Abstract
Systematic trading strategies are algorithmic procedures that allocate assets aiming to optimize a certain performance criterion. To obtain an edge in a highly competitive environment, an analyst needs to appropriately fine-tune their strategy, or discover how to combine weak signals in novel alpha creating manners. Both aspects, namely fine-tuning and combination, have been extensively researched using several methods, but emerging techniques such as Generative Adversarial Networks can have an impact on such aspects. Therefore, our work proposes the use of Conditional Generative Adversarial Networks (cGANs) for trading strategy calibration and aggregation. To this end, we provide a full methodology on: (i) the training and selection of a cGAN for time series data; (ii) how each sample is used for strategy calibration; and (iii) how all generated samples can be used for ensemble modelling. To provide evidence that our approach is well grounded, we have designed an experiment with multiple trading strategies, encompassing 579 assets. We compared cGAN with an ensemble scheme and model validation methods, both suited for time series. Our results suggest that cGANs are a suitable alternative for strategy calibration and combination, providing outperformance when the traditional techniques fail to generate any alpha.
Acknowledgments
Adriano Koshiyama would like to acknowledge the National Research Council of Brazil for his PhD scholarship, and The Alan Turing Institute for providing the environment to conclude this work. We are also grateful to Dr. Emre Kazim and Dr. Beatrice Acciaio for proofreading, reviewing and providing comments on this work. Finally, we are grateful to the anonymous reviewers for their constructive feedback.
Disclosure statement
No potential conflict of interest was reported by the author(s).
Notes
1 From now on, in line with the literature, we are omitting the subscript of D and G to keep the notation lighter.
2 Using would be more appropriate, but in line with the literature we kept the
notation.
3 We are highlighting this period in particular because our analyses and results concentrated on taking samples from 2001-2013.
4 The readers interested to understand more about the nonparametric statistical tests used in this work—Friedman, Holm Correction and Wilcoxon rank-sum test—should consult (Derrac et al. Citation2011).
5 A measure of dispersion calculated by taking the difference between the 3rd quartile (75%) and 25% 1st quartile.
6 When we rank the model validation schemes for a given asset, it means that we sort all them in such way that the best performer is in the first place (receive value equal to 1), the second best is positioned in the second rank (receive value equal to 2), and so on. We can repeat this process for all assets and compute metrics, such as the average rank (e.g. 1.35 means that a particular scheme was placed mostly near to the first place).
7 We decided to omit Ridge since all of the correlations were above 0.8.