An ensemble of models is the equivalent to an investor having a diversified portfolio of stocks. A single stock may have a higher return, but the risk of failure is much greater than a portfolio. A portfolio of models will on average have a lower mean squared error relative to a single model. The manager of an ensemble is not wed to a single point of view or technique but can take a diverse set of model approach to forecast the same target variable.
There is the view that ensemble modeling is a machine learning technique (for example random forest models and the use of bagging, boosting, and stacking approaches to model building), but the idea of bundling is an old forecasting tool. The idea of using more than one model and weighing the output to form better forecasts predates any machine learning. Machine learning has developed methods of employing more data and better search techniques to broaden the ensemble of possible models.
Ensembles can first be used on a simple basis, bundling a limited number of models, before employing machine learning. A measured approach will allow for better understanding of forecast dynamics. Econometric classes do not focus on ensembles because the focus is on testing specific models and hypotheses and not on just forecasting performance.
For trend-followers, the ensemble approach can have more than one trend look-back period to measure the trend. The position will be related to the number of models that are providing the same signal. This is consistent with the idea of preponderance of metrics.
A macro forecaster could use different variables to measure an interest rate change. One could be focused on inflation expectations, another could use trends, while a third could be related to the interest rate curve. Different models are based on different views of the world and how markets operate. An ensemble will say that different views of the world can compete and be bundled to limit the risk from holding just one view of the world.
No comments:
Post a Comment