In particular, they implement backtesting using three agents: An exchange agent representing the exchange which keeps the order book (e.g., Nasdaq or NYSE), a market replay agent that provides
liquidity by replaying historical orders and an experimental agent representing the trading strategy to be evaluated. While the experimental agent is based on an interactive agent-based simulator named as ABIDES, the market replay agent is based on LOBSTER message data. The authors visualize a short segment of the replay data in Figure 2 in the paper,
Figure 2: Price-level volume plot. Black line represents the mid price, Each point is the price at different price levels with the colour scheme indicating the size (log scale) present at each level
The experimental agent is configured to participate in the simulation in a manner similar to the market replay agent, with the orders submitted dependent on the experiment carried out. She uses a strategic “greed” parameter to determine what size order to place relative to the available liquidity. In the visualized experiment, the impact agent queries liquidity within 1% of the inside bid (if selling) or ask (if buying) and with greed = 1.0 places an order to capture all of it. [How exactly the greed level is defined is not very clear for me.]
Figure 3(a); Observe impact on the mid price by the replay-only experimental agent placing buy market orders at twice the best ask size
Figure 5(a): Observed impact on the mid price by the IABS experimental agent placing buy market orders with greed = 1.0
The above two figures show the typical simulation result observed by the authors. They conclude that in the backtesting environment the price trends rather quickly back to the baseline price, eventually reaching that price and remaining there [the authors also point out that whether the price finally stabilized at the baseline price seems to depend on the trading side. Although by looking their figures, I can not see this clearly.] . However, in the IABS experiments, the price stabilizes at a new level in each set of experiments, suggesting that the impact of the order is longer lasting or even permanent.
