Interview
Finding Signal in the Noise: Machine Learning and the Markets with In Young Cho
- Inyoung Cho joined Jane Street (referred to as "Chainstreet" in the opening) after a non-traditional path involving an identity crisis between biology and math, concluding that academic medicine was unsuitable and that finance required a different skillset.
- She describes the Jane Street interview process as intellectually exhausting, noting a literal physical sensation of her brain "overheating" during the on-site day, which paradoxically appealed to her.
- Her full-time role evolved from a broad "quantitative trader" title involving diverse tasks (learning OCaml, VBA, SQL, and handling broker calls) to a specialized research role leading machine learning initiatives.
- Jane Street's direct client interaction model, expanded in 2013, allows the firm to trade with known counterparties (like pension funds) to mitigate market impact and provide execution advice, contrasting with purely anonymous electronic exchanges.
- The "picking up the phone" workflow addresses scenarios where large, benign trades (e.g., quarterly pension rebalances) could disrupt anonymous markets; knowing the client's identity allows for better pricing and execution strategies.
- Research at Jane Street is broken into four stages: initial exploration using interactive tools, data collection (often manually verified for small datasets to understand errors), hypothesis modeling (using predictors/responses), and productionization.
- A critical tension exists between flexible, "quick and dirty" exploratory research (using Python/Notebooks) and robust, hard-to-modify production systems (often OCaml), with the former prone to reproducibility issues like inconsistent cell evaluation in notebooks.
- Tooling has evolved from crude methods (Excel, Bloomberg terminals) to sophisticated internal infrastructure that handles complex data cleaning, survivorship bias correction, and high-fidelity sub-second market visualization.
- Modern trading research increasingly relies on deep learning and tree-based models, shifting the discipline from "model interpretability" (using priors in low-data regimes) to "data discipline" (strict train/validation splits to prevent overfitting in high-data, high-noise regimes).
- Financial data is characterized by an "anti-inductive property" where market participants trade against known patterns, effectively removing signal and turning price action into noise, requiring models to extract signal from high-noise, non-stationary environments.
- Hardware constraints have become a primary factor in modeling decisions; for high-frequency strategies, model architecture is often shaped by hardware limitations (FPGA, GPU, CPU latency) rather than purely by predictive accuracy.
- Jane Street currently manages a fleet of thousands of high-end GPUs and processes tens of terabytes of market data daily to support the scaling of complex deep learning models.
- Future research directions include transfer learning across asset classes, multimodal data integration (text, images, geospatial), and applying foundation models (LLMs) to extract trading signals and improve internal developer efficiency.
- The firm is actively exploring how to leverage external LLMs and AI agents to build internal tools that enhance trader productivity, navigating the complexity of fine-tuning models versus integrating vendor solutions.
- Inyoung Cho notes that while the learning curve for new traders is steep due to complex tooling, the firm mitigates this through structured, foundational training that builds intuition incrementally before introducing advanced systems.