One advisor for the whole world?

Run your own experiment

The stated-investor arm of the paper adds one sentence about the investor to the portfolio question and measures how the recommended stock share moves (section 6). Here you can write your own sentence, for example about future income, and download a script that runs the comparison on OpenRouter with your own API key.

The script uses the paper's prompt, decoding settings and response parser. Nothing is sent from this page. Your key stays on your computer and your results are not collected.

1. Countries, models and repetitions

The investor arm is in English for every country. The stake is the paper's local-currency amount and the country sentence names the country.

Countries
Models

The paper's slugs are the defaults. OpenRouter retires slugs over time, so any slug listed at openrouter.ai/models works too. The paper uses age 40 and fifty repetitions per cell.

2. Arms

Each arm is a version of the prompt. Sentences go after "I live in [country].", where the paper places its statements, one sentence per line. The first arm is the reference that every other arm is compared with. Leave it empty to compare with the paper's plain prompt.

3. Preview

Grey: the Instructions block with the JSON schema, identical in every cell. Highlighted: your sentences.

4. The script

To run it you need Python 3.8 or later, the requests package and an OpenRouter key:

pip install requests
export OPENROUTER_API_KEY="sk-or-..."
python run_experiment.py --print-prompts   # check the prompts, no calls
python run_experiment.py --dry-run         # fake responses, no calls
python run_experiment.py                   # the real run, asks before spending

Responses go to experiment_results/results.csv. An interrupted run resumes where it stopped. At the end the script prints each arm's mean recommended stock share by model and its difference from the reference arm, by model and country, with standard errors.

Show the script

Before you read your results

  • Compare arms within your own run. Models behind a slug change, and the tracking waves show that the level of advice moves between collections. Your reference arm, run in the same session, is the right comparison, not the paper's numbers.
  • Responses are noisy. At temperature 0.7 the recommendation varies from draw to draw. With ten repetitions per cell, differences of a few percentage points are within noise. The paper uses fifty.
  • Wording matters. A sentence can shift advice through its tone as well as its content. The paper pairs each statement with its opposite (risk-tolerant and risk-averse, generous and limited pension) so that tone alone cannot explain the gap. Writing a pair is a good habit.
  • It is exploration, not replication. The paper's estimates use wild cluster bootstrap inference over thirteen models. The script's standard errors are simple and rough.