In Young Cho on Trading Research and Finding Signal in Financial Data

Ron Minsky introduces In Young Cho, a Jane Street researcher and trader who helps lead machine-learning work. They discuss her path into trading, how research is done, and what makes financial modeling distinctive.

Cho initially expected to become a doctor, while also pursuing competition mathematics. After deciding medicine was not right and trying internships in several industries, she heard friends praise Jane Street’s interview process and decided to try it.

The interviews were challenging enough that she felt mentally exhausted, but she enjoyed the experience. Despite knowing little about finance, she accepted an internship and eventually joined full-time after undergraduate study.

The internship offered a new puzzle or skill to learn each day. Cho valued both the work and the support of smart, intellectually humble colleagues who invested in helping interns learn and make the most of their abilities.

As a quantitative trader, Cho learned OCaml, VBA, shell tools, SQL, financial terminology, and how to speak to brokers. She recalls picking up a call and shouting “park” because she did not know it was a button on the phone.

Minsky asks why a technology-oriented trading firm takes broker calls at all. Cho explains that around 2013 Jane Street began building more direct client relationships alongside trading anonymously on electronic exchanges.

A large pension fund rebalancing an index portfolio illustrates the difference. A market maker seeing a large stream of anonymous buying may not know its full size or whether it reflects information the market maker lacks, and may quote more conservatively.

Knowing the client and the intended trade can help the firm estimate impact and provide pricing or execution advice. Minsky notes that information about a counterparty can be valuable to both sides when the trade is a routine portfolio rebalance rather than an informed bet.

Cho also wrote tools to visualize trades and analyze past opportunities. She describes these activities as a mix of software engineering, research, and live trading, even though they were all part of a trader’s day-to-day job.

Minsky adds that live trading also includes real-time decisions about automated systems and their behavior, not just calls with a counterparty. Both speakers note that the title “trader” can mean different things at different firms.

Asked what research looks like, Cho outlines a process beginning with exploration. Interactive tools must display information quickly and clearly so a researcher can develop and discard hypotheses without a long delay.

For the pension-fund example, exploration might involve examining the client’s portfolio and public filings to understand changes in positions and possible order flow. The purpose is to form a view of likely market impact and what price can be offered.

A hypothesis might be that most of a client’s trades have little impact but a particular position moves the market. Testing it requires collecting possible trade dates, sizes, times, and returns, along with other relevant information.

Minsky stresses the infrastructure needed to handle stock splits, ticker changes, and errors. Cho says manually building a small dataset can be a useful way to learn where errors arise before working with millions or billions of rows.

The modeling stage asks what to predict and which features matter. With a handful of observations, a simple linear regression may be preferable because a more expressive model can overfit. Larger datasets and complex interactions may justify more sophisticated methods, including deep learning.

A model and apparently credible out-of-sample predictions are only part of the job. Productionizing means deciding how to express the prediction in the market and ensuring the technology actually executes the intended trade reliably.

Minsky distinguishes two styles: simulate something close to a production trading system, or study a simplified version of the problem and later translate the result into production. He asks how Cho weighs those approaches.

For a new market or low-conviction idea, Cho prefers flexibility and fast hypothesis iteration over building a complete production system first. But quick studies made with nonproduction tools can later create reliability and deployment difficulties.

A mature system provides a robust experimental environment, but even a small change may require learning a large library ecosystem. Neither approach is inherently better; the right balance depends on the maturity of the work, the market, and the research question.

Minsky says some desks naturally research by changing their existing algorithmic systems. Simulation still has costs, including processing huge quantities of raw data. Lightweight summaries and short iteration times matter because many research ideas will fail.

Cho compares the evolution of research to a sketch gradually becoming a detailed painting. Earlier tools supported cruder studies, but an individual could understand each stage closely, from the meaning of every dataset column through regression and the trading implementation.

That hands-on experience was educational even when the work looks primitive in retrospect. As the tools and models became more capable, preserving that intuitive understanding for new researchers became harder.

Cho says training starts with fundamentals rather than presenting the full abstraction immediately. New traders build an understanding of individual dynamics before combining them into a strategy; researchers may spend time learning to read one example visualization before confronting hundreds.

Their tools have improved substantially. Cho estimates a study that took a week a decade earlier might take ten minutes with current tools, though the increased capability also brings complexity.

