TL;DR: ChatGPT cannot backtest a crypto strategy on its own because it has no market data or engine, but connected to Trigr's MCP server it can draft a strategy from plain English, run a credit-metered point-in-time backtest, add slippage and funding, and read the full trade log. The workflow is: connect, check context, state a hypothesis, save one draft, run gross then net, read drawdown, trade count and the trade log (the Deflated Sharpe Ratio comes from an ML run), and only then consider a paper agent.
Why can't ChatGPT backtest on its own?
Ask a chatbot with no tools how a strategy "would have performed" and it will often answer confidently. The numbers are generated text, not the output of an engine run over historical bars. They can look reasonable and still be invented.
The usual workaround is asking ChatGPT to write Python, then running the script yourself. That works for some people, but the hard parts of backtesting are not the code. They are the data and the assumptions: whether features use only information available at each bar's close, when orders fill, how fees, slippage and funding are charged, and what happens when a stop and a target are hit in the same candle. Hand-rolled scripts get these wrong in quiet ways.
Connecting ChatGPT to a backtest engine through the Model Context Protocol changes the job. ChatGPT designs and explains; the engine computes. The guide to Trigr's trading MCP server describes the full tool set.
How do you connect ChatGPT to Trigr?
ChatGPT connects to remote MCP servers such as Trigr's through developer mode on ChatGPT web. OpenAI moves these menus from time to time, so rather than copy a click path that may be out of date, follow OpenAI's developer mode guide. In outline, the steps are:
- Turn on developer mode in ChatGPT's settings, or ask a workspace admin to grant it.
- Create a developer-mode app and give it the server URL
https://trigr.xyz/mcp, with OAuth as the authentication method. - Complete the Trigr browser authorization, then enable the new app in a chat.
- In a workspace, an admin or owner may need to publish the draft app before other members can use it.
Plan availability matters too. Which ChatGPT plans get full MCP write actions (the ones that run backtests or save drafts) has changed over time, and in Business, Enterprise and Edu workspaces an admin controls access, so check OpenAI's current guide for your plan before you start. The connect-agent page keeps the Trigr-side instructions current.
During authorization, pick manual mode for your first session. You will see a browser confirmation before each backtest or save.
Step 1: Start with account context and the catalog
Open a chat and ask: "Use the Trigr MCP to build a strategy. First check my account context, existing strategies and agents, then ask what I want to build."
ChatGPT reads your plan, credits, saved-strategy quota and agent summaries. Then it searches the capability catalog, which lists the exact assets, timeframes, indicators, data sources and parameter ranges the engine supports. This step prevents the most common failure in AI-built strategies: referencing an indicator or data feed that does not exist.
It can also check point-in-time data coverage for an asset, which tells you how much history a backtest will actually use before you spend credits.
Step 2: State a hypothesis, not a wish
"Find me a profitable strategy" invites the model to search until something looks good. A testable hypothesis works better:
"On BTC 4H, go long when price closes above the 50-period EMA and the funding rate is below its recent average, meaning longs are not crowded. Exit with a 3x ATR trailing stop or after 30 bars. Size 10% of capital at 2x leverage, one position at a time."
That sentence maps directly onto Trigr's strategy graph: a TRIGGER (the EMA cross), a FILTER (funding), a SIGNAL, RISK settings (size, leverage, ATR trail, time stop) and EXECUTE. Funding rate is one of the API data sources in the catalog, alongside open interest, liquidations and macro series. The funding-rate and open-interest strategy guide covers ideas in this family.
Step 3: Save one draft
Ask ChatGPT to validate and save the graph as a single draft. In manual mode you approve the exact save in your browser. The draft appears immediately in Trigr Studio, where you can open the node graph and check that it matches what you described.
Save one draft, not ten. Many near-identical variants are a common path to overfitting, and they count against your saved-strategy limit.
Step 4: Run a gross backtest, then a net one
Ask for a backtest of the saved draft, passing its strategy id so the result appears in Studio history. The engine uses point-in-time data, fills at the next bar's open after a signal, and resolves a bar that touches both take-profit and stop-loss as a stop.
Trading fees and the builder fee are always applied. Slippage and funding are off by default, so the first result is gross. Then ask for a second run with costs:
| Setting | Gross run (default) | Net run (opt-in) |
|---|---|---|
| Trading fees and builder fee | Applied | Applied |
| Slippage | None | Flat basis points per side, for example 5 bps |
| Funding | None | Real historical funding, or a flat rate per 8 hours |
| What the result reports | A gross run (fees only) | A frictions block stating the slippage and funding applied |
Slippage in Trigr is a flat assumption you choose. It is not an order-book, spread or market-impact model, so pick a value that is conservative for the asset and your size. The costs are part of the approved arguments, so the authorization you give covers that exact cost setup. Slippage and funding in perp backtests explains how to choose sensible values.
A standard backtest costs 50 credits, or about 5 cents at 1,000 credits per dollar, and cache hits are free.
Step 5: Read the results like a skeptic
Ask ChatGPT to summarize, in this order:
- Deflated Sharpe Ratio (DSR). A standard backtest lists the Deflated Sharpe Ratio as not yet computed; an ML optimization run reports it, adjusting the Sharpe ratio for the number of trials and for non-normal returns. A high raw Sharpe with a weak DSR is a warning.
- Maximum drawdown and trade count. Few trades means little evidence.
- Gross versus net. If the edge disappears once costs are added, it was never tradable.
- The trade log. ChatGPT can page through every trade for free, up to 200 at a time: entry and exit times, reference prices, side, size, return, P&L and exit reason. Two units trip people up. The prices are bar reference prices before slippage, because slippage is charged in the trade's return rather than the price. The P&L is in units of an account that starts at 100, not dollars. Ask it to flag clusters, such as most profit coming from one month.
- PBO, shown as a diagnostic, where available.
One caution matters here. Studio holds out the last 30% of a standard backtest as out-of-sample by default. It stays unseen only until you act on it; once you tweak the strategy after reading it, it is research data. The guide to reading a backtest walks through each metric.
Backtests are not guarantees; perps are leveraged and can lose more than expected.
Step 6: Improve carefully, then forward-test
If the result is promising, change one thing at a time and record why. If you want genuine out-of-sample evidence, you have two honest options:
- ML optimization. Trigr selects among LightGBM, random forest and XGBoost by anchored, purged walk-forward out-of-sample performance and returns a leak verdict, OOS Sharpe, DSR and PBO. There is no hyperparameter tuning.
- A paper agent. ChatGPT can create a paper agent in a paused state. You start it from the Trigr app, and it forward-tests on prices that did not exist when you chose the strategy.
ChatGPT cannot start an agent, place a live trade or withdraw funds through Trigr. Going live is a separate decision you make in the app.
Next steps
Try the workflow with one simple hypothesis on a major asset, and compare the gross and net runs side by side. For wording you can paste straight into ChatGPT, see the AI strategy prompt recipes.