2026 - The Year Ahead on MoS

/

MoS is going around again for what is now its 21st year of operation, and there’s been a lot going on in the off-season, much of it assisted by AI, in particular ChatGPT.

I’m very aware of the ambivalence and sometimes clear hostility that many people feel about AI, and LLMs in particular, and can understand the outrage that artists and authors feel about their works being incorporated wholesale, with neither permission nor recompense, in the training material for LLMs.

Oddly, I think, there appears to be little or no backlash about the similar absorption of extant programming code into the corpuses of ChatGPT, Claude, Gemini AI and others. There seems to be a collective agreement that such coding effort does not embody creativity, or at least not in the same way or to the same extent as does, say, a novel or a painting. This sells short the art of coding, I’d suggest, but it is what it is.

In any case, as I say, ChatGPT has been a very helpful companion for much of the new work I’m about to describe. My feeling at this stage is that AI represents a clear benefit for people at my level of coding ability, which allows me to craft prompts for coding efforts using the appropriate terminology and to intuit when the outputs of those prompts are likely in error. What those prompts create are outputs that would either take me orders of magnitude longer to produce or, in some cases, would not be possible for me to produce at all given my current knowledge.

FORECASTING MODELS

Most of the head-to-head, probability, and margin forecasting models that appeared in the blog in 2025 will again appear in 2026, with the following changes:

  • MoSHBODS has been reoptimised with parameters chosen based far more heavily on recent seasons than has been the case in years past. Because I’ve wanted there to be some recognition in MoSHBODS (and MoSSBODS) of the entire span of V/AFL history, I’ve always included every season in the optimisation process, albeit that I’ve often downweighted the past in increasing measure as it ventured off towards and beyond the turn of the 19th century. On reflection, I think that’s produced algorithms that perform adequately across the period from 1897 onwards, but that aren’t as good as they could be at what I mostly want them to do, which is to predict future games in the current era, most particularly in the current season.
    With that in mind, this year I’ve chosen parameters for MoSHBODS that performed best (ie had lowest MAE) across the ten seasons from 2015 to 2024, leaving 2025 as a holdout for estimating out-of-sample performance. MoSHBODS is therefore deaf to football played before 2015 and this shows in its MAE performance for that part of history relative to the MoSHBODS of last year.
    That said, the resultant team ratings that come from MoSHBODS 2025 and MoSHBODS 2026 are very highly correlated (over +0.99), and the increases in MAE are mostly under 1% for any given season.
    ChatGPT’s role here was to very much speed up and simplify the optimisation process although, even with its improvements, it was difficult to do more than optimise one parameter at a time.

  • MoSSBODS, after 11 seasons of service, has been retired. In 2025 MoSSBODS was optimised to predict Game Totals and still managed to do worse than MoSHBODS on that metric as well as on Game Margins.
    When you think about it, MoSSBODS is actually just a special case of MoSHBODS where the Adjustment Factor, which determines how much weight is placed on Scoring Shots versus Score, is set equal to one, so it’s perhaps no great surprise that it struggles to outcompete its more flexible sibling.
    It was a useful entree into the Elo as margin predictor space, and allowed me to explore and explain the large role than randomness plays in teams’ on-the-day conversion rates.

  • In MoSSBODS’ stead, a new forecaster, based on a revamped MARS model, will get near top-billing in 2026. It’ll simply be called MARS and, like the MARS before it, it will have an Elo-style team rating element at its heart. I’ll probably write more about it in a future blog, but for here I’ll just note that

    • it outputs team ratings each week

    • an average team is rated 1,000

    • ratings are updated based on margin outcomes relative to expectations given the teams’ ratings and familiarity with the venue at which a game is being played

    • margin forecasts are also derived from the teams’ ratings and familiarity with the venue at which a game is being played

    • a proportion of team ratings carryover from one year to the next

    • unlike MoSHBODS, a fixed ratings update parameter is used for every round

    The algorithm has a number of parameters and, again, ChatGPT helped by creating a fast script that facilitated finding optimal values for the 2015 to 2024 seasons, leaving 2025 as a holdout

  • MoSHPlay, you might recall, combines MoSHBODS’ outputs with a Player Model, the latter aiming to forecast the AFL Playing Rating that a given player will record in his next game. That Player Model was significantly renovated in the off-season, with ChatGPT creating a much faster algorithm and facilitating the calculation of optimal parameter values for that algorithm. It produces clearly superior forecasts of AFL Player Ratings.
    As was the case last year, a Multiple Adaptive Regression Spline was used to combine MoSHBODS and Player Model outputs in an historically optimal way and incorpating “hinges” that only trigger for certain ranges of MoSHBODS margin forecasts and aggregate Home and Away team forecast AFL Player Ratings.

