It's March already and I've not written a thing here about my plans for MoS for the upcoming season.
Mostly those plans involve changing very little again this year, this in part a reflection of my belief - some might call it conceit - that, after 10 iterations of this thing, I should be nearing some sort of solution. To be honest though, it's also an acknowledgement that there's a lot already going on here at MoS, and every change just adds to the cognitive load of a visitor.
In summary, the changes for this season are:
- The designation of MoS Preferred Head-to-Head, Margin and Probability Predictors
- The addition of two new ensemble Margin Predictors. They'll also participate as Head-to-Head Tipsters
- The re-optimisation of the ChiPS Team Rating System, which has direct, flow-on effects for the C_Marg Margin Predictor and the C_Prob Probability Predictor
- The deletion of the MoS Margin Fund and the reweighting of the remaining Head-to-Head and Line Funds
All of the Head-to-Head Tipsters from 2014 will reappear in 2015. The Heuristic Tipsters will return for the duration of the home-and-away season, and the various Margin Predictors (including the two, new ensemble-based Predictors) will have their head-to-head predictions inferred from the sign of their respective margin predictions, while WinPred and ProPred will have their probabilistic opinions binarised on the basis of whether or not their assessments of the home team's chances are above or below 50%.
This year, as foreshadowed earlier, I'm designating a single Tipster or Predictor of each type as "MoS Preferred", and amongst the Head-to-Head Tipsters that moniker will be attached to Win_3 for season 2015. This designation serves no other purpose than to force me to select, ahead of time, the single Tipster that I think will perform best in the season ahead, which I've done here in Win_3's case largely on the basis of recent-season performances. Also, if others are looking for a single Head-to-Head Tipster with which to associate the MoS site, Win_3 is it.
As for the Head-to-Head Tipsters, all of the 2014 Margin Predictors will be returning for another year. They'll be joined by two new ensemble predictors, ENS_Greedy and ENS_Linear, the genesis of which, if you're technically inclined, you can read about in this blog post from the Statistical Analysis journal. The simplest way to think about these new Predictors is that they're based on a collection of statistical algorithms, the output predictions of some (ENS_Greedy) or all (ENS_Linear) of which are averaged to derive the Predictor's final opinion.
Specifically in this case, both ensemble predictors are constructed from 13 underlying predictive models (the 16 models from the blog post just linked but excluding nnet, svmLinear and gamSpline), though only ENS_Linear incorporates the opinion of all of them. ENS_Greedy uses just four of the models for its purposes: Sparse Partial Least Squares, weighted 77%; Ordinary Least Squares, weighted 14%; Weighted k-nearest neighbours, weighted 7%; and Bagged Earth, weighted 2%.
These new Predictors are the latest and, by far, most complex of my experiments with ensemble prediction in AFL, and their performances in 2015 will be an interesting empirical assessment of the applicability of this methodology in an AFL context. Their inclusion is the one indulgence in season 2015 of my desire for novelty - the only example of a difference relative to 2014 that isn't merely a change or a reduction.
In seasons to come I hope and expect that one of these ensemble predictors will become the MoS Preferred Margin Predictor, but for this year that title will be held by Combo_7, its choice also largely a reflection of recent-season form.
There'll be no changes in the lineup of Probability Predictors this year, with the same eight Predictors re-entering the competition (though I will admit to toying with the idea of deleting one or other of the two H2H Probability Predictors).
On form, any of the three Bookmaker-based Predictors would be sensible choices for MoS Preferred status, but in the end I've selected Bookie - OE for this role.
I've concluded, based on considerable empirical and theoretical evidence, to reject the null hypothesis (p < 0.01) that it's possible to make money reliably on SuperMargin wagering. So, this year, there'll be no Margin Fund (and so no Chasms of Death in the Ready Reckoners either). I'll probably miss the excitement that went along with knowing that an 8/1 payout was but a smother or an out-of-bounds away, but I'll not miss at all the stark realisation that a SuperMargin wager is never more than a score away from switching from in-the-money to out-of-the money. And, anyway, as we've seen, the overround in that market on all wagers appears to be prohibitively large.
No other Fund will be created to take the Margin Fund's place, which means that MoS Investors will be running with only a Head-to-Head and a Line Fund for the first time since 2011. Portfolio weightings will be 40% for the Head-to-Head Fund and 60% for the Line Fund, the latter Fund's higher weighting reward for the fact that no MoS Line Fund has lost money since 2009. The staking strategies for both of these Funds will be the same as last year, with the Head-to-Head Fund Kelly-staking using a variable cap that varies based on the Round, and the Line Fund level-staking and also using a Round-based variable cap.
(This post, though now 3 years old, is still a good place to start to understand the various Tipsters and Predictors on MoS. Then, this post, from a year later, will help explain some of the changes and provide details of the staking strategy I've just broadly described.)
Both MARS and ChiPS Ratings will again be reported in 2015, though I've taken the opportunity in the off-season to re-optimise ChiPS. Its parameters have now been chosen to minimise the Mean Absolute Error of its margin predictions across the past five seasons only.
Key amongst those parameters are estimates of Home Ground Advantage (HGA) for each team at each home ground, the values for which appear in the table at right. One of my main reasons for re-optimising ChiPS was to ensure that these HGAs better reflect teams' more-recent performances as, previously, ChiPS' HGAs were based on all performances from 1999.
Most of the HGAs are, as you'd expect, positive, though a few remained stubbornly negative as the optimisation process progressed, most notably the Pies' near 4-goal negative HGA when playing as the home team at Docklands, and the Tigers' more than 3-goal negative HGA when playing at the MCG.
