High-touch Trading UI

UX in light of automated trading

Although the financial world is increasingly automated with machine learning and high-frequency trading, there are always tasks that humans or machine-assisted humans outperform machines alone.

The High-touch Trading UI is an analytic dashboard that combines the power of data science and the strengths of human dealers to maximize trading performance. The project, developed during my time at J.P. Morgan Asset Management, was a compelling case of ML/UX and Human-AI Interaction.

Although the NDA clause in my employment contract prevented me from disclosing the details, I could share the high-level process and ideas about the project to demonstrate my involvement and contributions.

ML/UX Research

The project started with identifying where and when human traders beat machines in the world of automation. The user research and analysis of historical trade performance revealed that human traders are better at making illiquid trades, such as buying exotic derivatives or selling a vast amount of stocks, by utilizing their professional network, flexibility, intuitions, and negotiation skills.

However, traders also need data to make informed decisions for illiquid instruments. And several pain points have been found.

  • The lack of immediately available data
  • The low variety of data for making informed trade decisions
  • The disconnection between data and actions

The solutions to these pain points have been incorporated into the final design. For example, the visualizations are loaded progressively so that users can start consuming the information as soon as possible. We also spent a significant amount of time designing new visualizations for data that were found useful during prototype testing. Action buttons were also added to the UI so that traders could act on their insights immediately.

The UX of Model Interpretability and Evolvability

Two important aspects of Human-AI Interaction in this project concern the interpretation of the model output and the iteration of the model itself. The UX decisions made could be demonstrated through the following wireframe, which represents one of the panels on the dashboard.

A typical workflow is to choose the best desk (teams that provide specific services in brokerage firms) to send an order (trade request) to. The panel ranks the list of desks with the corresponding metrics that measure their performance. There is also a special entry representing the recommendation from an ML model.


One of the design challenges is to help users interpret the model suggestions.

  1. Should the model recommendation be ranked naturally in the list or be always put at the top?

    Our testing showed that when the model recommendation is not ranked first, users become more skeptical about it, even though it is entirely normal that model output deviates from the metric-based ranking. As a result, placing the suggestions always at the top created a mental separation between the model recommendation and the rest of the list. This helped users choose between the model suggestions and their personal preferences more objectively.

  2. How should the UI explain the rationale behind the model recommendation?

    Since the model assigned a weight to each input factor, we categorized the factors and put the most heavily weighted one next to the recommendation as a label. This helps users understand why the particular desk was selected by the model.


Different from a lot of B2C machine learning models, traders are both the sole provider of input data and the sole consumer of output data in this internal trading system. This presented a dilemma. Users need to make "random" decisions for a small subset of trades so that the ML model can learn new patterns and remove biases. However, when users know that a recommendation is random, they are disincentivized to pick it.

We solved this problem by presenting the random choices as model recommendations as well. Users could not tell if a recommendation was random or not. However, if they feel that the recommendations do not make sense, they can still select the best desks according to their own evaluation. This allows the model to gather additional data and evolve while minimizing the impact and frustration of making random trade decisions.

Please feel free to contact me with any questions about the project. I will provide more information as far as the NDA allows.