What is a Virtual High Volume Low Margin Model?
Odds Modelers are concerned with two global states of data:
- Market State: The volume and distribution of transactions across the marketplace; and
- Game State: The player data and immutable historical statistics that represent and are related to the current game matchup.
In effect the job of an oddsmaker is to learn some function:
F:(market data, game data) => [awaywp, homewp]
that takes some sliver of the global state data to a prediction of an outcome. The bettor then weights their prediction with their stake and believes that they have a few cents of edge on the bookmaker.
From the perspective of the book maker, the bettor is the market data:
The idea is to map weighted data (stake) to a histogram and measure the confidence that the statistical distribution of the sample set represents the distribution of betting appetites at larger volumes.
The bookmaker responds to Market State changes by converging the price to the sample mean with each additional bet and then raising limits when its variance is below an in-house calculated threshold. If the Game State changes then the market is reset as opinions reform around a new set of probabilities; such an approach is fairly common.
These temporary inefficiencies in game and market state can be exploited by the savvy bettor.
As volume grows, so do the limits resulting in the price converging to the closing line. This proportional information model is the state of the art. However, if we abstract even further we start to understand that bettors are just irrational agents optimizing a model from partial game state data.
Bookmakers just have a bigger picture thanks to various out of reach price aggregation and data streaming companies that flood it with state data. With so much data, the bookmaker is unlikely to let a bad price hang for long or raise limits irrationally on thin-sliced data. The plethora of real-time market data is enough to keep exposure for the most part, balanced over time.
The principal to keep in mind is more data is always better, and asymmetric data is best.
Virtual High Volume Margin Model (VHVM)
Given bookmakers have near global state data, do we even have to wait for bettors to provide information? Bettors are just irrational decision makers acting on a small sliver of state data. It begs the question, can we also make large ensembles of semi-optimal decisions with neural nets?
Instead of waiting for bettors to make bets, Fansunite.bet uses Elixir to spin up tens of thousands of bet simulations and make semi-optimized predictions. Each agent is given a partial view to game state data for a match.
We swap out the real bettors and rather try to anticipate their distribution of rational price-taking around a fixed game state. In the Virtual HVLM model, the trick is to inject randomness into the ensemble of agents so that you get a reasonable spread of bias. We can then quickly spin up millions of random agents and consistently adjust them in real time as bets come in.
The High Volume Low Margin model has pioneered the way for low margin sports bettors for 20 years, but now that data is so readily available, perfect game state/market state data can be had for a hefty price.
If you have all of the data, why would you ever need a human being to set a price when you can use thousands of neural networks instead?
Stephen Rothwell
FansUnite Head Trader & Machine Learning Expert
Learn More About FansUnite:
- Join our Telegram Community
- Follow our Blog updates
- Join the Newsletter
- Contact us
