The Neural Networks Survival Kit for Quants

Using examples ranging from portfolio construction to algorithmic trading, this talk explains neural networks as a non-parametric econometrics technique. Matthew also provides various examples illustrating the tradeoffs between using Deep Q-learning versus supervised deep learning for predictive modeling with signals such as news sentiment.

Important Considerations

• Traditional Statistical Modeling
• Stats vs Machine Learning
• What Does a Network Classifier Output
• Taxonomy of Most Popular Neural Network Architectures
• Geometric Interpretation of Neural Networks
• Half-Moon Dataset
• Why Deep Learning

Summary

• Neural networks aren't themselves “black-boxes”, although they do treat the data generation process as a black-box The output from neural network classifiers are only probabilities if the features are conditionally independent (or there are enough layers)
• One layer is typically sufficient to capture the non-linearity in most financial applications (but multiple layers are needed for probabilistic output)
• Recurrent neural networks are non-parametric, non-linear, extensions of classical time series methods
• TensorFlow doesn’t check that fitted Recurrent Neural Networks are stationary