Screening for GEMs

This proprietary screening model takes a three-pronged approach to find the best medium-term long and short ideas within the top100 largest stocks in the MSCI GEM Index. It is based on 1) a typical quantitative multifactor screening model and then two Artificial Intelligence Technology-based approaches using 2) a Deep Neural Network model, and 3) a Recurrent Neural Network model.

1. Multifactor screening model: This more closely resembles the traditional quantitative analysis multifactor screening models used by the buy-side. It comprises of 17 factors, grouped within traditional Quality, Valuation and Momentum buckets, and derives its outputs based on simple linear regression mathematics. The main benefit of this model is that it is very intuitive since the linear correlations between the factors and the output are easy to understand and track. 
2. Deep Neural Network model: Using Artificial Intelligence technology, this model is based on a Deep Neural Network architecture where hidden data layers allow for sophisticated non-linear correlations between key input variables, segmented between Macro, Technical, Fundamental and Value inputs. This is similar architecture to that used in Image recognition (amongst many others). The value-add from this model is that it can derive meaningful conclusions from the non-linear inter-correlation of thousands of inputs per stock. This model has trained on over 150,000 data points (and growing) for each stock analysed.
3. Recurrent Neural Network model: Also using Artificial Intelligence technology, this model is based on a Recurrent Neural Network architecture. This is similar architecture to that used in most Natural Language Processing applications. This model also allows for sophisticated non-linear inter-correlations between key input variables, segmented between Macro, Technical, Fundamental and Value inputs. However, the main benefit from this model is that it also incorporates a “time”-dimension, very suitable to analysing time-series data, which allows the model to “look-back” and contextualise inputs over time. This model is trained on up to 50,000 data points (and growing) for each stock analysed.

At its core, the model attempts to predict which stocks have a better probability of positive US$ share price performance over the following 3-month investment horizon.