Hybrid LSTM and PPO Networks for Dynamic Portfolio Optimization

Battle Count: 319

Highly relevant to quantitative trading as it directly addresses dynamic portfolio allocation across multiple asset classes including equities, bonds, and cryptocurrencies. The framework incorporates realistic transaction costs, sparse allocation constraints, and adaptive rebalancing. The hybrid predictive-policy architecture is a practical approach for systematic multi-asset strategies. However, the weekly rebalancing frequency and simplified cost model limit direct applicability to high-frequency trading contexts.

Moderate complexity. The LSTM component is standard (univariate, 64 hidden units, 30-week lookback). The PPO component uses Stable-Baselines3 with standard hyperparameters. The main complexity lies in the Top-K softmax projection for sparse allocation, the multi-asset state space construction, the simulated trading environment with transaction costs, and the coordination between LSTM forecasts and PPO state inputs. Separate models are trained for each K value. The pipeline requires careful handling of data alignment, normalization, and walk-forward prediction.

The paper specifies hyperparameters (learning rates, batch sizes, epochs, clip range, GAE parameters), software stack (Python, pandas, NumPy, scikit-learn, TensorFlow, Stable-Baselines3), data source (Yahoo Finance), and preprocessing steps (z-score normalization, log-returns, 70/30 split). However, no code repository is provided, and the exact random seeds and full training logs are not disclosed. The Top-K softmax projection algorithm is described in pseudocode (Algorithm 1 and 2), aiding reproducibility.