Highlights:

• So far, we have been focussing on traditional econometric approaches for modelling economically sound relationships. In contrast, machine learning (ML) employs algorithms to learn patterns directly from data. It can handle large datasets and complex relationships. As such, ML is a promising and flexible supplement to our existing models.
• In a traditional insurance asset management company, the specialists’ expertise is the central pillar of the allocation decisions. We here develop a concept to complement the existing human expertise with ML signals to improve the results of our processes.
• We generate ML signals that forecast the upcoming market regime simply defined as equities outperforming government bonds or vice versa. The signals are generated by an algorithm that systematically searches a large macroeconomic database for comparable situations in the past. These signals amend our existing tactical asset allocation process as an additional input layer. They are used to adjust the overall depth of the initially recommended active positioning, applying simple rule-based overlay strategies.
• In a true out of sample check since 06/2022 a tactical asset allocation (TAA) exclusively based on the machine learning signals would have clearly outperformed its benchmark and even naïve allocation strategies. For the intended use case i.e., complementing our more comprehensive TAA approach, we find in a long-term simulation study that the added value generated by particular combinations of machine learning overlays can be expected to range from 20 to 30 bps per year.

Download the full publication below