Claude Trading Bot: How to Use Claude to Build and Audit Forex Indicators

Claude, the AI assistant made by Anthropic, is a strong partner for traders who want to build their own tools. It can turn a clear trading rule into Pine Script, MQL5 or Python, explain what the code does line by line, and, most valuably, audit code for the mistakes that make backtests lie. This guide shows how to use it well.

What Claude is good at for traders

• Clarifying rules. Ask it to restate your strategy as a numbered list and point out ambiguities before any code is written.
• Writing code for TradingView (Pine Script v5), MetaTrader 5 (MQL5) and Python data analysis.
• Explaining code you found elsewhere, so you know what a downloaded indicator really does.
• Auditing for look-ahead bias, repainting, missing costs and position-sizing errors.
• Analysing your trade journal to find which setups and sessions actually make money.

What it cannot do is predict the market. No AI can. Treat it as a fast, careful engineer working to your instructions.

A workflow that produces usable tools

1. Specify. Describe the rules precisely: timeframe, exact conditions, which candle the signal appears on, stop and target logic.
2. Restate. Ask Claude to restate the rules and list anything unclear. Fix the specification first.
3. Generate. Ask for code for one named platform, using only closed candles, with every number as an input and plain-English comments.
4. Compile and correct. Paste error messages back and ask for fixes that do not change the logic.
5. Verify by eye on at least twenty historical examples.
6. Audit. Ask Claude to review its own code for look-ahead bias and repainting before you test it.
7. Test and freeze. Backtest with costs on separate development and test periods, then save the version.

The audit prompt that saves accounts

The single most useful prompt is:

"Review this code for look-ahead bias, repainting, survivorship bias, missing transaction costs and position-sizing errors. For each problem, give the line and explain on which candle the information would really have been known."

Look-ahead bias is the most dangerous bug in trading research because it makes losing strategies look like winners. When building the TebotechSignals engines, a research script of ours showed a steady profit across 32 currency pairs. A candle-by-candle replay revealed that it had quietly used knowledge of the future. The profit disappeared. We tell that story in full in The Hybrid Trader.

Smart Money Concepts with Claude

Claude handles Smart Money Concepts well when you define them as rules rather than impressions: a swing high is a candle higher than the N candles on each side; a break of structure is a close beyond the last swing; a fair value gap is when candle three's low is above candle one's high. Our guide to building an SMC bot gives the full rule table.

Build the risk module first

Ask Claude to write a separate safety engine that every order passes through: fixed risk per trade, a daily loss limit including open positions, drawdown limits that halve risk and then stop trading, exposure limits per currency, a news blackout and a kill switch. A bot with a mediocre entry and excellent risk control survives. A bot with a brilliant entry and no risk control does not. See algorithmic risk management for forex bots.

Claude or ChatGPT?

Both can do this work. The workflow matters far more than the model. Use whichever you are comfortable with, and keep the habits that protect you: precise specifications, audits for future data, honest tests with costs, and a live record before real money. The same steps for ChatGPT are in our guide to building a trading bot with ChatGPT.

Keep the human in the loop

The best use of Claude is not a bot that trades while you sleep. It is a bot that maps the market, prepares orders and enforces limits while you keep the final decision. Start there, and let your records tell you when the bot has earned more responsibility.