AI Based Investment Systems

Challenges

Noisy data - the signal to noise ratio in investment systems is typically low.

Non-stationarity - distributions change and so it can be challenging to train models that work well out-of-sample.

Small datasets - even with decades of data, the amount of data available to train models is small compared with other domains.

Approach

Curated data - we add human expert bias to the system through the selection of appropriate data for a given model.

Selection of techniques - some techniques are more applicable to finance than others, and both multivariate data sets and mixed frequencies are important characteristics.

Transparency - being able to interrogate the system and produce explanations for investors are important considerations. We've designed our AI system to be as explainable as possible.

Image by Denny Müller
Solution

The Global AI Allocator (GAIA), our machine learning-based prediction engine, combines unsupervised and supervised learning to create interpretable predictions across asset classes.

Predictions are made for a time horizon of one week to a few weeks ahead.

Model risk is mitigated by enhancing interpretability and using ensambles of models across different data pre-processing techniques and time horizons.

Implementation uses GPUs and FPGAs on cloud infrastructure.

Asset Predictions
 

Predictions for global assets together with key drivers / explanations

Stock Portfolios
 

Stock portfolios that depend on indicators or themes e.g. inflation expectations

Custom Indicators
 

NLP derived indicators such as risk appetite and sentiment based

Systematic Strategies
 

Long only, long short, and market neutral based strategies