Shuo Yin (尹硕)

Jun 2025 - Sept 2025, ICASSP 2026 In Submission

Authors: Shuo Yin, Jiahong Zhu, Tianlong Yang, Yuqiao Liu

Developed TS-Agent, a reinforcement learning–based agent for financial time-series forecasting, using LLM-generated strategy pool for exploration-enhanced fine-tuning, then seeding stepwise RL from selected strategies with custom RL module.

Sep 2025 - Present, ACL 2026 In Submission

Authors: Zhaolu Kang, Junhao Gong, Wenqing Hu, Shuo Yin, Zhicheng Fang, Yingjie He, Rong Fu, Eric Hanchen Jiang, Leqi Zheng, Xi Yang, Xiaoyou Chen, Richeng Xuan

QuantEval is a new benchmark designed to evaluate Large Language Models (LLMs) in financial quantitative tasks. It assesses three areas: knowledge-based question answering, quantitative reasoning, and quantitative strategy coding. Tests on many top LLMs show major performance gaps. To address this, we apply supervised fine-tuning and reinforcement learning to improve results.

Aug 2025 - Present, KBS In Submission

Integrated Agents, Genetic Algorithms, and Flow Networks to produce diversified alpha factors Developed factor alignment and filtering framework for A-shares and cryptocurrencies, implementing data cleansing, dual-chain dynamic iteration, and agent-based reporting

Aug 2025 - Present, CVPR 2026 In Submission

Authors: Rong Fu, Chunlei Meng, Shuo Yin, Zijian Zhang, Weizhi Tang, Shaobo Wang, Youjin Wang, Simon James Fong

This research presents a vision-based approach that abandons text and allows visual features to communicate through a graph of image prototypes. Using self-supervised learning, the model uncovers emergent semantic structures, where meaning is determined purely by vision. Text is only used for interpretation, not during training. This method demonstrates that vision alone can recognize and name the world without relying on textual input.

Jul 2025 - Present, CVPR 2026 In Submission

Authors: Haidong Kang, Shuo Yin, Huiquan Zhang

This work introduces OGDiff, a novel OSDG framework that replaces expensive second-order meta-learning with a conditional diffusion process. This avoids inner-loop backpropagation, reducing computational costs and improving stability. The MSDFF module is also introduced to better capture both fine- and coarse-grained features, enhancing recognition of known and unknown classes.

Feb 2024 - Sept 2024, Published in Nonlinear Dynamics (SCI Q2)

Authors: Feier Chen, Shuo Yin, Jiahang Zhang, Yi Sha & Huaxiao Ji

Based on MF-DFA and multi-peak analysis, the multifractality of LNG spot freight rates was verified, and factor signals were established, including spectral decomposition components, peak positions, and long-range correlation strength. Using methods such as LSTM, LNG futures return predictions were made, incorporating indicators like the Hurst exponent to achieve strategy switching and window smoothing. The backtest results showed an annualized return rate exceeding 36%.

Sep 2023 - Feb 2024, Published in SPIE

Authors: Kaixiang Zhang, Haoran Zhang, Qi Lan, Zesen Chen, Shuo Yin, Yongqi Wei, Zehao Wang, Jianlong Yang

This research presents a zero-shot OCT segmentation method that overcomes the limitations of traditional deep learning approaches, which require re-labeling and retraining for unseen Regions of Interest (ROIs). The method only needs box or point prompts for ROIs in the first frame, enabling automatic segmentation in subsequent frames.

Sep 2023 - Mar 2024, Bronze Medalist

• Derived high-frequency market microstructure factors from order book and auction data, including buy-sell imbalance ratios, price-volume triple barriers, reference price lags/differentials, statistical moments, depth-weighted spreads, and volatility proxies
• Built LightGBM model with rolling training and Optuna-based hyperparameter optimization; achieved MAE of 5.473 on validation set

Jul 2022 - Feb 2023, SJTU Outstanding PRP Project

Built a real-time database by web scraping public opinions and disclosure information from trading platforms using proxy pools to enhance crawler efficiency. Fine-tuned BERT NLP model to establish a text sentiment prediction framework, providing ESG ratings for 3,000+ listed companies based on mainstream ESG rating indicators