Adaptive Regime-Aware Stock Price Prediction Using Autoencoder-Gated Dual Node Transformers with Reinforcement Learning Control

Battle Count: 92

Highly relevant to quantitative trading. The paper addresses regime-dependent prediction, which is critical for trading strategy robustness. 72% directional accuracy (vs. 54% naive baseline) and 0.59% MAPE for 1-day predictions are practically meaningful. The regime-aware routing maintains MAPE below 0.85% during high-volatility periods where baselines exceed 1.5%, directly relevant to risk management. The adaptive threshold learning avoids hand-crafted regime rules that become stale. However, the paper does not include transaction costs, market impact, or execution constraints, and the 20-stock universe with survivorship bias limits direct trading applicability. The framework is more of a prediction architecture than a complete trading system.

High complexity due to three interacting components (autoencoder, dual node transformers, SAC controller) with a 4-stage training pipeline. Requires graph construction with learnable edge weights, expanding-window normalization, sentiment data pipeline (BERT fine-tuning + X API scraping), and RL training with delayed/noisy rewards. The node transformer architecture (6 layers, 8 heads, 512 dim, 252-day sequences) is computationally intensive. SAC training adds another optimization loop. Approximately 33% longer training than baseline. Inference is simpler (frozen weights, single forward pass per component), but the overall system requires significant engineering effort for production deployment.

The paper provides detailed architecture specifications, hyperparameters (Table I), training pipeline (4 stages), feature engineering details, and evaluation metrics. However, no GitHub repository is mentioned. The sentiment data collection via X (Twitter) API is noted as potentially non-replicable due to API policy changes. The MSE dataset is publicly available, but the CSS dataset construction details are referenced to prior work. Temporal splits, normalization procedures, and imputation strategies are clearly specified. Baseline hyperparameter search ranges are provided in Table X.