# Forecast combination approaches or: How to beat the simple average?

Say, i have a set of two forecasters (or 'experts'), who provide a forecast for wind power time series and i have to pick the best expert (or a somehow-weighted combination of both experts) to have an overall better forecast. Here, i can do a simple uniformly weighted average, which often beats many other forecast combination methods. (Diebold 2017, p.405ff)

Other elaborated approaches were suggested in the literature, most prominently variance-covariance based methods originally developed by Bates & Cranger,1969. Especially in the field of ensemble learning, there are algorithms to robustly combine experts (Littestone & Warmuth, 1994). Further developments of picking the best expert advise based on its prior performance via exponentially weighted aggregation (see Cesa-Bianchi & Lugosi 2006, p.7ff) also exist. Especially the last method is often used in online machine learning literature, where it is opposed to batch learning alghorithms to overcome computational limitations. (if you are interested: Opera Package which implements the online learning in R)

Now to my problem: Recently, i have gained the opportunity (for my master thesis) to access an additional set of 7 weather experts who provide meterological forecast data like wind speed, direction, temperature and air pressure. My previous mentioned set of 2 experts who provide wind power forecasts also rely on meteorological experts and probably they use the same weather data as an input for their predictions, but i am not sure and i don't have insight to their models. They are black boxes to me.

In my understanding i have a univariate regression problem, because i want to predict a future value based on a different set of experts; in the next step I would like to create a model which uses the wind power predictions of my 2 wind power experts and additionaly considers the meteorological data of 7 weather experts.

I have read a lot of literature (mostly journal articles) on that combination topic but cannot yet decide how to combine my initial 2 forecasts with additional meteorological data. Obviously as a first step because of different units, i need to standardize, which is a topic on its own - still straightforward. But the next steps are a bit confusing, because i have so many approaches to build my regression model, especially considered in the broad field of statistics.

In the field of computer science, Recurrent Neural networks - especially Long Short Term Memory for time series forecasting - seems to be a promising research avenue nowadays. I have to deep dive further in that topic to assess whether they fit to my problem or not.

Any recommendation for suitable methods is highly appreciated!