2023 - Round 12 : Overs/Unders
Apologies that this post is a little late this week but I have been travelling on business (and. frankly, if you’ve been waiting to see the models’ views before making your own bets, you’ve not been following along carefully enough).
Overall across this truncated round, the MoS twins have roughly the same expectations about likely average total scoring in Round 12, although there is, again, some variability of opinion about individual games.
As highests and lowests, we have:
HIGHEST SCORING GAME
MoS twins: Essendon v North Melbourne
Bookmakers: West Coast v Collingwood
LOWEST SCORING GAME
MoS twins: Melbourne v Carlton
Bookmakers: Gold Coast v Adelaide
HIGHEST SCORING TEAM
All: Collingwood
LOWEST SCORING TEAM
All: West Coast
WAGERS
Investors have two wagers this week, one under and one over.
The estimated overlays for the games where we’ve a wager are about 7 and 13 points.
PREVIOUS RESULTS
Honours for the lowest Mean Absolute Errors (MAEs) this week were again spread around, with MoSHBODS taking Game Margin and Away Team Scores, and the TAB taking Game Totals and sharing Home Team Scores with Sportsbet.
That left all four titles in the same hands as last week, and the leads as follows:
Game Margins: TAB 5 points ahead of MoSHBODS and 31 points ahead of MoSSBODS
Home Team Scores: TAB 8.5 points ahead of Sportsbet, and 82 points ahead of MoSHBODS
Away Team Scores: MoSHBODS 11 points ahead of MoSSBODS, and 28 points ahead of the TAB
Game Totals: Sportsbet 3 points ahead of the TAB, and 35 points ahead of MoSHBODS
MoSSBODS had three wagers last week, one of each kind with the TAB, and a lone under with Sportsbet. It won the under with the TAB and lost the over, and it lost the under with Sportsbet to leave it now at 14 and 15 for the season.
Across all games, MoSSBODS landed 6 from 9 against both bookmakers, and MoSHBODS landed 5 from 9.
That moved MoSSBODS to a 49 from 99 (50%) record against the TAB, and a 47 from 99 (48% ) record against Sportsbet. MoSHBODS now has a 52 from 99 (53%) record against both bookmakers..
ACCURACY by estimated overlay
The idea of calculating an overlay (the difference between a model’s forecast and the Total on offer) and only betting when its absolute size exceeds some threshold is based on the implicit assumption that a true wagering edge is more likely to exist when a model’s forecast is very different from the Total being offered.
The one practical adjustment we make to this is the case where the estimated overlay is large, the recommended wager is an over, and rain is forecast. We do this because we know that, in this instance, there is important information about which the model has no knowledge.
What, then, are the relationships between overlay and accuracy for the two MoS models against each of the bookmakers?
For the results in the top half of the table we’ve deemed a large estimated overlay to be one that is, in absolute terms, 4.5 points or larger, and we see, for example, that MoSSBODS has registered a large negative overlay relative to the TAB’s offered Total (hence recommending an unders wager) on 24 occasions. In 13 of those games the actual Total did come in under the TAB line, giving MoSSBODS a 54% record for those wagers.
(Note that I’ve chosen the 4.5 cutoff here purely to put roughly equal numbers of games in each bucket)
Looking across the entirety of that table we see that the best results actually come in those games where either model records a small positive overlay against the TAB, and when MoSHBODS does the same against Sportsbet. MoSHBODS also performs relative well when it records a small negative overlay against the TAB’s line, although the sample is small.
In the bottom half of the table I’ve used, instead of 4.5 pointsd, a cutoff of 6 points to align with the threshold that we’ve been using for under/over wagering.
There we see that MoSSBODS does best with small positive overlays against the TAB and large positive overlays against Sportsbet (although with only a small sample). MoSHBODS now does well with small positive or small negative overlays against either bookmaker.
What to make of all this? Well, firstly, we should recognise that the sample sizes are generally quite small in each cell, especially so in some cases, so drawing any firm conclusions is ill-advised. Nonetheless it does look at though that small overlay result for MoSHBODS might be real given that it is based on 60-odd games.
If MoSHBODS is particularly well-calibrated for small estimated overlays, how might we explain the relatively poorer outcomes for it when it registers larger estimated overlays? That’s relatively easy for the large positive overlays in that we can attribute it to the effects of rain, although we’d need to link rain data to each game in order to bolster (or otherwise) that assertion. We could also postulate that any sufficiently large estimated overlay - positive or negative - was actually a sign that the model was blind to something important about the match, and that the magnitude of the estimated overlay should serve as a warning to us. Lastly, we could also point out that the Large Under results for both bookmakers are based on sample sizes of only 17 games.
That argument still isn’t entirely convincing, though, especially as it’s only come after we’ve observed the apparent anomaly. The best test will be to keep an eye on the performance of the MoSHBODS model in each of these categories across the remainder of the season and to observe the extent to which we see regression to the mean.