Arguments against model or forecast combination? Do you know any references providing arguments against model or forecast (models output) combination?
Could not find anything
 A: Combining models is call Ensemble learning.
In general, these methods are useful.
When you can build models that specialise in areas the previous model don't perform well, as in boosting, you can improve the performance.
When you have an unstable classifier and you modify the dataset in order to get some models, as in bagging you can also improve.
However, adding models increase the model complexity.
Therefore, it usually belongs to a larger concept class, and your assurance of learning by Occam's razor is lower.
Having a more complex models also typically leads to a higher VC dimension, hence your dataset is less likely to estimate well the underlining structure.
The best practice is to start with simple models and increase their complexity if needed. In case that the models predictions are close to each other and focus on the same areas that ensemble methods are likely not to help mush or even cause damage. 
A: Our new review discusses these questions in detail: 
Dormann, C. F., Calabrese, J. M., Guillera‐Arroita, G. , Matechou, E. , Bahn, V. , Bartoń, K. , Beale, C. M., Ciuti, S. , Elith, J. , Gerstner, K. , Guelat, J. , Keil, P. , Lahoz‐Monfort, J. J., Pollock, L. J., Reineking, B. , Roberts, D. R., Schröder, B. , Thuiller, W. , Warton, D. I., Wintle, B. A., Wood, S. N., Wüest, R. O. and Hartig, F. (2018), Model averaging in ecology: a review of Bayesian, information‐theoretic and tactical approaches for predictive inference. Ecol Monogr. Accepted Author Manuscript. doi:10.1002/ecm.1309
One of the main conclusions is that bias/variance -> error of the average is influenced by the 


*

*bias

*variance / covariance


of the contributing models, as well as the error (uncertainty) of the estimated weights for these models. 
Depending on how things play out, MA will have higher or lower error than alternatives (picking one / the best / the full / a regularized model). 
A further issue is that calculating uncertainties / CIs for averaged predictions is challenging, which can be an argument against MA if you care about that. 
A: One argument against it is if one of your models is much closer to properly identifying the time series patterns than the rest in the ensemble. Had you used that model versus the combination, it would likely have performed better. Other problems: knowing which models to include, what ensemble size and how to combine is vague. While there is evidence that the basic average works well for time series combinations, it's still not settled (my opinion). I've seen median (or trimmed-means) and linear combinations work quite well in practice.
