Dynamic Calibration of Order Flow Models with Generative Adversarial Networks#

Summary:#

  1. The study discusses a limit order book (LOB) and a Poisson order flow model for it.

  2. The past models, including limit order submission, cancellation, and execution, had shortcomings in order to predict the queue size and the order arrivals at different price levels. The approach of this paper, which proposes a Poisson order flow model, is well motivated and uses independent Poisson processes for different order types.

  3. The research methodology involves the calibration of Poisson order flow models on short-term limit order book data, with additional information given by the set of conditioning variables: session data, a variable depending on the current time and day of the week, and a variable describing the current LOB state.

  4. The study aims to generate synthetic order flows with GANs, which can contain additional patterns compared to a model calibrated on historical data only. The GANs generate realistic patterns for the tick frequency, the spread-imbalance histograms, and the distribution of limit order queues. The synthetic order flow generated by the models matches the stylized facts in the data.

Background:#

  1. Subject and characteristics

    • The study discusses a Poisson order flow model for the limit order book (LOB).

  2. Historical development

    • The past models have had shortcomings to predict queue size and order arrivals at different price levels.

  3. Past methods

    • Past methods include limit order submission, cancellation, and execution.

  4. Past research shortcomings

    • Past models were not able to predict the order arrivals or queue size accurately, which was a challenge.

  5. Current issues to address

    • The study addresses the challenge of generating synthetic order flows, which contain additional patterns compared to a model calibrated on historical data only.

Methods:#

  1. Study’s theoretical basis

    • The study proposes a Poisson order flow model for the LOB, which involves independent Poisson processes for different order types, and considers queue size, price levels, and order events.

  2. Article’s technical route (step by step)

    • The technical route includes calibrating Poisson order flow models on short-term LOB data, using generative adversarial networks (GANs) to learn the distribution of these models, based on additional information given by a set of conditioning variables, and generating synthetic order flows with GANs.

Conclusion:#

  1. Work significance

    • The study proposes a method to generate synthetic order flows with additional patterns, which can improve model performance and provide insights into market dynamics.

  2. Innovation, performance, and workload

    • The study proposes the use of GANs to learn the distribution of Poisson order flow models and generate synthetic order flows with realistic patterns. The synthetic order flows generated by the models match the stylized facts in the data.

  3. Research conclusions (list points)

    • The study proposes a Poisson order flow model for the LOB and uses GANs to generate synthetic order flows with additional patterns.

    • The GANs generate realistic patterns for the tick frequency, the spread-imbalance histograms, and the distribution of limit order queues.

    • The synthetic order flow generated by the models matches the stylized facts in the data.