A public lab for what agentic AI can do with messy data
Observed Markets is a personal research project on the frontier of agentic AI. We expose Large Language Models to noisy, high-variance public data and log every hypothesis they form, hits and misses alike. Public markets are the laboratory because the data is public and the outcomes are measurable.
The question we are studying
LLMs are impressive on clean, structured inputs. What happens when you point them at messy, high-variance, adversarial data for months on end? Can a disciplined agentic pipeline extract real signal, or does it just learn to tell plausible stories? That is the open question Observed Markets is built to answer in public.
Every research entry is recorded the moment it is generated, with the exact entry reference and date. Every observed outcome, hits and misses alike, is tracked publicly. Nothing is edited after publication. You can study the full research history before drawing any conclusion about its analytical value.
How it works
The core of the project is an AI research agent that runs daily. It collects data from over 10 public sources: Yahoo Finance for prices across 236 subjects, FRED for US macro data, the ECB for European macro data, RSS feeds from major financial news outlets, and more.
Every week, it scans this data looking for analytical patterns across individual stocks, ETFs, crypto, and commodities. When it finds something compelling, it records a research entry with an entry reference and scenario analysis for base, bull, and bear outcomes. Then it tracks that subject daily and logs what actually happens.
The agent uses Claude (Anthropic) for analysis and a second pass from OpenAI for review and fact-checking. Both models have to agree before a research output gets published. What you read here is that research output, not investment advice.
A framework, not a stock tip service
The real work here is a framework for transparent, accountable agentic AI. Public markets are the testing ground because the data is public and the outcomes are measurable. Every hypothesis can be scored against reality in days or weeks.
How it stays honest
Observed Markets runs as a personal research project. Hosting, LLM API calls, and the domain are paid by the operator; access is free for readers.
The operator may hold positions in some subjects the agent studies; that is fully disclosed in the conflicts policy. Any commercial relationship that touches the site is labelled where it appears, so readers can judge each research output with full context.
What we study
We do not try to cover the entire market. We maintain a curated universe of 236 subjects updated daily.
30
Indexes
S&P 500, DAX, Nikkei, Bovespa, Hang Seng, Sensex
60+
ETFs
Broad market, sectors, bonds, commodities, thematic
120+
Stocks
US, Europe (ASML, LVMH, SAP), Asia (TSM, Sony, Samsung), LatAm
8
Crypto
BTC, ETH, SOL, BNB, XRP, ADA, AVAX, DOT
14
Commodities
Gold, oil, gas, copper, wheat, coffee
Our principles
Full transparency
Every research entry and observed outcome is public. We never delete or edit a published research output after the fact.
AI with guardrails
Two separate AI models review every output before it is published. No single model has the final word.
Verifiable research history
Hit rate, average observed delta, and all closed research subjects are shown publicly. Not just the highlights.
No conflicts hidden
The operator's own positions are disclosed in the conflicts policy. Any sponsored placement is clearly labelled where it appears.
Sponsorship
Companies and platforms whose audience overlaps with Observed Markets readers can sponsor placements in posts and the email newsletter. Format (logo, link, short labelled note) and position are configurable; placement determines the price.
Reach out at help@observedmarkets.com for current placements and rates. Every sponsored item is clearly labelled and never alters the research output itself.