How to research trades with an AI spreadsheet agent

Chat assistants are good at explaining trading concepts and bad at doing trading research. Ask one "do favorites win as often as their price says?" and you get a confident paragraph with no data behind it. You cannot check it, and you cannot change an assumption and see what happens.

An AI spreadsheet agent works differently. It has tools to collect data, write it into cells, and build formulas and charts on top. The answer is a workbook, not a paragraph. This guide covers how to get good research out of one, using Jordan as the example.

#How Jordan works

Jordan is a spreadsheet with an agent panel beside the grid.

1. You ask a question in the panel. You can reference a sheet or a range with @ so the agent knows what you mean, and it reads the whole workbook for context.
2. Jordan collects data. For market questions it searches Kalshi's public market data and writes the results to a sheet, one row per market. For a historical study it walks the full history of a series and writes it as one contiguous table.
3. Jordan builds the analysis with ordinary spreadsheet formulas, tables and charts that reference the data.
4. You review a changeset. Every edit is proposed as a set of changes that you can preview cell by cell, then accept in full, accept in part, or reject. Accepted changes can be undone like any other edit.

Jordan is read-only toward markets. It cannot place, modify or cancel orders. If you connect your own Kalshi account, it can read your balance, positions, fills and settlements so you can analyze your own trading. Your credentials stay in a separate service and are never passed to the AI.

#Why "the sheet is the dataset" matters

An AI model can only hold a limited amount of text in its working memory, and it cannot reliably do arithmetic on a list it has memorized. A backtest over thousands of markets cannot be done "in the model's head."

So Jordan puts the data on the sheet and lets formulas do the math. That has three benefits for you:

• Every number is traceable. Click a summary cell and you see the formula and the range it covers.
• Nothing is silently dropped. The row count is right there. If you expected 2,000 markets and see 400, you know before you trust the result.
• Assumptions are editable. Change a fee cell, a price threshold or a bin width, and the analysis updates without asking the agent again.

#Writing good research prompts

The agent does what you ask. Precise questions produce precise workbooks. A good prompt names four things.

A complete example:

Collect settled Kalshi NFL game markets onto a sheet. Bucket them by last YES price in 10-cent bins, and build a calibration table comparing each bin's implied probability with the share that resolved YES, with counts. Add a chart of implied vs. actual and a one-line takeaway above the table.

Then iterate in small steps: "add a ±2 standard error column," "exclude markets with volume under 100," "split the table by season." Each step comes back as its own reviewable changeset.

#How to review a changeset

Accepting changes without looking defeats the purpose. A quick review takes a minute:

1. Check the row count. Does the data sheet have about as many rows as you expected?
2. Check the key column. Is the price column the price you asked for? Is the outcome column 1/0 the right way round?
3. Check one row by hand. Pick a row and confirm the derived columns match what you would compute.
4. Read one summary formula. Make sure ranges cover the whole table and criteria are right.
5. Look for lookahead. Does any input use information that would not have been known at trade time? See How to backtest a trading strategy.

If something is off, reject the change or ask for a fix. Nothing touches your workbook until you accept.

#Questions to start with

These work on public data, without connecting an account:

• "Do Kalshi NFL favorites win as often as their price says?" A calibration study. Background: Market calibration.
• "What does Kalshi think the Fed will do next?" An implied distribution for the next rate decision. Background: How prediction market prices work.
• "What's priced into this week's NFL games?" Moneylines, spreads and totals side by side, with implied probabilities.

Once you have a result you believe, carry it forward: compute the expected value after fees, and size any hypothetical position with fractional Kelly.

#What an agent will not do for you

An agent makes research faster. It does not make conclusions true. It is still your job to:

• decide whether a sample is large enough,
• question whether a historical pattern will persist,
• account for costs you will actually pay, and
• make your own trading decisions.

Jordan is a research tool, not an adviser. It does not recommend trades.

#Frequently asked questions

#What is an AI spreadsheet agent?

An AI spreadsheet agent is an AI model with tools to read and edit a spreadsheet. Instead of answering in text, it collects data into cells and builds formulas, tables and charts, so the result can be inspected and modified like any other workbook.

#Can Jordan place trades on Kalshi?

No. Jordan is read-only. It can read public market data and, if you connect your own account, your portfolio data. Trading operations are not available through Jordan at all.

#Does Jordan need my Kalshi account?

No. Public market data, including markets, events, series and price history, works without an account. Connecting your account adds read access to your own balance, positions, fills and settlements.

#Can I export my Jordan workbook to Excel?

Yes. Jordan workbooks can be exported as .xlsx files, and formulas use familiar spreadsheet functions.