TechPrufer
Get started
Hudson River Trading
Hudson River Trading logoHudson River Trading · Feature trail

AI Labs Summer Projects

2 posts · Sep 2021 – Sep 2026

What this trail covers

HRT AI Labs hosted interns who contributed to various AI research projects. Projects included Signal Architecture Search using Genetic Algorithms, optimizing Hyperparameter Search with Graybox Bayesian Optimization, and accelerating Quantitative Finance with GPU Computation by investigating CUDA performance for tall & skinny matrix multiplications. The interns' work focused on applying AI techniques to trading problems, exploring novel algorithms and performance optimizations. This post details the application of diffusion models to generate sequences of stock returns, exploring conditional and cross-sectional modeling, and defining evaluation metrics. It also discusses using market metadata as latent features for conditioning and inpainting capabilities. The post highlights diffusion models' potential for synthetic time-series data generation, avoiding mode collapse issues seen in autoregressive models.

All Hudson River Trading trails
  1. Intern Spotlight: HRT AI Labs ProjectThis post details the investigation into applying reinforcement learning to optimal execution problems in algorithmic trading. It describes modeling the problem as an optimal execution task where an RL agent uses short-term return predictions to trade towards a predefined target position. The post contrasts Soft Actor-Critic (SAC) and Proximal Policy Optimization (PPO), highlighting SAC's tendency to learn non-smooth policies and PPO's enhanced training stability due to on-policy data collection. A Memory-Augmented PPO is introduced to address partial observability. The evaluation compares the trading profit and loss (PnL) of these algorithms against a rule-based baseline under varying signal accuracy levels, showing Memory-Augmented PPO as the top performer.
  2. Intern Spotlight: HRT AI Labs ProjectThis post details the application of diffusion models to generate sequences of stock returns. It explores conditional modeling (e.g., generating returns conditioned on market metadata like volume and volatility) and cross-sectional modeling (e.g., generating correlated returns for multiple stocks). The post also discusses inpainting capabilities (generating missing stock return channels conditioned on other stocks) and the challenge of defining meaningful evaluation metrics for synthetic time-series data. It highlights diffusion models' advantage over autoregressive models like transformers and GANs in avoiding mode collapse.
  3. Intern Spotlight: 2023 HRT AI Labs Summer ProjectsThis post details three distinct intern projects within HRT AI Labs: 1. Signal Architecture Search using Diverse Multi-Solution Genetic Algorithms, which involved representing signals as DAGs and using a fitness function to encourage diverse signal generation. 2. Speeding up Hyperparameter Search with Graybox Bayesian Optimization, which models beliefs about individual functions contributing to a sum rather than just the sum itself, leading to more efficient exploration. 3. Accelerating Quantitative Finance with GPU Computation, which investigated the performance of CUDA's cuBLAS for tall & skinny matrix multiplications and evaluated a custom CUDA kernel (TSM2X) on newer hardware, highlighting the ongoing challenge of performance engineering.
  4. Intern Spotlight: 2022 HRT AI Labs Summer ProjectsThis post details the application of Large Language Models (LLMs) to financial news for sentiment analysis and prediction, and the use of Transformers for modeling both tick-by-tick order book events and cross-sectional return data. The LLM section discusses prompt engineering and the impact of model size and examples on accuracy. The Transformer section covers data representation.
  5. Applying Artificial Intelligence to TradingThis post introduces the application of Artificial Intelligence to trading by defining AI in the context of HRT's business. It contrasts the trading problem with board games like Chess and Go, highlighting differences in zero-sum nature, rule knowledge, state observation, and data availability. The post then decomposes the trading problem into three subproblems: Prediction, Optimization, and Execution, suggesting how AI techniques can be applied to each.