Reasoning on Time-Series for Financial Technical Analysis

Battle Count: 285

Highly relevant to quantitative trading. The paper directly addresses stock price forecasting using technical analysis signals, produces interpretable reasoning traces for practitioners, and demonstrates practical portfolio construction via Markowitz optimization with strong Sharpe ratios (1.719). The framework bridges the gap between black-box time-series models and interpretable financial analysis, which is critical for trading strategy development and risk management. The CAPM and Fama-French factor analysis further validates the trading signal quality.

High complexity due to: (1) multi-stage RL training pipeline (cold-start RL, rejection sampling + SFT, RL for reasoning), (2) cross-modal alignment between time-series and text embeddings using PCA and multi-head attention, (3) joint conditional training with classifier-free guidance, (4) feature regularization losses across transformer layers, (5) LoRA fine-tuning of multiple LLM components, and (6) integration of 10+ technical indicators as textual annotations. Requires 4x A5000 GPUs and approximately 100-120 GPU-hours for full training.

Code is publicly available on GitHub. Detailed hyperparameters, training pipeline stages, computational resources (4x NVIDIA A5000 GPUs), and LoRA configurations are provided. Statistical significance tests are reported. However, the multi-stage RL pipeline with rejection sampling and the specific technical indicator annotations add complexity to full reproduction.