Continuous Program Search

Battle Count: 63

The paper is directly focused on algorithmic trading strategy discovery. It introduces a trading-specific DSL (GPTL) with entry/exit signals, evaluates strategies using Sharpe ratio on five liquid futures contracts with realistic transaction costs (0.1% slippage, 0.05% fees), and uses a rigorous walk-forward evaluation protocol spanning 2008-2025. The core contribution—improving evolutionary search efficiency for trading strategies—has direct practical implications for quantitative strategy development. The interpretable program representation is particularly valuable for trading where understanding strategy logic is essential.

Implementation requires: (1) designing and implementing a typed DSL with grammar-aware tokenizer, (2) training a transformer-based VAE with block-factorized latents, (3) implementing behavioral geometry diagnostics (decode validity, AST edit distance, action divergence), (4) training a flow-matching model on logged evolutionary traces, (5) implementing the (μ+λ) evolution strategy with geometry-aware mutation masks, and (6) building a deterministic backtesting engine with walk-forward evaluation. The multi-component pipeline and interplay between representation learning and evolutionary search add significant complexity.

The paper provides detailed grammar specifications (GPTL), complete hyperparameters (μ=34, λ=66, 100 generations, 128-dim latent), training procedures (AdamW, lr=1e-4, 50 epochs, batch size 128), walk-forward evaluation protocol with exact date ranges, and algorithm pseudocode (Algorithm 1). Five independent seeds per configuration are used. However, no code repository is explicitly linked, and the flow model training details (architecture specifics, training hyperparameters) are somewhat abbreviated.