There are a handful of other, smaller negative HGAs, though bear in mind that the non-Victorian teams will frequently enjoy the 6.5 point Interstate Status benefit when playing at home and the Victorian teams will do the same, albeit less often, when they play non-Victorian teams at home.
The largest differences between the HGAs shown here and those used for ChiPS in 2014 are for:
- Adelaide and Port Adelaide when playing at the Adelaide Oval. Last year, since the two teams were unfamiliar with the venue, I treated it as a neutral venue for both, which meant that they suffered a -2 point HGA when playing there.
- Richmond, whose -1 point HGA at the MCG has been reassessed as a -19 point HGA.
- West Coast, whose +13 point HGA has been reassessed as a -3 point HGA.
- Melbourne, whose +10 point HGA at the MCG has been reassessed as a -4 point HGA.
- Collingwood, whose -9 point HGA at Docklands has been reassessed as a -23 point HGA.
- Essendon, whose +3 point HGA at Docklands has been reassessed as a -7 point HGA.
- The Bulldogs, whose -1 point HGA at Docklands has been reassessed as a +13 point HGA.
It's important that these changes in estimated HGAs be considered along with the changes in ChiPS underlying team Ratings. When a team plays at home, ChiPS' expectation about their performance is driven by a combination of its assessment of the team's underlying ability, its assessment of the HGA for the team at the venue, and by its assessment of the underlying ability of its opponent. So, for example, ChiPS would expect a home team Rated 1,020 with a +10 point HGA to beat an opponent Rated 1,000 by 30 points, but would also expect the same result were it, instead, to Rate the home team 1,015 but assess the HGA as +15 points.
For some teams then, ChiPS' re-estimation of their HGAs has come with a countervailing reassessment of their underlying Rating. ChiPS has, if you like, reallocated the source of previous victory margins between HGA and underlying ability.
To illustrate this, the table at left reveals the end of season team Ratings under the ChiPS Ratings System as it prevailed in 2014, and the ChiPS Ratings System as it would have functioned in 2014 with the new ChiPS parameters.
Notably, compared to 2014, Adelaide and Port Adelaide are slightly down-rated, since part of their 2014 performance is now attributed to the larger-than-previously estimated HGA at the Adelaide Oval. They'll still be assessed as strong teams when playing at home on account of the high HGAs now assessed for them, but they'll be assessed as less potent when playing away.
Conversely, Richmond, West Coast and Essendon are slightly up-rated since more of their apparent underperformance in 2014 is being attributed by the new ChiPS to their unusual, negative HGA effects.
All that said, the correlation between the end-of-season Team Ratings from the 2014 and the 2015 ChiPS Ratings Systems is +0.98, so it's not as if the two Systems don't share a common set of core beliefs about the rough abilities of all the teams.
For Margin Prediction purposes, ChiPS has eight other parameters, five of which relate to how it weights results from different portions of the season when those results are used to update Ratings. The summary of the changes in these parameters is that under- or over-performance in games from the first six rounds of the season will tend to move ChiPS Ratings more this year than they did last year, as will results in Rounds 12 to 17, and in the Finals. Results in Rounds 7 to 11 will have about the same effect as they did last year, while those from Rounds 18 to 23 will have slightly less effect.
The Carryover parameter, which determines how much of a team's previous-season Rating it carries into the new season, has increased under the new ChiPS, up from about 60% to almost 80% (note though that the larger early-season result weightings will make for more rapid adjustment to those carried over Ratings). Also, the Cap parameter, which determines the maximum victory margin recognised by ChiPS when making Rating adjustments, remains unchanged at 200 points - effectively meaning that there is no cap.
Finally, the Delta Form parameter, which determines how much influence the difference in the competing teams' recent form (as measured by the change in their ChiPS Ratings over the two most recent games in the current season), has reduced from about +0.65 to +0.44.
One additional parameter is required to convert ChiPS Margin Predictions into Probability Predictions: the standard deviation of the zero-mean Normal distribution from which the final adjusted game margin is assumed to be drawn. The previous ChiPS used a standard deviation of about 34 points; the new one uses 36.5 points.
I notice that I've a few new e-mail subscribers this year. So, firstly, welcome to you all. During the course of the season you should expect to receive a few e-mails each week as new posts go up on various portions of the site. There's a fairly regular publication schedule here on MoS:
- Wednesday night: posting of the week's wagers and tips on the Wagers and Tips blog.
- Sunday night: posting of the week's wagering and tipping results accompanied by appropriate gnashing and wailing (though MoS does have a "no excuses" policy that's mostly enforced). That's also posted here in the Wagers and Tips blog.
- Sunday or Monday night: posting of the updated Team Dashboard over on the Team Dashboard section of the site.
- Monday or Tuesday night: posting of the updated ChiPS and MARS Team Ratings with accompanying commentary here in the Wagers and Tips blog.
Publication of these blog posts will sometime need to time-shift to accommodate differences in the scheduling for particular rounds or delays in the posting of relevant wagering markets. Very occasionally, life will otherwise intervene and cause publication delays, but these have been rare.
In addition to these regular posts I'll also occasionally post to the Statistical Analyses portion of the site on, as the name suggests, topics related to the statistical analysis of AFL results. Posts there tend to more technical in nature (I'm a Statistician/Data Scientist by day, with all the attendant jargon risks and inability to communicate complex ideas clearly that this entails), but I try to ensure that the practical implications of these analyses are plainly laid out in a non-technical way. Comments and suggestions are, genuinely, always welcomed. You can e-mail me or leave a comment on the relevant posting.
Finally, as the Finals draw near (and occasionally otherwise) I'll start simulating the remainder of the season and estimating teams' Finals prospects over on the Simulations section.
I'm looking forward to the season ahead and to your comments and feedback. Here's to a good one.