How to test stock recommendations
The GPT Investor generated stock recommendations with web-searching agents, executed 19 real-money experiments, and reported returns against SPY.
news.ycombinator.comSuggested prompt
Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.
For [research experiment], use web research to propose stock recommendations with an explicit [holding period] and documented rationale. Record sources and methodology. Build a tracking plan comparing [portfolio] with [benchmark] over matching periods; leave trade execution to me.
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source author
- ishtiaqrahman
- Original account
- Read the source ↗
- Published
- Source link checked
- Reported result
- The author reported 19 experiments investing CAD 5,700, with an 11.54% portfolio return versus 8.89% for SPY over the experiment.
- Human involvement
- The author executed trades, closed positions, and published methodologies and results. A separate review of recommendations is not described.
A linked account supports attribution. It does not mean we independently reproduced its result.
Report an issue with this case
Open correction emailOpens your email app with the case link included. Nothing is sent automatically.
How the agent did it
- Generated duration-specific stock recommendations with web-searching LLM agents.
- Allocated CAD 300 to each of 19 experiments through Wealthsimple.
- Closed positions at the end of their specified holding periods.
An investor ran a year-long experiment called The GPT Investor, using LLM-powered agents to generate stock recommendations and comparing the resulting portfolio with SPY. The experiment began on May 16, 2023, and involved real trades on the NYSE and NASDAQ rather than paper trading.
The author tried different methods and technology stacks, using GPT-3.5 and GPT-4 through platforms including ChatGPT, Godmode, and BabyAGI UI. The common process enabled autonomous web searches and asked for stock recommendations over specified holding periods. The author documented the methods and results for subscribers.
Across 19 experiments, the investor allocated CAD 300 to each, totaling CAD 5,700. Trades were executed through Wealthsimple. Positions were closed at the end of their specified durations, and the returns were compiled into a combined portfolio. The experiments ranged from seven days to one year and produced recommendations for 31 stocks.
The author reported an 11.54% portfolio return, or CAD 658.20, compared with 8.89%, or CAD 507.10, for SPY. These are the author's results for this experiment; the supplied text does not establish future performance.
Who it's for: Researchers examining how agent-generated recommendations can be tracked against a benchmark with explicit holding periods and recorded trades.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Hacker News · ishtiaqrahman
Frequently asked questions
Did the agents execute the trades themselves?
The source says agents generated recommendations using autonomous web searches. The author used a Wealthsimple brokerage account to execute trades and closed positions at their specified endpoints.
What returns did the author report?
The author reported 11.54% for The GPT Investor portfolio versus 8.89% for SPY across the experiment, with CAD 5,700 invested in 19 allocations of CAD 300.
More details
Remaining steps
- Compiled portfolio returns and compared them against SPY.
About this account
The agents autonomously searched the web and generated recommendations; the author handled brokerage trades and tracking.
Platform: Hacker News. Attribution: https://news.ycombinator.com/item?id=40493026. Prompt: Suggested (written by Agents In Use, not the author's). Sourced by the Agents In Use pipeline.
Credit: ishtiaqrahman on Hacker News
Source published
Added to Agents In Use
Source checked
Tools and topics
GPT-3.5 · GPT-4 · ChatGPT · Godmode · BabyAGI UI · Wealthsimple
Investing experiment · Stock research · Benchmarking
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.