Reads your chart
Draw your levels and boxes as you always have. The agent reads them as prices, writes the rules they imply and draws its trades back.
Describe an idea in plain English, or draw it on a chart. The agent writes it in Python, backtests it on CME futures data since 2010 and runs the checks most results fail.
No card needed. Free credits every month.
Draw your levels and boxes as you always have. The agent reads them as prices, writes the rules they imply and draws its trades back.
Pine Script compiled from the tested file, and a port for NinjaTrader or any other platform on request.
Results split by the morning's range, trend, session and news days, and the one filter the numbers argue for.
Every study, file and run is a row in the chat, with what it read and how long it took. Nothing happens off the record.
A demonstration on a simulated ES market, not a result.
Every winner goes through the protocol institutional quants use, in order, with the last two years of data locked until the final test. A result that only fits its own data does not get through.
Validation report
ES opening range breakout
52 trials, 2016 to 2025
Fit on two years, trade the next six months, roll forward. Only the traded months count, and they kept 83% of the fitted Sharpe.
Every way to hold out two of six groups, the days at their edges dropped so nothing leaks. Five paths through the data instead of one, every one positive.
Eight years cut in half 12,870 ways. How often the best setting on one half ranked in the bottom half of the other.
The winner's Sharpe against the best that luck reaches in 52 tries, every failed idea counted. The chance the edge is more than that.
A year rebuilt 10,000 times from the out-of-sample days: the drawdown to plan for, and how often a year loses.
The last two years, locked before the first look and opened once, at the end. Nothing was tuned on them.
Fit 2 years, trade 6 months, 12 times
Stop 32, target 56 chosen in 9 of 12
A demonstration on simulated data: the research log's 52 trials over ten years of a simulated ES market, 2016 to 2025, through the protocol.
Robustness checks
Each one kills a kind of result that only looked good.
Every setting one step either way. A peak with cliffs around it is noise that landed well.
Slippage, fees and latency raised until the edge is gone. You learn the cost it survives.
The same stops, targets and hours, with entries at random. If chance does as well, the signal adds nothing.
The same rule, unchanged, on another contract and other bar sizes. An edge in one place only is where the data was cut.
The code is read for both before a run is spent. No decision may see a bar that had not closed.
The ten best days taken out. An edge that lives in a handful of days was a handful of lucky days.
Every idea, why it was tried, what measured it and what became of it, the failed ones too. It never runs a test twice, and every attempt counts against the winner.
What was tried and why, what measured it and what became of it. Ask about any line and the agent opens the runs behind it.
Four ideas and a sweep of 48. The winner's Sharpe is deflated for all 52, the dropped ones too.
The diff of what was kept, and the runs that measured both sides of it.
Features
The chat, the files and the runs go back to how they were before it.
A new chat from any point, the original left as it was.
Up to five on one equity chart, figure by figure.
What changed in the file, and what it did to the result.
It runs every version through the challenge ten thousand times, reads how the attempts end, and changes the strategy and its size until the odds hold.
94% Kept
of 10,000 attempts at a Topstep 50K Trading Combine pass, in 21 trading days at the median.
Atlantive agent
Skipping narrow mornings lifts it to 94%, passing in 21 trading days at the median. Failures fall from 28% to 6%.
A demonstration: the page's strategy on a simulated ES market, through the product's prop engine. On 28 trades the engine's interval for version 2 runs from 21% to 100%.
Choose any frontier model for each chat, set how hard it thinks, and cap what a month can spend.
Claude, GPT, Gemini, Grok, DeepSeek and the rest, chosen for each chat, with how hard it thinks.
Free credits every month, $10 for now, under a cap you set. One answer can only spend a share of what is left.
It never places an order or touches a brokerage account. It measures, you decide.
A chat app can write a strategy. It has no futures data to test it on, and it can quote a number no run produced. The agent gives the model you already use the market, the engine and the method.
| ChatGPT or Claude On its own | Atlantive agent The same models, free to start | |
|---|---|---|
| Writing it | ||
| The code | In any language you ask for | A short Python file in your workspace |
| Your chart | Reads a screenshot of it | Reads your drawings as prices and turns them into rules |
| TradingView | Writes the Pine from scratch | Compiles the tested file to Pine Script |
| Testing it | ||
| Market data | Only what you paste or upload | CME futures, one-second bars since 2010 |
| The backtest | In a general Python sandbox, when it has one | A futures engine: next-bar fills, fees and each market's slippage |
| The numbers | It can state a figure no run produced | Only from runs, each linked to its run |
| Trusting it | ||
| Overfitting | Yours to ask about | Walk-forward, purged cross-validation, deflated Sharpe |
| Out-of-sample days | Yours to keep apart | Held out, run once, logged as seen |
| Failed ideas | Forgotten by the next chat | Logged, never retried, counted against the winner |
| Prop firm rules | Not built in | Pass odds from the strategy's own trades |
What else the agent does
How your systems move together, how to split the risk and the drawdown they make as one.
