Forecasting model for small dataset with many parameters? I have a dataset of 20 columns/parameters (x's) and not many rows (10-20 historical values for each x) I need to use to predict a y column for the future 5 years. Each row represents one year. What can I do with such little data? Is a VAR model the way to go? Is it an issue that some of my parameters have more data than others?
Any direction welcome, I'm a newbie.
Thanks
 A: If the other predictors presumably influence your $y$, but not vice versa, then a lasso might be useful. The problem then is that you would need to forecast your predictors themselves.
If your $y$ also influences the other columns, then a VAR might be indicated, ideally with some kind of regularization. This automatically forecasts the entire system, but of course you get nothing for free, so such a model would really need more data than the lasso approach.
Missing data are of course a problem. You can try to fill these in (but be aware this will simulate knowledge to the model that you really do not have, so you will be too certain of your conclusions), or delete rows with missing data, which will pose problems for estimating the time series dynamics and potentially leave you with too little data.
In any case, with so little data your forecasting accuracy will be limited. I would always also fit a pure time series model to your $y$. It may well be surprisingly competitive with a model that uses predictors.
Related: Best method for short time-series
