Baseball Prediction using Machine Learning - Predicting Runs Scored: First Model

This is the eleventh video in a series where we will attempt to predict the winning probabilities for (MLB) Major League Baseball games using modern machine learning techniques. In this video, we shift our focus to predicting the number of runs scored. This requires us to wrangle our data into a new format where we have one teams hitting features, the opposite teams pitching features and try to predict the number of runs scored. We want to predict the distribution of runs scored, not just a point estimate, so that we can use these models to get probabilities for the over/under. As such we will employ a probabilistic regression technique called Coarsage (available in the StructureBoost package) to model these predicted conditional distributions.
Data: www.retrosheet.org, www.oddsshark.com
Notebooks: github.com/num...
github.com/num...
Personal links:
Consulting: www.numeristical.com
Github: github.com/num...

[@tshock22]

Another great vid. Thanks!

[@numeristical]

You bet!

[@prophecysports]

i built something similar and i'm getting similar results as you... the majority of the probability mass is centered around 2 runs so when trying to produce an actual prediction taking the class with the highest probability yields 2 runs for most games. not ideal... were you able to improve this? curious to see how others are tackling this problem. would love an invite to the discord!

[@numeristical]

Here’s an invite - will be easier to engage in discussion on the discord! discord.gg/2pTX472x

[@malone1020]

in 11b: preds_test = cr1.predict_distributions(X_test) I get: ValueError: Buffer dtype mismatch, expected 'long' but got 'long long'

[@malone1020]

maybe a linux vs. windows thing? long for linux, and long long for windows? do you use linux?

[@numeristical]

yeah - I don’t get this error but others have reported it. I can fix it (meant to do it with the latest release but forgot). Give me a couple days and I can upgrade the package to fix it

[@numeristical]

@Brian Malone This should (hopefully) be fixed now. Please upgrade to structureboost 0.4.1 and let me know if that resolves the problem

[@malone1020]

@@numeristical upgraded to 0.4.1 but still getting the long long vs. long error. appreciate you putting in the time to investigate and correct!

[@numeristical]

@Brian Malone - ok, give it a shot now - structureboost 0.4.2. If you still get the error, try to email me the error message so i can diagnose a bit better. Fingers crossed!