Issue #20

Why Structured Dynamic Learning.

At Noviscient, we believe the future of investment management will be the use of increasingly sophisticated machine-learning technologies to enable different business models that create real value for all investors.

Technically, our objective is to understand and model the financial markets, so we can make better decisions along our value chain of trading, selecting fund managers, constructing portfolios of securities and risk management.

Deep Learning

Deep learning has captured the public’s attention through its remarkable success in areas such as autonomous vehicles, image recognition and speech translation. However, it generally operates on static problems where massive data sets are available. In a certain sense, deep learning models are 'dumb'. There is minimal structure in these models and all the information has to come from the (big) data. This absence of structure is also why interpreting deep learning models is difficult.

Challenge of Financial Data

Financial markets have quite different characteristics that makes it challenging for the deep learning approach.

• Noisy data – with low signal to noise ratio and multiple levels of uncertainty (measurement, model and parameter)

• Limited data – spread across many securities with decaying information content

• Time-variation – through both natural and adversarial dynamics

• Dependence – across space (cross-correlation) and time (auto-correlation)

Structured Dynamic Learning

We have developed and continue to work on a technology platform called Structured Dynamic Learning as an alternative to deep learning. It uses model-based learning within a probabilistic framework to directly address the challenges of financial data.

Model-based learning introduces domain knowledge within a coherent framework for selecting, comparing, expanding and analyzing models. Learning the components and (causal) dependencies in the model may be manual or automated.

Our probabilistic framework uses probabilistic programming to represent variables with distributions. This generative, Bayesian approach (based on integration rather than optimization) enables decision-making based on expectations of functions of variables which we believe better translates to good decision-making.

Fundamental Advantages

Our Structured Dynamic Learning technology provides a number of fundamental advantages when dealing with financial data.

• Captures all model uncertainties in a coherent framework allowing us to operate with full distributions which provides a much deeper understanding of risks.

• Probabilistic (Bayesian) modelling makes it simple to introduce structure in the form of priors on parameters and likelihoods. It also facilitates online updating of distributions to enable fast and adaptive decision-making. 

• Probabilistic programming is a generative approach that allows us to deal with complex models and high dimensional problems without having to resort to analytic shortcuts and heuristics.

Outcome

The outcome is a superior ability to explain and predict financial data which means we can make better business decisions along the whole finance value chain. We will be creating real value for our investors through better decision-making that includes:

1. Faster and more accurate manager selection (identification of trading strategies with alpha)

2. Dynamic allocation of capital for granular and effective risk control

3. Customization of portfolios to precisely match investor preferences on demand

Struggling to raise capital? It’s time to do it the modern way.