Over the last five years, Machine learning has become an intriguing yet interesting topic in quantitative investment. Amongst the numerous families and flavors of algorithms, Deep Learning stands as the one which is as fascinating as it is complex. However, the obsession from researchers for this topic is continuously increasing and results show promising territory for empirical asset pricing and stock selection purposes. Despite the significant promises of neural nets in finance, the pros of those algos seemed clouded and sometimes shadowed by the inherent complexity and the resulting lack of embedded transparency that they exhibit. Naturally, this lack of inherent “interpretability” led to the inflated “Machine Learning is a black box” myth, which turned to be a limit in the trust and then the adoption of those algos in production.
This presentation will review the following points:
•Recap the basic notions of Machine learning and their related specificities for ML in finance
•summarising what are the notions of explainability and interpretability in the field of deep learning models for stock selection.
•Introducing a taxonomy of methods for interpretability
•Equity US stocks universe use case to explain the different techniques and results for:
- global models
- Local models
- Partial dependence
•Expanding the use case by focussing on two deep learning models for stock selection
- Multi-Layer Perceptron (MLP)
- Convolutional Neural nets (CNN)