We have built our quantitative models using behavioural analytics with machine learning to understand how the participants, who generate positive and negative capital flows for each portfolio component, interact. Our parameter selection software runs on the AWS Supercomputer infrastructure. We build decision trees – each trained on a huge volume of historic price data for each component, to reveal repeating behavioural patterns.

Our models are tactical in nature as opposed to long-term econometric models. We use both directional and mean reverting models.

For each portfolio component we use separate long and short models. Each model comprises a parameter set for position entries and one for position exits so that there are four sets for each component.

All historic patterns are stored so that the code, which is run daily can search for and match them in the future.

Our proprietary code ranks parameter combinations by the Alpha generated. We select combinations which are part of a group with a flat kurtosis of returns, so that in the event of parameter shift the Alpha generated is not significantly degraded.

The simplest way to think about the philosophy behind the models is that if a market is trending… at some point there will be a reaction to the trend and at the end of that reaction there will be a move back in the direction of the trend. At each turning point a pattern of daily bars will be created.

The machine learning code can distinguish between significant patterns and non-significant patterns.