Man Group's Head of Quant said something that stuck with me:
"The challenge is the sheer volume of data and possible market relationships that has grown faster than any human team can evaluate by hand."
So they built AlphaGPT. It generates signal hypotheses, writes the code, and runs the backtests. Autonomously. Hundreds of ideas per week instead of 20 per quarter.
Bridgewater went further and built a $2 billion fund where AI makes the primary trading decisions.
Jane Street spent $6 billion on GPU infrastructure last year to train proprietary models.
I'm not going to pretend I know exactly what's running inside these systems. But the public statements from the people building them tell a fairly consistent story and it's not the one most people assume when they hear "AI trading."
The firms winning aren't replacing their quants. They're making each quant about 10x faster.
This article is the complete framework for running the same architecture on Polymarket today.
PART 1 - WILL AI REPLACE QUANTS?
The question everyone asks wrong.
Man Group went public with AlphaGPT in July 2025. The system generates signal hypotheses, writes implementation code, and runs backtests autonomously. Several dozen signals have already been approved for live trading after passing human review.
The challenge in quantitative investing is the sheer volume of data and possible market relationships that has grown faster than any human team can evaluate by hand.
A strong research team might seriously test 20 signal ideas in a quarter. AlphaGPT tests hundreds in a week.
But not a single signal from AlphaGPT touches real capital without a researcher making a deliberate decision about it.
Bridgewater built an AI Reasoning Engine combining LLMs, machine learning, and reasoning tools. Their co-CIO called it "a big jump." But humans still oversee risk management and execution.
Citadel's CTO said it plainly: "We don't want PMs offloading their human investment judgment to AI."
Ken Griffin himself said AI boosts efficiency but is unlikely to produce market-beating returns on its own.
The firms winning are making their quants 10x faster. Not replacing them.
PART 2 - FIVE USE CASES WITH REAL EDGE
USE CASE 1: AGENTIC SIGNAL DISCOVERY
Man Group's AlphaGPT runs four agents in a loop:
- Agent 1 generates a signal hypothesis.
- Agent 2 writes the implementation code.
- Agent 3 acts as pure challenger - finds every reason the signal might be fake or overfitted.
- Agent 4 evaluates the backtest and decides whether to send it to human review.
On Polymarket this maps directly:
- Agent 1 generates a probability estimate from news, related markets, and base rates.
- Agent 2 compares to current market price.
- Agent 3 challenges: what would have to be true for this to be wrong?
- Agent 4 evaluates EV and sends go/no-go to the human.
USE CASE 2: ALTERNATIVE DATA EXTRACTION
For prediction markets, every statement from a Fed official, every geopolitical development, every economic data release contains signal. The AI converts unstructured text into a structured probability shift.
USE CASE 3: MONTE CARLO SIGNIFICANCE TESTING
Standard backtesting uses one path through history. One path is not enough.
USE CASE 4: REGIME-AWARE POSITION SIZING
f_adjusted = f_kelly x regime_factor x (1 - drawdown_factor)
USE CASE 5: DEPLOYMENT MONITORING
PART 3 - THE COMPLETE PIPELINE
Start here if you're not on Polymarket yet: polymarket.com/?r=atlas
$28 billion traded. 9,000+ markets. Every resolved contract is a ground truth data point for your model.
6 stages. 5 automated. 1 always human.
Stage 1 - Data ingestion: historical resolution rates, price time series, related market correlations, volume metrics.
Stage 2 - Signal hypothesis: specific, testable, with economic rationale and the conditions under which it fails.
Stage 3 - Adversarial challenge: a separate agent whose only job is to break the hypothesis before any time is invested building it. Man Group calls this the most valuable part of AlphaGPT.
Stage 4 - Walk-forward backtesting: every parameter estimated using only data available at trade time. This single requirement eliminates the most common source of inflated backtest performance.
Stage 5 - Monte Carlo significance testing: if your signal sits in the top 5% of 10,000 random alternatives, you have evidence of real edge.
Stage 6 - Human review gate: cannot be automated. Write down three conditions that will make you stop and review the system before you start.
PART 4 - BEFORE AI VS AFTER AI
Before AI:
An idea came from reading or observation. Writing implementation took hours or days. Setting up proper backtesting took more time. A researcher might seriously test 20 strategies per year. Position sizing was calibrated by intuition.
After AI:
Time between idea and rigorous evaluation compressed from days to hours. You run adversarial review on your own hypotheses before investing any time building them out. You test 12 variations of a promising signal and evaluate all of them rather than picking one by intuition.
Man Group described this precisely: the technology helps them test more ideas. Researchers spend time evaluating signals that have already survived automated challenge rather than spending that time on implementation work.
For Polymarket specifically, the compression is even more valuable. Markets resolve on fixed dates. The window to enter at a good price is finite. The faster you go from hypothesis to validated signal, the more opportunities you actually capture.
THE SUMMARY
AI does not predict markets.
It compresses the time between a trading idea and a rigorous test of that idea from days to hours. It runs adversarial review that most systematic traders never apply to their own hypotheses.
Man Group: the LLMs have accelerated the pace of change. But their quants are still there. Every signal that reaches capital has had a researcher sign off.
Jane Street invested $6 billion in GPU infrastructure to multiply what their researchers can do. Not to replace them.
The AI gave them scale. The judgment stayed human.
The edge in prediction markets right now is not better information.
It is testing more ideas faster than everyone else and only acting on the ones that survive adversarial review.
That's the whole system.
