How to use Claude to test market context briefings
A developer compared Claude trading decisions using current briefings, prices, web search and stale briefings. The latest reported run favored current context, but the author highlighted bearish-window bias and inconclusive sample size.
news.ycombinator.com
Prompt to try
Editorial adaptation, not the source author’s exact prompt. Replace bracketed inputs and review tool permissions first.
Design a controlled comparison of Claude decisions over [historical market window]. Use four arms: [current briefings], [price-only data], [web search], and [time-shifted briefings]. Record one decision per tick, returns, drawdowns and problematic decisions. Report sample limitations and possible future-data contamination.
Reach people exploring AI agents
See placements →What happened
In the latest reported six-month run, current briefings returned +7.83% versus -8.14% for prices only. The author attributes the edge mainly to defensive short campaigns and states that bull-market performance remains untested.
Reported by KalskiTheDan on Hacker NewsHow the agent did it
- Pre-compute market analysis into short markdown briefings.
- Set up current briefing, price, web search and stale briefing arms.
- Run Claude for one trading decision per tick in every arm.
Where a person stepped in
The author designed and compared the experiments and aborted one run at tick 33 after an inappropriate short decision.
A person approved each step before the agent continued.
A solo developer built a financial context API to give AI agents pre-analyzed market briefings covering on-chain information, macro conditions, and regime classification. Instead of returning raw price JSON, the developer pre-computed analysis on the server and supplied markdown briefings of roughly 400 tokens. The aim was to reduce the preliminary interpretation an agent needed before making a decision.
The author then compared four experimental arms. One received current briefings, another received only prices, a third used web search, and a placebo received briefings shifted five to seven months while presented as current. All arms used Claude and made one trading decision per tick. In the latest reported run, 202 ticks covered six months: the briefing arm returned 7.83%, compared with a loss of 8.14% for the price-only control.
The author says the advantage was mainly defensive, with the treatment's return coming from two short campaigns during crashes. Rallies and sideways markets showed matching or weaker performance. Earlier problems included repeated buying and selling without a net position change and an aborted run after bearish context prompted an inappropriate short position.
The author explicitly cautions that bull markets remain untested, a single run is not statistically conclusive, and web search risked future-dated contamination.
Who it's for: Researchers comparing agent context formats who can track decision behavior and experimental limitations.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Hacker News · KalskiTheDan
Frequently asked questions
What were the four experimental inputs?
The arms received current market briefings, price data only, web search, or briefings shifted five to seven months and presented as current. All arms ran Claude with one decision per tick.
Did the author establish a general trading advantage?
No. The author reports mainly defensive gains in bearish windows, says bull markets are untested, and calls 202 ticks insufficient for statistical conclusiveness within one run.
What problems appeared in earlier runs?
The author reports repeated BUY and SELL decisions without net position change. Another run was aborted at tick 33 after the agent interpreted macro bearish context as a reason to short rather than hold cash.
More details
Remaining steps
- Compare returns, drawdowns and decision behavior across runs.
About this account
Claude made one trading decision per tick in each experimental arm using the context supplied to that arm.
Platform: Hacker News. Attribution: https://news.ycombinator.com/item?id=47530186. Prompt: Suggested (written by Agents In Use, not the author's). Sourced by the Agents In Use pipeline.
Credit: KalskiTheDan on Hacker News
Source published 26 Mar 2026
Source checked 14 Oct 2026
Tools and topics
Claude · Financial context API · Web search
Agent evaluation · Market context · Experiments
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.