47 points per game better than Combo_NN_2's, which means the sum of its Absolute Prediction Errors across the 38 games we've had is just 18 points less than that of Combo_NN_2's. If every result favoured Combo_NN_2 this week it could pick up 64 points on Bookie_3 and thus take the lead.

I calculated the figure of 64 by summing the absolute differences between the margin predictions of Bookie_3 and Combo_NN_2. This is equivalent to calculating what's called the "manhattan distance" between the predictions and represents an interesting approach to answering the question "quantitatively, how different are the margin predictions of any two Margin Predictors?".

Here's the matrix of pairwise differences, or manhattan distances, for this week:

Based on this metric the most distant - and therefore most different - margin predictions come from ProPred_7 and Bookie_9. Their predictions are 111.7 points apart, or almost 16 points per game. ProPred_7 is also over 100 points distant from Combo_NN_1.

The most similar predictions belong to H2H_Unadj_3 and H2H_Unadj_10. Their predictions are only 1.7 points apart.

Calculating this same distance measure, averaged per game across all six rounds of the season so far, we see that the pattern of pairwise distances for this week is similar to that for the entire season.

One notable feature of to glean from this table is how relatively different have been the margin predictions of Bookie_3 and Combo_NN_2, which we can contrast with how similar have been their MAPE performances. Also, notice how distant Combo_NN_1 is from all other Margin Predictors. The Predictor most similar to it is Bookie_3, but even its predictions have, on average, been almost 8 points different per game.

Another way to measure the pairwise similarity between Margin Predictors is to calculate pairwise correlations. One way to think about these is that they tell us the extent to which we can predict one Predictor's predictions based on another Predictor's.

These correlations tell a very similar story to the manhattan distances.

Manhattan distances and correlation matrices tell us how similar are any two Predictors, and the colour-coding that I've added hints at some higher-order clustering of Predictors which a factor analysis makes clearer still. 

(Please click on the image for a larger version.)

(Here I've plotted the first two factors of the factor analysis. Combined these factors explain over 96% of the variability in margin predictions across all Predictors.)

The clustering is now quite distinct. The H2H clan form one cluster, the twin ProPreds and Wins form another, relatively close to the H2H household, the two Bookie Margin Predictors live on the other side of town with Combo_NN_1 and Combo_7 in residence nearby. Combo_NN_2, as all our analyses have been showing, is loosely related to the other strongly performing Predictors, but clearly stands apart from them.

On then to the Probability Predictors where there's not much to speak of this week other than to mention in passing the mildly contrarian views of ProPred and H2H in the Sydney v Carlton game and the more strongly contrarian views of H2H in the Roos v Port game.