Relevance
9/10
Directly addresses quantitative trading with a novel state representation derived from Auction Market Theory for RL-based trading. Demonstrates substantial out-of-sample returns (+163.6% TSLA, +116.5% NVDA) at 1-second resolution with zero leverage. Introduces microstructure pattern theory as a structured input modality for sequential decision-making on financial time series. The prospect-theory-grounded reward shaping and risk-bounded execution layer are directly applicable to live trading systems. However, the high drawdowns and limited asset coverage temper practical deployability.
Implementation Complexity
7/10
Requires institutional tick data processing (nanosecond to 1-second aggregation), computation of AMT primitives (VPOC via rolling VWAP, Value Area boundaries, LVN detection, CVD, tape velocity, order-flow imbalance), Z-score normalisation, dual PPO agent training with custom reward functions, and a multi-layer risk-bounded execution engine with volatility-adjusted sizing and circuit breakers. The 5-layer pipeline (tick ingestion, OHLCV+Delta aggregation, 20-dim state engineering, dual PPO inference, risk engine) requires significant infrastructure. Use of Gymnasium and Stable-Baselines3 reduces some complexity, but the custom environment and reward shaping add substantial engineering effort.
Reproducibility
3/5
The paper provides detailed hyperparameter tables (Table 3) with selection mechanisms, specifies the Gymnasium environment, Stable-Baselines3 PPO implementation, and Databento data source. However, no GitHub repository or code release is mentioned. The institutional tick data from Databento is commercial and not freely available. The 80/20 chronological split and evaluation protocol are described, but a formal one-factor-at-a-time sensitivity sweep is explicitly reserved for future work. The synthetic transaction fee and reward constants are documented but their exact tuning process on validation slices is not fully reproducible without the data.