0

So I am working on a project where I want to forecast how a team will perform in their next match in a number of specific categories (goals scored, time spent in certain parts of the field, passes, ball control, etc.). This would then go up against another team's forecasted performance in a prediction model to simulate the outcome of a match between the two teams.

I have data on all of the team's previous matches but am unsure of which path to take. I have looked into VAR, ARIMAX, and Random Forests but because there is no resource that really compares them at the same time I am getting a bit confused. What is the benefit of using a machine learning model vs a time series model? what questions should I be asking to figure out which path to take? Thanks for any help!

Example of how data is stored at the moment:

|----------|----------|...|----------|
|  Date    |   Goals  |   |Time in _ |
|----------|----------|...|----------|
| 2011-1-1 |    2     |...|   20     |
|----------|----------|...|----------|
JJ Stamp
  • 13
  • 4
  • 1
    Could you indicate the sense in which you believe there is a "time series" involved? Please see https://stats.stackexchange.com/questions/142593 and https://stats.stackexchange.com/questions/126791. What might you mean by "time series model"? – whuber Nov 14 '18 at 19:44
  • I am pretty convinced that however I forecast this data lag will have to play a part because teams usually thrive on momentum while they play. A poor performance could lead to more poor performance and vice versa. I meant a time series model as models that are usually applied to time series exclusively which would exclude random forest. – JJ Stamp Nov 14 '18 at 21:00
  • 1
    Unless you have actual time series data--and that means, among other things, that you have sequences of data obtained at regular time intervals--you cannot even hope to apply standard time series methods for the analysis. This is why it will help you to describe your data in more detail. – whuber Nov 14 '18 at 23:20
  • While the data is not at regular intervals, there is time between seasons, the data is obtained over time. If I'm understanding the links you posted above the main constraint is that the data has to evolve with time. Is there a part of my past explainations that makes that ambiguous with relation to my dataset? If so I apologize but having an example table above that shows the time aspect and mentioning the importance of lags in my analysis in the past comment I'm confused on where I'm not explaining the passage of time within my dataset. – JJ Stamp Nov 14 '18 at 23:46
  • 1
    What you seem not to have gleaned from the links is that almost all time series procedures--especially ARIMAX--rely on there being equal intervals between successive times of observations. "Data is obtained over time" does not otherwise constitute a time series and therefore usually requires different approaches such as mixed models. That leaves one wondering what you might mean by "time series model" in your question. – whuber Nov 15 '18 at 00:28
  • I think I understand it better now. What I had meant by time series model was things like AR, MA, and ARIMA models. I however was not aware of the equal intervals constraint so I can see where my problem doesn't fit into those models. I'm sorry for the lack of understanding on my part. That being said does that mean my only options come in the form of mixed/ensemble models? If that's the case does that make random forest the best choice? – JJ Stamp Nov 15 '18 at 00:51
  • You might look into methods of *longitudinal data analysis,* among other possibilities. – whuber Nov 15 '18 at 13:07

0 Answers0