Gedanken
Run the thought experiment before you risk the capital.
Gedanken tests market research in thousands of simulated markets where the right answer is planted in advance. It finds out which methods find real edges and which ones fool themselves, then rebuilds its own methods under a judge it cannot change. The engine is general; each market gets a domain pack built from its own causal story. Two packs run today: stock factors, and the chain from natural gas to fertilizer producers' earnings. Power and AI data centers come next.
Can your factor research tell alpha from luck?
We generated 300 stock markets (200 stocks, 10 years of monthly returns, fundamentals reported one to two months after each quarter) with no edge at all, and 300 where value and quality carry a real, planted premium.
A typical search tries 76 factor variants and keeps the best. In markets with no edge, it claims one 99% of the time. Two mistakes compound: trying many variants, and stamping fundamentals at quarter end, before investors could see them.
The disciplined test commits to one hypothesis, on data as investors saw it. It claims a false edge in 4% of those markets and still finds 90% of the real ones.
Switch between the four approaches. The dashed line is a 5% tolerance for false discoveries.
When everyone asks the same AI, follow it or fade it?
More investors now ask LLMs for views. Research shows that early on, LLM reads of the news predicted returns, and that the edge shrank as adoption grew. Separate work finds LLM-guided retail trading moves together and then reverses.
Gedanken simulates that crowd. Slide adoption up and watch following the LLM stop paying while fading it starts to pay. A fundamental signal, hedged when its view matches a crowded LLM view, holds up throughout.
The crossover depends on assumptions about how much the crowd moves prices, so the real level has to come from live data. That is what the daily consensus log is collecting.
A research system that improves itself and cannot grade its own homework
An LLM, the Architect, proposes one change at a time to the research method. Sentinel, a fixed judge, scores each change on 1,000 fresh markets with no edge and 400 with one, and accepts it only if false discoveries stay under 5% and the gain is significant. The bar rises with every attempt, so tinkering gets harder.
Sentinel's first version had a little slack. The loop found it immediately and accepted the one change that spent it. Sentinel was tightened, by a human, and the loop then found a real improvement: sizing positions by signal strength raised detection from 25% to 33% with 2% false discoveries.
Sentinel's code is hash-locked. The loop refuses to run if it changes without a human signature.
| # | Change the Architect proposed | Verdict |
|---|
Déjà Vu: noticing when the real world leaves the simulation
Simulations are only useful while they resemble the market. Déjà Vu fingerprints the market every two weeks and measures how far it sits from the simulated worlds. To test it, we built 22 years of synthetic market data with a hidden structural break: from August 2021, European gas decouples from US gas and spikes, the way it did in the real 2021 to 2022 crisis.
The general alarm caught the break about 11 months in, because the shock built slowly. That failure became a permanent, targeted check on producers that should move together, which caught it in 196 days with no false alarms. Synthetic data with a known answer; no real-market claims.
Questions I like to dig into
Brian Curry founded KC AI Lab in 2016, led data science and ML work at Koch Ag & Energy Solutions, Tractor Supply and Vail Resorts, and runs Vector1. Gedanken is how he works: build fast with AI, then prove it with evidence that does not depend on trusting the code. Vector1 is a non-commercial research lab. If one of these questions is on your mind too, he is happy to compare notes.
How often does a research process fool itself?
Known-answer tests for signals and models: how often a process invents alpha, and how often it misses a real edge.
What happens when everyone asks the same AI?
How views that millions of people get from the same AI tools move prices, and when that crowd is worth following or fading.
How do you build AI foundations before the playbook exists?
Forecasting, causal inference and gated agent workflows: building early, proving it with evidence, and handing it over cleanly.