Stock Price Formation: Precepts from a Multi-Agent Reinforcement Learning Model

Mesoscale effects of trader learning behaviors in financial markets: A multi-agent reinforcement learning study

“…The “SYMBA” MAS stock market simulator we used for this work [ 53 , 54 ], was calibrated to the London Stock Exchange data between the years 2007 and 2018. In this model, the agents autonomously manage their portfolio via a long-only strategy based on two reinforcement learning algorithms: one performing price forecasting and another one performing stock trading.…”

“…In a previous publication [ 53 ], we detailed the cautious calibration procedure of SYMBA to real stock market data, also see Supplementary Material. In [ 54 ], we then studied how its agents learn and acquire new trading strategies over time. SYMBA hence emulates the microstructure of a financial stock market through a bottom-up approach to system complexity, via these autonomous economic agents (e.g.…”

“…Each agent learns to forecast and trade autonomously by reinforcement learning with a long-only equity strategy, and each agent’s discretionary asset pricing process relies on learning to weight both chartist and fundamentalist inputs. These features allow us to plumb the financial market from the macro- to the microscopic level: in [ 53 ], we compared the overall stock market return statistics to real world data, while in [ 54 ] we characterised the trading performance of the agents from the learning dynamics of their trading strategies over time.…”

“…IN SYMBA all of the agents are autonomous and are endowed with reinforcement learning [52], by which they forecast stock prices and send individual transaction orders to a centralised order book. In a previous publication [53], we detailed the cautious calibration procedure of SYMBA to real stock market data, also see Supplementary Material. In [54], we then studied how its agents learn and acquire new trading strategies over time.…”

“…For the sake of clarity, the entire code of SYMBA is available to the community on GitHub [67]. We also provide as Supplementary Material a detailed sum up of the main results of the calibration procedure from [53] to validate the model. In Section 4, we study the impact of the reinforcement learning rate on agent performance and overall market dynamics.…”