How to ask for a backtest

You can start with one short sentence, and the Agent will fill the gaps with defaults it lists for you. You get to a result you trust faster, though, when your first message already says the things a trader would say to another trader.

What to put in your first message

Say this	Example
Market	ES, NQ, MNQ, SPY, AAPL
Bar size	1 minute, 5 minute, 15 minute, daily
Direction	long only, short only, both
Entry	”RTH opens above the prior day’s VAH, price touches the VAH, long on the first bullish 15-minute close”
Stop	”below the low of the entry candle”, “20 points”, “1 ATR”
Target	”2R”, “40 points”, “the prior day’s high”
When to be flat	”flat by 16:00 ET”

Add these only if they are part of your plan:

Filter: a condition that must also be true, such as “only on negative gamma days” or “only above VWAP”.
Time limits: “no entries after 11:00 ET”, “skip the first 5 minutes”.
Size: “2 contracts, the first exits at 1R and the second at 2R”. Without it, a future trades 1 contract.
Dates: “last 2 years”, “2022 to today”. Without them, the test uses all the history we hold.

A complete first message, ready to copy:

ES 15 minute, long only. If RTH opens above the prior day's VAH, wait for price
to touch the VAH, then go long on the first bullish 15-minute close. Stop below
that candle's low, target 2R, no entries after 11:00 ET, flat by 16:00 ET.
Only on negative gamma days.

What happens next

The Agent reads your rules back

You see a card with each rule. Anything you did not say is marked Assumed, with the default it used. Check this card first: it is what will run.

A history check runs

Before the backtest, the Agent counts how often your entry happened in the past and how price moved after it. It takes about a minute and is not charged. The check counts the entry trigger on its own: a filter such as “only on negative gamma days” or “RVOL above 1.5” is applied in the backtest, not in this count, and the Agent says so. For a gamma filter, the “Performance by gamma regime” card shows the same signals split by positive and negative gamma.

You approve the run once

One approval card shows what will run and its price in credits. Nothing runs or is charged until you approve it. If you ask again while that card is open, you get the same card back, not a second one. A card you leave unanswered expires after about 10 minutes and stays in the chat as an Expired receipt, so you can see it lapsed. Nothing ran and nothing was charged; ask again to get a new card.

The result arrives in the chat

It usually takes 2 to 6 minutes. You can leave the chat open or come back later.

Opening range breakouts

Say “ORB” or “opening range breakout” and name only what you care about. Whatever you leave out is filled the same way every time and listed as Assumed:

You did not say	What is assumed
The bar size	5 minute bars. In “15 min opening range”, the 15 minutes is the length of the range, not the bar size.
How many trades a day	One.
When to be flat	15:55 ET.

Anything you do state is used as you wrote it, for example “on the 15m chart”, “up to 3 trades a day”, “flat by 15:45” or “hold overnight”. Ask for a different bar size or cap in one sentence, such as use 15 minute bars or allow 2 trades a day.

Change one thing at a time

After a result, say only what changes. The Agent keeps every other rule exactly as it was.

does it improve if there are no entries after 11am?
same 11am cutoff but no gamma filter
make the stop 1 ATR instead
go back to the 11am version and rerun it exactly

You do not need to repeat your entry. If you remove a filter, the Agent says what was removed and what stayed the same. To compare with an earlier version, ask it to restore that version: it reruns the same rules and returns the same trades.

What to expect from the answer

The tested window. The result names the dates it actually tested. Rules that read gamma start where our gamma history for that market starts (for ES and NQ, January 2024), and the result says so.
A stop at a level is worked out from your words. “Below the 2nd swing low of the last 20 bars, 2 ticks of margin” is computed on every entry. If the run could not work it out, the answer says so plainly and that your filter was not the cause.
Every trade keeps its stop. A trade is never left open without the stop you stated. A limit entry is not placed when price is already past its stop, is cancelled when price trades through the stop before it fills, and if it fills through the stop it is closed at once as a stopped trade. The result counts how many entries this affected.
A filter that let nothing through is read again once. If your filter’s code let none of the signals through, it is rewritten once from your words (never loosened). When that rewrite produces the result, the result says so, and a rerun can read the filter differently.
Write your plan in your own language. “Sin stop por precio” or “Ninguno” is read as no stop or no filter, the same as in English, and an exit condition you wrote runs in place of a stop.
A run that runs out of time is tried on the most recent half. A very long window can take longer than one run allows. When nothing could be kept, the Agent offers the same rules on the most recent half of the window, raises the card for it, and says that is what the result covers.
Too few trades to judge. Fewer than about 30 trades is a sample, not an edge. The Agent says so instead of calling the result good or bad.
Where each signal went. If the history check found 230 signals and the backtest took 21, the result explains the other 209: your filter, your time cutoff, a position already open, and so on.
Assumptions said out loud. Every default the Agent chose is listed, so you can change it in plain words.
No invented results. If something cannot be tested exactly as written, the Agent says what it did instead and asks before it guesses.

When something looks wrong

0 trades, or far fewer than expected: ask where did the signals go?. The answer names the rule that removed them.
A flat stretch in the equity curve: ask how many trades per year?. A year with none usually means a filter or the data had no value for those days.
The rules read differently from what you meant: correct them in one sentence and run again. You do not need to start a new chat.

See Backtests for the full list of what rules can say and how each number is defined.