AlphaRJM: Reward-Jump Memory for Stochastic Return-Guided Alpha Discovery

Battle Count: 61

Highly relevant to quantitative trading. The paper directly addresses formulaic alpha discovery, which is a core task in systematic equity strategies. It evaluates on standard Chinese equity universes with realistic forecasting horizons (20 and 42 trading days). The pool-dependent reward mechanism mirrors real-world alpha combination. The method produces interpretable symbolic expressions (unlike black-box neural models), which is valuable for risk management and regulatory compliance. Results show consistent improvements in IC, ICIR, RIC, and RICIR across multiple universes. The cumulative wealth analysis demonstrates practical portfolio-level benefits.

The implementation involves multiple coupled components: (1) GRU-based expression encoder with positional embeddings, (2) permutation-invariant pool encoder with item-level features, (3) event-driven Reward-Jump Memory with learned event encoder, jump-context network, and gated jump heads, (4) action-conditioned SDE return critic with Euler-Maruyama integration, (5) distributional Bellman learning with energy-distance matching, mean calibration, and jump regularization, (6) contiguous sequence replay with burn-in memory reconstruction, (7) epsilon-greedy exploration with annealing schedules, and (8) target network synchronization. The architecture requires careful coordination between fast token-construction and slow event-driven memory timescales. However, all components are well-specified with concrete dimensions and hyperparameters.

The paper provides detailed implementation specifications including architecture dimensions, hyperparameters (Table 3), Euler-Maruyama discretization, exploration schedules, replay buffer settings, and optimization details in Appendix B. Pseudocode (Algorithm 1) is provided. The computational environment is specified (Apple M4, Python 3.12.13, PyTorch 2.13.0, Qlib 0.9.8.dev32). However, no public code repository is explicitly linked for AlphaRJM itself. Baseline implementations reference public GitHub repos for AlphaSAGE, AlphaQCM, and AlphaGen. Theoretical proofs are provided in Appendix F.