Skip to main content
Give Itoflow a signal idea, or a dataset of your own, and a universe. The research agent writes the test in Python, runs it on data as it stood at each date, and tells you whether the signal held up out of sample. When it finds no edge, it says so. What you’ll get: information coefficients by month, in-sample and out-of-sample results, overfitting checks, a long-short backtest with costs, and the code behind each number.

Try it

With Itoflow connected to your assistant, paste this into your assistant:

Open in Claude

Start this in Claude with the Itoflow plugin.

Open in ChatGPT

Start this in ChatGPT with the Itoflow plugin.
Using another client? Copy the prompt with the button on the code block.To test your own dataset, keep it in a connected GitHub repository and name the file in the prompt, or attach it in the Itoflow app.

What a result looks like

We ran the prompt above on October 28, 2026. The agent fixed the method first: 98 months to train, from January 2015, and 43 months to test, from March 2023 to September 2026. Its answer:
  • In the test months, the mean rank information coefficient was +0.034, with a t-statistic of 1.28. Its 95% interval ran from -0.018 to +0.087, so it included zero. 53.5% of test months were positive.
  • After costs, the long-short decile portfolio returned 30.7% over the test months, with a Sharpe ratio of 0.39 and a deflated Sharpe ratio of 0.77. It gained 42% in June 2026 and lost 35% in July.
  • Only 89 to 98 of the stocks had a full year of prices, so each decile held 8 or 9 stocks. The agent said so rather than fill the gap.
  • It judged momentum not confirmed out of sample and did not retune the signal after seeing the test. It named the hindsight in the universe: today’s largest stocks leave out the large caps that failed or shrank, so the test months are held out for the signal but not for the choice of stocks.
What our quant-research agent actually does — and where it still needs a human, July 9, 2026, shows an open request: find and validate a signal for 12 US large caps. Out of sample, neither 6-month momentum nor short-term mean reversion passed, and the agent reported no validated edge. It didn’t flip the signals or switch to a universe where they looked better.

How it’s checked

  • The method is fixed first. The agent writes a research contract with the universe, dates, split and evidence before the test runs, and the system checks it.
  • The out-of-sample result is worked out again. The system recomputes the out-of-sample choice itself rather than trust the agent’s claim.
  • Overfitting checks are code. The deflated Sharpe ratio, the probability of backtest overfitting and false-discovery-rate control run in the engine.
  • A second agent reviews it. An independent reviewer reads the plan and the files, and can send the work back.

Limits

  • No tool returns an index’s past members, and the agent can’t yet pick a universe from data known at each past date, such as the 500 largest stocks then. Asked for one, it says it can’t build that universe rather than use today’s list. With a fixed list, as in the prompt above, it states how hindsight in that choice could bias the result.
  • Data covers stocks and ETFs on 24 exchanges, plus crypto. Futures, options and individual bonds aren’t covered.
  • A file attached in the app can be up to 5 MB. Keep larger datasets in GitHub or serve them from your own MCP server.
  • A signal that passes is a research result, not a forecast.

Next steps