Machine Learning and Tactical Asset Allocation – Part II: Decomposing the Machine Learning Signals

Highlights:

• In part I of this Core Matter we showed that integrating machine learning signals as an additional input layer into an already existing and proven tactical asset allocation process added value.
• So far, we mechanically applied the signals according to the chosen overlay strategy, thus accepting the ‘black box’ machine learning is widely associated which. That said, there is scope to shed light into that box. 
• The algorithm that is used in our model is called: k-Nearest-Neighbours. Based on macroeconomic timeseries data, our signals are designed as a weighted majority vote amongst the representatives in the neighbourhood of the current query point i.e., the algorithms identifies periods in the past that are most similar to the current one. We can take this information about the neighbourhood from the model to contrast it with our own intuition or put it into a historical perspective by comparing it to past forecasts.
• The darker part of the box is the identification of the drivers behind the model’s choice. It is not directly related to the k-Nearest-Neighbour algorithm itself but to one step in the data pre-processing. The model’s choice is not directly based on the initial data but on a reduced number of (Kernel-based) principal components without knowing anything about their composition. We “unmask” the so-called Kernel-Trick and are thus able to map the model’s choice back to the initial data.
• There is always a need to critically review pure model  results. Hence, with the knowledge of the structure of the neighbourhood and the main drivers that led to its choice we offer a complete set of analytics allowing for a human valuation of the machine learning signals.

Download the full publication below