Dynamic Calibration of Order Flow Models with Generative Adversarial Networks

Abstract

Classical models for order flow dynamics based on point processes, such as Poisson or Hawkes processes, have been studied intensively. Often, several days of limit border book (LOB) data is used to calibrate such models, thereby averaging over different dynamics - such as intraday effects or different trading volumes. This work uses generative adversarial networks (GANs) to learn the distribution of calibrations - obtained by many calibrations based on short time frames. The trained GAN can then be used to generate synthetic, realistic calibrations based on external conditions such as time of the day or volatility. Results show that GANs easily reproduce patterns of the order arrival intensities and can fit the distribution well without heavy parameter tuning. The synthetic calibrations can then be used to simulate order streams which contain new dynamics such as temporary drifts, different volatility regimes, but also intra-day patterns such as the commonly observed U-shape that reflects stylized behaviour around open and close of market hours.