# Year-long experiment: Claude researches the news before every market open and trades a $1,000 paper portfolio against the S&P 500. Here's Day 1 — retained record

> Retained reddit record from reddit:singularity.

Canonical: https://highsignal.app/data/records/153c1251708e28d8

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:singularity)  published 2026-09-29 23:11 ingested 2026-09-30 02:41

[original source ↗](https://www.reddit.com/r/singularity/comments/1wtp4tr/yearlong_experiment_claude_researches_the_news/)

## retained text



```
Most actively managed funds trail the S&P 500 over the long run. So I wanted to see how an AI agent does at the same job, with real research, over a full year, against a scoreboard that can't be argued with.

**The setup**

* Claude (Sonnet 5.5 at high effort, on a normal Claude Pro plan) manages a **$1,000 paper portfolio for 365 trading days**.
* Every weekday **before the open** it gets its portfolio, a one-line price screen of all 503 S&P 500 stocks, and its own last 10 days of decisions. Its only tools are **web search and web fetch**: no code, no files.
* It reads the overnight news and places up to 5 orders. They fill at the **opening price**, which it can't see when deciding.
* After the close, the portfolio is valued and compared with **$1,000 of SPY bought at the same open and held**.

**The rules** (enforced in code, not just in the prompt)

* Individual S&P 500 stocks only. No ETFs, options, leverage or shorting.
* Max 5 trades a day and max 30% in any one stock. Orders that break a rule are rejected and logged.
* Every buy has to be backed by something it read that morning, and it cites its sources.
* Doing nothing is allowed. Overtrading is the classic way to lose to the index.

**Day 1**

|Value|Today|Change|
|:-|:-|:-|
|**Claude**|**$1,000.79**|**+0.08%**|
|SPY buy & hold|$996.57|−0.34%|

It bought four starter positions and kept 56% in cash:

* **NFLX $100** (−0.1% by the close): down 44% from its high; Deutsche Bank upgraded it to Buy at about 18x 2027 earnings.
* **TMO $120** (+0.5%): Thermo Fisher raised its 2026 guidance; a steady compounder if rising rates hit the market.
* **GOOGL $120** (−0.3%): bought the \~15% pullback on AI-spending worries.
* **XOM $100** (+0.7%): oil above $100 with the Strait of Hormuz shut; also a hedge for the growth names. Kept small because a peace deal could reverse it.

Its take on the day: *"Stocks fell on Monday … as the 10-year Treasury yield hit 5.23% (highest since 2007), oil stayed above $100, and markets priced roughly a 70% chance of an October Fed hike. I'm holding about 56% in cash while building the portfolio in stages."*

One day means nothing, of course. I'll post a weekly update with the chart, the trades and Claude's reasoning, including the bad calls.

**FAQ**

* *Real money?* No. It's paper trading. Not financial advice, and I don't trade these picks myself.
* *Lookahead?* It decides before the open and fills at the open. It never sees a price after its decision.
* *Can I see the prompt and the rules?* Yes, ask in the comments. Every decision, source and fill is timestamped in a git log, which I'll publish in full at the end.

**Prediction time: after 365 trading days, does Claude beat SPY, and by how much?**
```



document fetched 2026-09-30 00:33 · canonical https://reddit.com/r/singularity/comments/1wtp4tr/yearlong_experiment_claude_researches_the_news