A PDF tearsheet: the figures, the charts, the checks and a verdict.
Written in Python, drawn on your chart, used by any strategy.
Filters for economic releases and headlines: skip the Fed's days, or trade only them.
Size, stops and daily limits tuned on the strategy's own trades.
From a histogram of trades to a parameter surface in 3D.
Finds the paper, cites the page, and matches its trade counts before its returns.
The compiler's Pine Script, never Pine of its own.
Python to Pine Script exportThe Atlantive agent is an AI backtesting agent for CME futures. It turns a trading idea written in plain English, or drawn on a chart, into a short Python strategy, backtests it on historical futures data and reports what it measured: the trades, the equity curve, the drawdown and whether the result holds up. Along the way it studies the market, reads the research, keeps a log of every attempt and plays a finished run through a prop firm's rules.
The Atlantive agent runs on the same kind of models, Claude, GPT and Gemini among them, and gives them a futures market and an engine. A chat app has only the data you paste and can write a figure no run produced. The Atlantive agent backtests on CME futures data, quotes only numbers its runs returned, keeps every version of the file, and checks a winner before it calls it one.
No. You describe the idea in your own words and the Atlantive agent writes the code. What it writes is an ordinary, short Python file that you can open, read, edit and run again, so nothing about how the strategy trades is hidden from you, and a trader who codes can change the file by hand.
Yes. Draw your levels, zones and trendlines on the chart as you always do, and tell the Atlantive agent what you do there. It reads each drawing as prices and times, writes the rules it implies as a strategy, backtests them, and draws its own trades back on your chart, so you see where the rules agree with your eye and where they do not.
Yes. The Atlantive agent splits a backtest, or a trade list you bring, by regime: trending or ranging days, high or low volatility, the session, and days with economic releases or news. It shows where the strategy makes and loses money, proposes the filter the numbers argue for, and runs the filtered version once on days it never saw before calling it better.
The Atlantive agent backtests on futures traded on CME Globex: equity index, rates, currencies, energy, metals, agriculture and crypto, such as ES, NQ, ZN, 6E, CL, GC and BTC. Strategies run on price bars from one second to one month, with history back to 2010, and every market fill pays a commission per contract and a tick of slippage. The backtesting engine page has the details.
The Atlantive agent puts every winner through the anti-overfitting protocol quant funds use. The last years of data are locked before the first look; the rest goes through walk-forward analysis, combinatorial purged cross-validation and the probability of backtest overfitting; the Sharpe ratio is deflated for every attempt, failed ones included; a Monte Carlo sets the drawdown to expect; and the locked years are opened once. Six checks then attack the edge itself, from costs to random entries.
In Atlantive a strategy decides at a bar's close and its order fills at the next one-second bar, never on the bar it decided on, so no decision sees data it could not have had. Strategies are written in a closed trading vocabulary that is checked before they run, and every run records what examined its days before it, so a held-out period already looked at is never reported as unseen.
The Atlantive agent runs on the model you choose for each chat: Claude, GPT, Gemini, Grok, DeepSeek, Qwen and other frontier models, with how hard it thinks set per chat. The model writes and reasons, and the numbers always come from Atlantive's engine, so a cheaper model changes how well the agent works through an idea, not how a backtest is measured.
Signing up for Atlantive is free and open to everyone, with no card asked for, and for now every account receives $10 of agent credits each month. A credit is a dollar of model use at the provider's price, and one answer may spend only a share of what is left, so a single question cannot use up the month. Backtests and prop simulations spend no credits.
Yes. Your strategies, code, files, backtests and conversations in Atlantive remain yours, and Atlantive does not sell them or use them to train models. The Privacy policy names the providers that process them to produce the agent's answers.
Yes. The Atlantive agent plays a finished backtest through a prop firm's challenge rules, such as a 50K Topstep Combine, or rules you set yourself. The prop firm simulator resamples the run's own trades into thousands of attempts and reports the chance of passing each phase, the days it takes and the chance of blowing the account, and the agent then fits the contracts and the daily stop to the rules.
Yes. A strategy the Atlantive agent writes compiles to Pine Script, TradingView's language, with one click. The TradingView export translates the file that was tested rather than asking a model to write Pine again, and it refuses, by name, anything a chart cannot reproduce. For NinjaTrader, TradeStation or another platform, ask the agent to port the rules to its language: a port is a translation, so test it there again.
No. The Atlantive agent researches strategies and never places orders or connects to a brokerage account. What it reports are measurements of how rules would have behaved on past prices, not forecasts or recommendations, and nothing it says is investment advice. Trading futures carries a risk of loss, as the Risk disclosure explains.