Statistical Arbitrage in Polish Equities Market Using Deep Learning Techniques

Battle Count: 75

Highly relevant to quantitative trading. The paper presents a complete statistical arbitrage framework with three distinct factor construction methods, detailed trading signal generation via Ornstein-Uhlenbeck mean-reversion modeling, and comprehensive backtesting on a real market. The introduction of LSTM for pairs trading replication is novel. Results show practical profitability (up to 20% combined return, Sharpe ratios up to 2.63) with realistic transaction costs. The adaptation of Avellaneda-Lee methodology to a smaller, less liquid market is practically valuable for emerging market quant strategies.

Moderate to high complexity. Requires: (1) PCA implementation on returns correlation matrices with eigenvalue decomposition, (2) Stacked LSTM network training with custom loss function (MSE + L1 penalty) using Adam optimizer, (3) Ornstein-Uhlenbeck parameter estimation via AR(1) fitting (Yule-Walker equations), (4) Daily signal generation with threshold-based trading logic, (5) Portfolio construction with 60 stocks and multiple replicating factors, (6) Transaction cost modeling with leverage. The LSTM component requires GPU training infrastructure. The overall pipeline involves multiple interacting components that need careful coordination.

The paper provides detailed mathematical formulations, parameter settings (W=120, r=15, hidden sizes=64, batch size=16, training window=120 days), and trading rules. However, no code repository is mentioned. The LSTM architecture details are provided but hyperparameter optimization is described as 'rule of thumb' based. Market data from GPW (Warsaw Stock Exchange) would need to be sourced separately. The backtesting methodology is well-documented with specific thresholds and assumptions.