I've created a random forest model which predicts if a team will win a match (assuming no draws), and returns a probability of that occurring.

The features that are trained on include various statistics, such as a team's average win rate over the last 4 years coming into the match, the name of the opposition, and the final ranking of that team in a previous tournament. We also include the same statistics for the opposition (for example, opposition's average 4 year win rate). We then use these features to learn on a column which shows if a team won or lost that match.

Since we have a limited number of games, it was decided to double the number of games by also printing the statistics from the away team's perspective. Therefore, each match appears twice, once from the perspective of the home team, and once from the perspective of the away team. We therefore introduce a home flag. Therefore, the home flag will be True from the perspective of the home team, and will be False for the game from the perspective of the away team.

We then choose 20 features using CARET, and train a forest using those features.

Having trained the random forest, we now use it to predict match results. We only care about the predictions concerning the home team, and filter by home flag is True. However, analytically speaking, would there be any merit to also use the game from the away team's perspective (ie: home flag is False), and then blending the two probabilities, ie: $$\mathbb P(FinalHomeWin) = {(\mathbb P(HomeWin) + (1 - \mathbb P(AwayWin))) \over 2}$$

Therefore the questions are: is this the best approach, and would blending the probabilities improve accuracy?

To give an example: imagine England is playing Japan at home (ie: in England). We have trained our model on previous games where the home team flag is both True and False. However, when we predict this match (which hasn't been played), we only consider the prediction on the probability that England will win, and ignore the probability that Japan will win.

  • $\begingroup$ Two comments: 1) in my experience there is a home-field advantage, but it isn't where folks think it is, and sometimes it is a disadvantage. Account for it, but let the data tell you what the coefficients are. The P(final_home_win) should be a function of both P(homewin) and P(Awaywin) but let the data drive the relationship, by both hometeam and away. 2) GBM are better for outliers and random-forests are great for robust fit. All else being equal - the RF will tell you how to go, but the real world isn't equal. Consider a well build and heavily validated GBM. $\endgroup$ – EngrStudent Jan 11 '17 at 18:33
  • 1
    $\begingroup$ This is very helpful, thank you. There is no assumption in our model as to whether home is an advantage or disadvantage, we just treat it as a factor to consider when training; I would presume that the tree structures would be able to learn if this is an advantage/disadvantage for each team? Good point about the driving the probability blending as a function of the data, this gives me something tangible to pursue. $\endgroup$ – fiorenza2 Jan 13 '17 at 9:55

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.