The two-person YC lab says delta-neutral strategies gained 140%, but it has not published the period, capital base, leverage or methodology.
By Ryan Merket · Published
Primary source: Y Combinator
Why it matters #
Prodigy Research is testing whether a domain-specific foundation model can monetize through proprietary trading. Its 140% claim will remain marketing until the lab publishes enough data to assess risk and repeatability.
Brothers Michael Wang and Yuhua Wang launched Prodigy Research earlier in August around a direct proposition: train a foundation model for quantitative finance, let it trade live capital and judge the system by the returns.
The San Francisco lab, part of Y Combinator's Summer 2026 batch, said in its launch post that its delta-neutral strategies returned 140% during the batch. Prodigy Research said the S&P 500 gained 1.01% over the same period while the Nasdaq fell 3.67%, and claimed its systems never recorded a down week.
Those figures come entirely from Prodigy Research. The launch materials do not state the exact measurement period, starting capital, gross or net exposure, leverage, transaction costs, asset classes or maximum drawdown. Prodigy Research also has not published independently audited performance. A 140% gain on a small experimental account carries different weight from the same return on institutional capital, particularly when execution costs and market impact enter the calculation.
That gap matters because the performance claim is doing much of the work in Prodigy Research's introduction. The model itself remains largely undescribed.
A proprietary benchmark against frontier models
Prodigy Research calls its system a foundation model and agent harness for trading and financial reasoning. It says the system outperformed Claude Fable 5 and GPT-5.6 Sol, two general-purpose frontier models, on financial and quantitative research tasks.
Prodigy Research has not identified those tasks or released prompts, scoring criteria, sample sizes or model outputs. It also has not disclosed the model's architecture, parameter count, training data, compute budget or the extent to which the system uses reinforcement learning. Without that material, the comparison cannot be reproduced or separated from the agent scaffolding surrounding the underlying model.
The distinction is important in quantitative finance, where a system can look strong because of its tools, data access, execution layer or benchmark design. A model trained or evaluated on information that overlaps with a test set may also appear to reason about markets when it is retrieving patterns already present in its data.
Prodigy Research's public site describes autonomous agents searching for trading opportunities and says the model is already operating in live markets. It does not present an API, a software subscription or a customer-facing product. The public pitch points instead toward proprietary trading: use the model internally, capture the resulting alpha and avoid selling the same signals to clients who could trade them away.
That structure puts pressure on Prodigy Research to show a repeatable track record. Quantitative strategies often degrade as capital increases, competitors discover similar trades or market conditions move outside the period on which a model was trained. Delta neutrality can reduce broad directional exposure, but it does not remove leverage, liquidity, concentration, execution or model risk.
Two brothers from DeepMind, Jane Street, Apple and Salesforce
The founders' backgrounds explain why Prodigy Research is attempting to combine model research with live trading rather than build another financial-document assistant.
Michael Wang previously worked on a Jane Street trading desk and later became a senior researcher at Google DeepMind. Prodigy Research's website says he helped train Gemini 3.1 Pro and held a Series 57 securities trader license. Google published the Gemini 3.1 Pro model card in February 2026, describing it as a multimodal reasoning model for complex, agentic and algorithm-development tasks.
Yuhua Wang previously built AI systems at Apple and worked on Salesforce's flagship AI platform, according to the founders' YC profiles. Prodigy Research says the brothers brought more than 16 years of combined experience across quantitative finance, model research and large-scale AI infrastructure. It also says they left seven-figure jobs to start the two-person lab.
The brothers are applying the familiar foundation-model strategy to a domain where a useful model can generate revenue without waiting for enterprise procurement. Prodigy Research can trade its own capital while improving the system, keeping both its data and strategies private.
The same approach makes outside evaluation unusually difficult. Prodigy Research is asking investors and potential recruits to accept a large return, a clean weekly record and superiority over leading models without the evidence that would normally support those claims. Its next meaningful proof will come from sustained performance across changing market conditions, with enough disclosure to distinguish a durable trading system from a short, favorable run.