AFL Mullumby Defender

Hmm after weekly manually calculating the MAE of all the other Squiggle Models  I have decided to use the 2017 data to predict early 2018 rounds - we will see how well this works.

Well we only had one bet, and the Cameron KO in the 1st quarter of the highest ranked defender (who happened to be a full back) probably swayed the result. We lost by 1 goal and the restructure caused to Brisbane has reignited the red card debate. So, for those still on the train I have indicated the two games that made the cut, until reliability was applied (in blue) as well as two selected games for round 15.

https://www.dropbox.com/s/25uv6h3dmvlprxo/rounds%207%20to%2013%20Mullumby%20Defender.xlsx?dl=0 Dropbox link to excel

After the numerical analysis, I have been directed to "Kelly Criterion" which corroborates the work I did AND provides the theoretical basis to examine past Mullumby Defender Data to work out the best betting strategy for the future. http://www.elem.com/~btilly/kelly-criterion/ Reliability 0% means not using the 2018 developed reliability (at respectively winning and losing selected bets), so MORE bets, but lower % won. The column is filled with the ACTUAL percentage of the selected bets won that season. next column is number of games selected in the season Proportion is the "Kelly recommended" proportion to bet. Return over bets is calculated return for proportion, probability and number of games. Reliability 35% is the statistics with bets of reliability less than 35% discarded Reliability (% wins when selected for the winner * % losses when selected as loser for the loser) 40% is the Kelly recommendation for the reliability 35 % method, but in...

Well, Tables first Very much middle of an impressive field, of course, Defender focusses on detecting bets it performs best with. So how do the models like the line bets on offer? Interesting that ALL the models like the single defender selected bet for the round, and that there is disagreement on every other line bet. Also, interesting that for the first time Defender is 1 out on calling Melbourne to win.

Worth noting that both Collingwood and Hawthorn "would" be bets except that Mullumby Defender automatically rejects all bets with lines less than -44 (but gee I like them). No bet will be placed on them.

The wandering drunkard is the base problem for a LOT of Statistical Physics. Have a read https://medium.com/i-math/the-drunkards-walk-explained-48a0205d304 Essentially the mathematics of a "item" with a probability of stepping right (winning) or stepping left (losing), then repeating forever. So if we have a succession of bets lined up that we "know" the probability of winning (say 0.6) then what % of the betting pool should each bet be to maximise profit? If you bet 100% of the pool each time, very quickly you have no money at all. Conversely, if you bet 0.000001% of the pool each time, well your pool will barely change size. So for a given probability of winning, what is the best fractional bet for "the next step"? A numerical analysis solution (over a thousand simulated bets, starting with $100) yields if p=0.7 win    fraction bet = 68% and above => $0 return over time    Otimal bet fraction  = 35%   return (2x10^35) if ...

13 of 14 ----- I can only say Goooooo Cats today

As you know Mullumby Defender incorporates a strength and a defence parameter for each team. I have added a mechanism whereby more recent results are more "valuable" in the calculations that generate these parameters - weighted averaging. Long story short. Defender is now measurably "better" in terms of MAE, and modelled betting. Also closer to bookies line, rightmost column.

Hey All Long story short - 7 of 10 models collated at https://squiggle.com.au/ fall in agreement on the two selected Mullumby Defender Line bets this weekend (yellow highlighted)

Good Day All some musings on a major change to Mullumby Defender, not to the calculation of the predicted scoreline and margin, but rather to the confidence in it from game to game. Currently, the model assumes the true "home margin" result will lie in a normal distribution around the predicted value with a standard deviation equal to the MAE of the season. Reasonable as a overall tactic, but what about doing better for individual games? For the Mullumby Defender selected games, the success rate for line bets (50% bookie rated) prior to reliability is around 60-65%. After adding the "reliability" criterion which relates the reliability of each team in the past to be successful when selected in a game, the success rate.... well we are on 11 wins from 13 bets in 2018, and the historical modelling puts it around 75-85%. This means apples are not apples, some predictions are better than others. Further evidence is found in the Mean Absolute Error for the relia...

