< back

PGSGAN: Policy Gradient Stock GAN

Masanori HIRANO, Hiroki SAKAJI, Kiyoshi IZUMI

International Journal of Smart Computing and Artificial Intelligence, vol.8, no.2, p. IJSCAI832, 2024


We introduce a novel generative adversarial network (GAN) designed to generate realistic trading orders for financial markets.Past models of GANs for creating synthesized trading orders have always been focused on continuous spaces because their architecture has constraints coming from the learning algorithm. Contrary to this, actual orders are placed on discontinuous space, which comes from the actual trading rules such as a minimum unit price for orders. In this study, therefore, we adopt a different approach supporting the actual trading rules and placing generated orders to discontinuous state space. The modification led to the inability to apply the standard GAN's learning algorithm, requiring us to use a policy gradient, a solution commonly used in the realm of reinforcement learning, as our learning strategy. Our experiments handle massive amounts of order data over half a year. Our experimental results indicated that the proposed model surpassed previous models in terms of the distribution pattern of the generated orders. A noteworthy advantage of integrating the policy gradient into our model is that it provides the ability to monitor the GAN's learning progress by analyzing the entropy of the generated policy.


Generative adversarial networks (GAN); Financial markets; Policy gradient; Order generation;




  title={{PGSGAN: Policy Gradient Stock GAN}},
  author={Masanori HIRANO and Hiroki SAKAJI and Kiyoshi IZUMI},
  journal={International Journal of Smart Computing and Artificial Intelligence},