Data infrastructure must account for corporate actions, stock splits, ticker changes, and historical universes. Simply looking at companies that exist today can introduce survivorship bias when reconstructing what traded many years ago.

The firm’s tooling can attach metadata such as market capitalization, volatility, and internal historical valuations to observations at the right times. Cho nevertheless urges researchers to trust but verify: check data and report suspicious results rather than assume every value is flawless.

Visualization tools can inspect market events at finer time scales and build interactive historical or real-time plots. Minsky contrasts this with manually manipulating spreadsheet rows and using basic graphing controls.

Researchers can draw on libraries of tree-based models, deep-learning models, and specialized regressions. Python notebooks make it easy to organize an interactive study and annotate the reasoning behind code, but were not designed as production systems.

Minsky finds notebook execution order frustrating: cells can be run in an arbitrary sequence, leaving visible code inconsistent with the state that produced a result. Reproducibility is also difficult because large research jobs can be too expensive to rerun casually.

Cho describes the tension between researchers wanting quick changes and engineers needing reliable, performant systems. People who work across both domains can bridge the gap and prevent those differences from becoming territorial.

Minsky argues that reproducibility is necessary inside research too. If a model changes, researchers need to know whether the experiment changed one intended variable or several things, such as underlying data sources and code, at once.

Cho agrees that hard experience often teaches this lesson. At Jane Street the boundaries between research and production roles are not rigid, and people who understand both can help build better tools and mutual understanding.

The conversation turns from simpler linear models to neural networks and tree-based approaches. Cho notes that researchers accustomed to small datasets and strong prior intuitions must change habits as models and datasets grow.

When data is scarce, researchers may restrict the hypotheses they test and preserve as much data as possible for fitting. With larger, more complex models, Cho stresses discipline around training and validation sets, out-of-sample evaluation, and avoiding duplicated hypothesis searches across groups.

Minsky frames these as different forms of discipline: constrain the model and hypotheses in a low-data setting, or use careful data and evaluation procedures when models are less interpretable. Cho says the shift can be uncomfortable for researchers who want an intuitive story for model behavior.

A simple model may be understandable enough to adjust after an unusual event such as a pandemic. Minsky calls this a form of intervention-oriented interpretability: understanding the model sufficiently to decide how to change it when the world changes.

Cho describes financial machine learning as unusually hard. Her metaphor is having 100 units of data of which only one contains useful signal, without knowing which one. It is an illustration of low signal-to-noise, not a measured ratio.

Millions of interacting participants determine prices for thousands of instruments, so any feature set may explain only a small fraction of returns. Minsky adds that market participants trade against observable regularities, which can remove the very patterns researchers learn from the past.

Cho calls this an anti-inductive property of markets. Unlike a stable benchmark, financial data can shift after crises or other events; sometimes the market operates differently because the world itself has changed. A robust model must cope with such regime changes.

The firm handles tens of terabytes of market data per day, Minsky says, and predictions span very different time horizons. Some systems care about changes on the order of ten nanoseconds, while other decisions involve humans and much longer time scales.

Cho says hardware is not an afterthought for the fastest models. Latency and throughput limits can determine the design problem: first understand what can be served on available hardware, then seek the best model within those constraints.

Minsky lists CPUs, FPGAs, GPUs, different memory systems, interconnects, and storage as relevant constraints. He connects this to the broader machine-learning lesson that architectures which scale efficiently on available hardware can become especially powerful.

He does not think scale makes prior knowledge irrelevant to every trading problem. In data-poor domains, researchers can carry insights from related markets; elsewhere the ability to scale changes both model and hardware choices.

Asked about the future, Cho is cautious about predictions but excited about recent progress. She says Jane Street had built access to thousands of high-end GPUs and improved its understanding of models, data, and serving hardware.

The firm records large quantities of market and other data and has advanced its use of deep neural networks for prediction. Cho expects further work on transferring knowledge from well-understood asset classes to data-scarce ones and combining market data with text or images.

Minsky distinguishes extracting information from familiar modalities such as text and images from modeling financial time series directly, where much relevant practical work is not published. He expects multimodal and reasoning advances to create opportunities across the wider trading process.

Both are also interested in tools that make employees more effective: assistants or agents able to work with unusual internal languages, systems, and datasets. Cho says decisions about what to build, fine-tune, or obtain from partners remain difficult and open.