Some requests for actual predicted scores have been made to me - so

Note Gold Coast selection is removed because of poor reliability at winning selected games Gold Coast win 30% of games selected by Defender to win St Kilda lose 89% of the games Defender selects them to lose Product =  27% which is less than the optimised parameter of 35% - so no bet

Well, Defender put up one bet, which it won. But overall competition with other models. MAE from round 7 and money won on $10 line bets and the league table Matter of Stats 28.2 Graft 28.2 Ranking A 28.4 Squiggle 28.5 Plus six 1 28.5 HPN 28.6 The Arc 29.0 Mullumby Defender 30.1 Stattraction 30.5 Swinburne 30.6 Massey Ratings 30.6 FMI 30.7

Go Freo Fremantle 71 to Adelaide 68 Well, that's good news. We will see how Mullumby Defender is tracking in the lowest MAE model stakes after round 12, but the more important(?) statistic is that for 2018 Mullumby Defender is now at 11 line wins from 13 games. Remember that line bets are what bookies say are 50:50 so getting 85% right is actually just stupid good (except I hope it is just SMART GOOD :) )

Hi I have started applying my Defender Model this year (optimised parameters over last 4 seasons). Essentially each team is allocated a strength in points and a Defence %. The predicted scoreline is Team A strength modded by TeamB Defence and vice versa. Add in blowout adjustments for both strength calculations and predicted scorelines and individual home state bonuses and you have my model. Based on previous season analysis Defender also self-selects the games it is more likely to win line bets, typically 1-2 per round. This is from analysis of an assumed normal distribution around the Defender margin with a standard deviation of the MAE, PLUS a reliability score based on how "winning" that team has been in previous selected games (and "losing" for the selected loser). The reason I am emailing you however is that for the games defender self-selected HPN scored only 4 wins from 12 games. However for the games NOT selected by Defender HPN scored 23 wins from 3...

Well, given the 11 other models at Squiggle I decided to see how THEY went picking Line bets. I tabulated success over the 44 games in rounds 7 to 11, and also just for the 12 selected games by Defender. While defender made $1.60 (all $10 bets) if all games were bet on, and 5 models went backwards HPN is the standout, making $78 over the 5 rounds. Of course over the 12 Defender selected games:      1)     Defender made $72      2)     HPN actually loses $43 WTF When Defender is right HPN is wrong and vice versa. I decided to look at HPN on all games EXCEPT the games Defender self selected - WOW HPN makes $121 over 23 $10 line bets. While I can only fine tune and fully understand defender, I may be able to make use of HPN, plus if the trend continues over time the combination of the models may be greater than either.

Round 12, Mullumby Defender AND other models collated at   squiggle.com   - Defender selected 4 bets- three cancelled on the reliability of teams left with Freo Line bet. However no Fyfe, so perhaps no bet. Interesting that ALL the models bar 1 have Fremantle winning that line bet   :) Yellow is selected bet by Defender Orange - bets initially selected by the algorithm, but discarded based on "reliability" of the particular teams in question. zzzzz

Of course, comparing Mullumby Defender on ALL games kind of defeats its whole purpose. Defender algorithms select games for line betting automatically with no human input. So how does Defender compare to the other models on the 12 games it has selected for bets out of 44 games over rounds 7 to 11? Looking good, wouldn't you say?

I have been directed to look at squiggle.com and it has many other AFL models collated including the famous Swinburne University Tipping. Given the years of research in Swinburne and Massey (I know little of the others) I guess 8th place on the rounds since 7 (when Defender starts each season) is respectable. (ranked on Mean Average Error)

Good Day All I have discovered that Bet Fair have a paid service that provides THEIR analysis and predicted margin for games. I am able to compare with Mullumby Defender because they release the past rounds predictions https://www.betfair.com.au/hub/news/afl-predictions-2018/ So, how does Mullumby Defender compare to the paid service? Lets look at Mean Absolute error over rounds 7 to 11 for ALL games So, actually, Bet Fair does better by a point per game, but DOES IT? Lets just look at the games Mullumby Defender says to place money on, the SELECTED Bets. green were successful line bets, red losing bets. GEE WHITACKERS! Mullumby Defender beats Bet Fair by 5.1 points per game!

The West Coast at 3/4 time were 48 points up, on target for a 64 point win. Mullumby Defender had them on for a 74 point win which is pretty close. The winning team WCE pulled their boots off the Saints throats and coasted to win by only 13 points. This has inspired a new constraint "minimum line" The round so far: Out of 7 GAMES 3 TIPPED WITHIN 4 POINTS So 2018 rounds under the "new" system Still hanging on Adelaide, though looking to be a loss. So total after 5 rounds of betting $245 I will actually reset to around $147 because of the West coast and not having the >-44 points in the system