:   .., ..
:  
:  119
:   -
:  2026
:   .., .. // . - 2026. - . 119. - .257-283.
:   , , , ,
(.):  optimal execution, moscow exchange, reinforcement learning, implementation shortfall, price impact
:   , ~ . , (RL). 2025 ~ (AFLT). , . , ,~-- , , -- . , , . , RL- , .
(.):  The problem of optimal execution of exchange orders is among the most important for institutional investors in both international and Russian financial markets. This paper presents an approach to discovering an optimal trading strategy based on deep reinforcement learning (RL). For environment analysis and modeling, we use historical streams of incoming orders on the Moscow Exchange for April 2025 for the ordinary shares of Aeroflot (AFLT). The key innovations of the proposed model lie in a more comprehensive representation of the current market state and in the use of more sophisticated price-impact models, which bring trading conditions in the simulated environment closer to those observed in reality. The main components incorporated into the modellimit order book state, a propagator, and a volume-dependent temporary-impact functionserve as important factors in the agents decision-making process. To enhance the transparency of the decision-making system, we examine the importance of the environments state features by applying tools from game theory and computing Shapley values for each feature. The results demonstrate that the RL-based solution within the proposed framework outperforms trivial strategies in terms of execution shortfall, and that its decisions indeed depend on the current state of the order book.

PDF

: 88, : 19, : 8.


© 2007.