Hook
On May 15, 2025, a Y Combinator S26 batch startup named Prodigy Research published a self-reported performance sheet: 108% net returns over two months, zero losing weeks, and a claim that its AI trading agent had outperformed the top 10% of Jane Street traders. The S&P 500, during that same window, moved less than 2%. The code never lies, only the auditors do—and here, there are no auditors. The numbers are a statistical anomaly begging for a forensic dissection.
Context
Prodigy Research is the rebranding of Prodigy AI, a project that previously focused on building autonomous trading agents for prediction markets like Polymarket and Kalshi. The company is helmed by brothers Michael Wang (former Jane Street trader and Google DeepMind researcher) and Yuhua Wang (Apple AI engineer). Their YC backing is the primary credibility signal, but the project’s entire narrative rests on a single, unverified data point: two months of 108% returns with no weekly drawdowns. No third-party audit, no strategy code, no capacity disclosure, no leverage ratio. Just a spreadsheet and a press release. This is the classic pattern of a high-risk narrative project—complexity is just laziness wearing a tech suit.
Core: The Forensic Teardown
Let’s start with the math. A 108% return in two months implies an annualized return exceeding 1,000%. For context, Renaissance Technologies’ Medallion Fund, the most successful hedge fund in history, averaged 66% annualized before fees over 30 years—and it operates under strict secrecy, audited returns, and billions in capacity. A 1,000% annualized return, especially in a flat market, is statistically impossible unless one of three conditions holds: extreme leverage, a one-time arbitrage opportunity, or selective reporting.
The claim of “zero losing weeks” amplifies the red flag. Every trading strategy—even market-making firms with perfect hedging—experiences periodic drawdowns due to liquidity gaps, technical glitches, or fat-tailed events. A two-month streak of positive weekly returns in a volatile asset class like prediction markets is not just improbable; it’s a mathematical tell. Based on my audit experience during the 2022 LUNA collapse, I learned that systems claiming flawless performance are either under-reporting risk or over-reporting returns. The Terra crash was a math error, not a market crash. Prodigy’s “no losing weeks” is the same species of math error.
The Strategy Conundrum
Prodigy’s stated focus is Delta-neutral strategies in prediction markets. Delta-neutral means the portfolio is hedged against directional price moves, earning returns from volatility, funding rates, or basis spreads. In a two-month window where the underlying index barely moved, the theoretical maximum from pure Delta-neutral strategies is in the single digits—certainly not 108%. The only way to achieve such returns is through extreme leverage on funding rate arbitrage or by exploiting tiny pricing inefficiencies in illiquid markets. Polymarket’s total trading volume peaked at $2 billion during the 2024 elections and has since declined sharply. The capacity for a large-scale strategy is minuscule. Prodigy has not disclosed its principal size, but if it is under $1 million, the 108% return is economically insignificant. If it is over $10 million, the market impact would have been observable on-chain—and it wasn’t.
The Verification Gap
Prodigy’s technical claims are equally unverifiable. They allege they have trained “the world’s most powerful quantitative finance foundation model,” yet no benchmarks, model weights, or inference latency data are provided. They claim to have beaten “Claude Fable” and “GPT-5.6 Sol” in trading tasks, but these are not standardized benchmarks. Comparing LLM performance to trading profitability is a category error. The model’s ability to generate coherent text has no proven correlation with edge detection in stochastic markets. The code never lies, only the auditors do—and in this case, no one has looked at the code.
Contrarian: What the Bulls Got Right
To be fair, the team’s background is genuinely impressive. Michael Wang’s combination of Jane Street (top-tier market making) and DeepMind (frontier AI research) is rare. The brothers’ tight collaboration likely enables fast execution. The prediction market niche is also a legitimate arena for AI-driven strategies: odds are public, information asymmetry is high, and the market structure is inefficient enough to reward faster data processing. YC’s screening process, while not a technical audit, does filter for founder quality and market timing. The bulls would argue that the returns, while unverified, are plausible given the team’s edge in information processing and the nascent state of prediction markets. They might point to Polymarket’s rapid growth and the increasing role of AI agents in DeFi as a tailwind.
However, the counterpoint is that the same bulls are ignoring the statistical impossibility of the “no losing weeks” claim. The lack of any independent verification—even a simple on-chain trade history—is a deliberate choice. If the strategy were real, a transparent audit would attract more capital. The fact that they haven’t published one suggests either the returns are a one-time anomaly or the strategy is not scalable. Forensics reveal the truth markets try to bury.
Takeaway
Prodigy Research is a fascinating case study in narrative engineering. The team has all the right credentials, the right accelerator, and the right buzzwords. But the data does not support the story. A 108% return with zero losing weeks, unverified, in a flat market, is not a trading victory—it’s a statistical outlier that demands explanation. Until they release audited trade logs, a transparent strategy outline, and a third-party validation, this project belongs in the category of “sophisticated hype.” The market will eventually correct the narrative, and when it does, the cost of trust will be paid by those who bet on the story rather than the code. Complexity is just laziness wearing a tech suit.