State-Space Design

State-space design is the failure mode arising because the state vector of a trading MDP — which variables to include (inventory, time, order imbalance, volatility, regime label) — is a hand-picked modelling choice. Too sparse a state breaks the Markov property by omitting decision-relevant history; too rich a state enlarges the space (curse of dimensionality) and invites overfitting to noise. It appears in this vault because Lalor & Swishchuk 2025 note that optimal market-making state spaces “in much of the literature … often include[s] variables such as inventory, order imbalance, market quality measures, differences between bid/ask prices and many more,” with no principled criterion for the choice.

Connections

• Markov Decision Process Trading Model — suffers_overfitting_risk, source: https://arxiv.org/html/2410.14504v2