Arima-Garch out of the Lab, into Trading
The other problem is once we run a set of experiments it could lead to other experiments based on our results from this first pass to create robust models. Using past experimentation and statistical and machine learning, we can cut this search space greatly, but this requires a lot of research and expertise.
The number of experiments and cost of them shows the value of experience in using this algorithm to predict the markets. Knowing what parameter space to look at and what regressors to use is invaluable in this process. Large institutions that are using this technology won’t share this. Studying one market is not so bad but, baskets of stocks or commodities can be time-consuming. Also, the interaction of using external regressors is not as easy as you think. How well an intermarket works as a regressor is not always correlated on how well that intermarket worked in an intermarket divergence model.
In the next segment of this story on our Arima-Garch hybrid model, we will report hypothetical results on our model trading the S&P 500.