Extending the use of Deep Learning in Markets – three interesting papers.

Deep Reinforcement Learning for Trading, Zhang, Zohren and Roberts (2020). The authors use a variety of momentum indicators (including MACD and RSI) across 50 different futures contracts to represent the state space and then compare the results of three different RL algorithms (DQN, PG and A2C on top of an LSTM) with those generated from the individual momentum indicators.  They address the risk of overfitting through a combination of training on a cross-section of the futures contracts; constraining the number model parameters; and employing dropout.  

Their results in FX and commodities are particularly encouraging, despite high position turnover.  One nice feature of this work is that it targets a market position directly (on a volatility adjusted basis), rather than via a prediction/classification step first.  One question would be around use of momentum as the benchmark given the overall poor performance of this factor through the analysis period (2011-2019), but nonetheless provides some really good insight into balancing trend with mean reversion signals.

Mitigating Overfitting on Financial Datasets with Generative Adversarial Networks, De Meer Pardo, Cobo Lopez (2020). Use a Wasserstein GAN to generate additional synthetic data in order to address the issue of overfitting when applying deep neural networks to financial time series (given the relative lack of data).  In this paper they analyse the VIX index using 1000 samples, rolling forward 10 days at a time.  A couple of nice points from this paper:  

1. Potential use cases for testing different market regimes where you would ordinarily be constrained by the small sample size; or be required to specify the data generating process (if applying a stochastic model to generate synthetic data).  
2. The paper avoids issues with assessing the quality of the GAN by directly measuring its impact on the test-set performance of a 3rd model (in this case ResNet architecture).

Generating Financial Markets with Signatures, Buhler, Horvath, Lyons, Perez, Arribas and Wood (2020).  Examine the use of Market Generators on sparse financial time series data (as low as 250 observations). Compare classic models’ use of stochastic sample generation under explicit assumptions of underlying data distribution with newer ML approaches.  They expand on some of the work discussed in the prior paper by focusing on Variational Autoencoders (VAEs), being less data hungry than GANs, and the use of signatures to encode path valued data (as opposed to a finite set of synthetic samples).  Signatures can be likened to principal components insofar as their ability to encode a richer set of information in projecting an infinite-dimensional unknown distribution onto a finite N-dimensional vector space.  An area that is a real challenge for path dependent time series data.