Hi Edson and Mher, I admit that this is not an usual way to do and maybe it is not correct either but You have to train and predict all Your data and guess these values separately in order to enhance the chance to win. When predicting with linear regression, the average of linear relationships are not enough particularly for goals prediction. From "testSet" the modified dataset should be created with separate predictions. The teams will be entered as numbers, hopefully they are identified by the Predictor. Main assumptions:

1. Time-dependency and data which is not periodical {-1,0,1} but having a sort of ramp (cumulative and using a NonliearFit consisting polys and cos/sin). Later, differences between time-based predictions will give outcomes and goals.

2. The prediction collects the team performance over the time in series and head-to-head between particular teams

3. Prediction should be targeted on next game only (and updated after every real game developing prediction models)

I will give a real example/notebook, hopefully within a week. This example shows that even having wrong predictions put in the "Outcome" through the main predictor will be relatively stable and increases the chance to predict right. above You see an individual model of data Then prediction:

The prognosis dataset can be created:

and Outcome prediction:

Which is rounded to zero as in testSet. Sorry for pictures this time but hopefully I will have time to automate the code and post it here.

Attachments:

Hi, nice post. However, I guess something's wrong. Please correct me if I got it wrong. You are training the classifier by providing the home and away goals. In addition, home goals and away goals are also present in your test set as predictor variables. So, basically you are trying to predict something that you already know.

Are you using "hgoal" and "vgoal" as predictor variables for "Outcome"?