  • Lastly, ENS_Linear was rebuilt this year to account for the new MARS ratings. It remains an ensemble of five base learners each of which take as inputs MARS team ratings, bookmaker pricing and handicaps, venue familiarity and location, recent team form, and whether a game is part of the home and away or the finals part of the season.
    It too has a number of parameters and, yet again, ChatGPT helped by creating a rapid script that facilitated finding optimal values for the 2015 to 2024 seasons leaving the 2025 season to be the holdout comparator

Turning then to the performance of these models on that holdout year, as you can see in the table at right:

  • the 2026 versions of MoSHBODS and MoSHPlay produce superior Margin MAEs to their predecessors for the 2025 season

  • the 2026 ENS_Linear model is slightly worse than the 2025 version on Margin MAE, although it is still highly competitive

  • the MARS Model had no 2025 version, but the 2026 version would have done quite well in 2025 on margin forecasting

  • the 2026 MoSHBODS and MoSHPlay models produce superior LPSs compared to their 2025 equivalents

  • the 2026 MARS model lags a little on that LPS metric, but still does better than both the 2025 MoSHBODS and 2025 MoSHPlay models

WAGERING IN 2026

The men’s 2025 season ultimately proved to be a profitable one, but only after a lengthy mid-to-late season sequence of losses, the exact source of which I can’t pin down.

In any case, buoyed by the fact that the final result was profit, in 2026 I plan to again wager in the head-to-head, line, and total markets, but with MoSHBODS informing all of those wagers.

A review of how it would notionally have performed in 2025 has led me to alter the wagering thresholds as per the table at right.

In the head-to-head market we’ll now be looking for an estimated edge of at least 5% before considering a wager on either a home or an away team, and we’ll be a little more risk-taking when a home team wager seems especially lucrative, but a little less risk-taking when an away team wager looks equally inviting. I still vacillate about the wisdom of head-to-head wagering on away teams, and worry about the highly variable nature of the meaning of “away team” in some games, but a sensible, practical solution escapes me. At some point I do think I’ll come up with a measure of “homeness” and “awayness” for a given game, and use that to further refine the thresholds.

In the line market the only change is a lowering of the threshold on the minimum estimated edge requirement for a wager on the away team.

Lastly, in the totals market, we’re now happier to run with any wager that is estimated to have a non-zero edge but we’ve significantly capped our appetite for wagers that seem to have overly positive EV.

The weights of the three Funds will remain:

  • 65% Line Fund

  • 25% Head-to-Head Fund

  • 10% Over/Under Fund

Wagers, when made, will continue to be sized as one-fifth Kelly bets based on the original Fund size, not the current.

TEAM DASHBOARDS

These will continue to be produced in 2026 using exactly the same format as last season. If you have any requests for new inclusions, left me know.

TEAM RATINGS

There will be a weekly blog post updating the latest team ratings of MoSHBODS and MARS, with MARS being less of an afterthought and more prominent this season and providing an interesting alternative to MoSHBODS.

SIMULATIONS

There will also be a weekly blog post summarising the results of multiple simulations of the remaining head-to-head season, as well as the Finals including the Wildcard Participation Appreciation Round. ChatGPT has also created a significantly faster R script for producing these simulations (about 5,000 an hour), so there might be the opportunity, at least occasionally, to do more than the usual 10,000 replications.

AFLW

MoS will also be providing forecasts and wagering updates for the 2026 AFLW season and, at some point, look at reoptimising the models for those tasks.

That’s it for now. Please let me know if you have any questions, and best wishes for the men’s and women’s seasons ahead.