Using Generative Models for Backtesting Trading Strategies

Recently, generative models are increasingly being used in financial markets, particularly in modeling high-frequency data such as order books and trade flows. Modeling this data is critical for backtesting trading strategies that influence the behavior of other participants. Thus, in our research we look at how, using generative models, we can predict the reaction of other market participants to our execution of large transactions. Our task is divided into several subtasks. First, we construct an order book and trade flow from the complete order log on the Moscow Exchange. Second, we calculate features based on the extracted data, including the current time of day into the features. Third, we train a classifier to predict the probabilities of the next events. Fourth, we train a Generative Adversarial Network (GAN) separately to simulate trade flow and order book changes. Fifth, and most importantly, we demonstrate the quality of the trained models by testing known properties of high-frequency data. We then demonstrate the use of the model for backtesting the execution of large trades and show its effectiveness for this task.