I Upgraded The Betting Algorithm... Again

Exploiting non-implicit correlations to squeeze out an edge over the sportsbook.

In the last update to our algorithm, we ran into the scenario where our model was essentially the exact same as the sportsbook’s, so we had a hard time finding major opportunities for mispricings.

Well, to solve this problem, we enlisted some help from our friends in the finance department.

Non-Implicit Correlations

In finance, there are many different forms of correlation (e.g., implied, auto), but the two main ones are implicit and non-implicit correlations. Let’s take a glance at the differences between the two:

• Implicit Correlation

  • These are your no-brainer, obvious correlations. If the price of Coca-Cola rises, it’s likely that the price movement in Pepsi will be correlated and also rise. A news report of Ford having to recall 1 million vehicles will be strongly correlated to the stock price moments later.

• Non-Implicit Correlations

  • This area is where the correlations become murky. If Boston Scientific receives a $47 million grant to manufacture a product that needs a rare element, the stocks of miners of the rare element are likely to rise. If there is an uncertain macro environment that may influence corporate earnings, then volatility around earnings are likely to rise.

In theory, non-implicit correlations are less likely to be priced in, and thus, leaves room for opportunity.

Let’s see if we can factor that into our algorithm:

But first, a not-so-shameless plug:

• If you’re interested in deploying the full, end-to-end system yourself, I formalized all of the progress into one comprehensive format. Each step is covered, from setting up the database to calculating the underlying data and training the models, all the way up to forward testing strategies before live action, leaving no stone unturned.

  • Machine Learning for Sports Betting: MLB Edition

If They’re a Hitter, What Else Might They Be?

To first identify any non-implicit correlations, the first place to look was for any overlaps in model predictions. The goal was to see if a player is say, likely to record a hit today, what else are they likely to do (or not do)?: