Models combinations My goal is time series forecasting. I have created a number of models to make predictions. I know that forecast quality can be improved by combining predictions from different models(linear combinations of forecasts, stacking etc) So the question is : How to  determine how  should i subset and combine my common models?
Are there some common or recommended approaches? Or it is mostly  question of intuition, lack and large number of attempts?   
 A: There are many ways to do what you want. The simplest is to take the average of all individual forecasts (empirically, it doesn't showed to be worse than theoretically optimal combinations). Another way is to make in-sample forecasts:


*

*Split you data (from time 1 to T) into estimation period (from 1 to M) and validation period (from M+1 to T);

*Use rolling windows or expanding windows schemes to produce forecasts on the validation period, and thus producing T-M forecast errors $y_{t}-y^{(forecast)}_{t}$.

*Use this forecast error series to calculate some relevant measure (for instance, Root Mean Squared Forecast Error or Mean Absolute Error). You can construct weights in such a way that best individual models receive more weight. For example,
$weight_{k}=RMSE^{-1}(model_{k})/ \sum_{i} RMSE^{-1}(model_{i})$.


There are more elaborated techniques such as Dynamic Model Averaging or Selection that will do what you want. The estimation can be tricky if you are not familiar with bayesian inference though. Here the weights are the predictive probability of an individual model.
It should be mentioned that the Model Confidence Set is also used in several applied studies.
I hope it helps your studies!
A: Simple average works best. It's been shown empirically, but you can see it yourself. This phenomenon even has a name: forecast combination puzzle. 
All the weight optimization schemes are a waste of time.
Here is one recent paper explaining this puzzle: Gerda Claeskens, Jan R. Magnus, Andrey L. Vasnev, Wendun Wang, The forecast combination puzzle: A simple theoretical explanation, In International Journal of Forecasting, Volume 32, Issue 3, 2016, Pages 754-762, ISSN 0169-2070, 
A: 
I know that forecast quality can be improved by combining predictions from different models

I suggest you review the concept of bias variance trade off, and understand currently you are under-fitting or over-fitting. 
Note that, average models will increase bias and decrease variance. It will improve the performance when you are over-fitting data.
