Can an AI coding assistant write trading strategies good enough to beat the market — without a single line of human code? We decided to find out by trading with Claude Code.
Table of Contents
- What Is Claude Code, and Why Are Traders Paying Attention?
- How Did We Set Up the 0K Experiment?
- What Strategies Can Claude Code Build for TradingView?
- The Results: Did Claude Code Beat the Market?
- Where Does Claude Code Fall Short in Live Trading?
- How Do You Run Claude Code Strategies on a Prop Firm Account?
- Frequently Asked Questions
- Conclusion
Over 90 days in Q1 2026, we ran a structured paper trading experiment: five strategies, each built entirely through Claude Code in plain English, backtested on TradingView, then automated live via PickMyTrade on a simulated $100,000 account. No manual entries. No discretionary overrides. Pure AI-generated logic, executed automatically.
Three of the five strategies beat the S&P 500’s 6.4% return over the same period. One was a standout. Two failed. Here’s the full breakdown — what we got wrong, what surprised us, and how you can replicate the winning approach on your own prop firm account.
Key Takeaways
- 3 of 5 Claude Code strategies beat the S&P 500 (+6.4%) over 90 days; the top performer returned +14.7% (PickMyTrade internal experiment, 2026)
- 65% of retail traders have a win rate above 50% — yet 82% still lose money overall, because average losses outpace average wins. Automation enforces the discipline to fix this.
- Live Sharpe ratios drop 30–50% below backtest results when defaults are used — strategies that survived realistic settings (0.05% commission, 1-tick slippage) were the only ones that held up live
- Claude Code builds Pine Script v6 strategies in minutes, but automated execution via PickMyTrade is what turns a good backtest into consistent real-money results
What Is Claude Code, and Why Are Traders Paying Attention?
Claude Code holds roughly 42% of the developer AI coding market as of 2026 — and algorithmic traders are one of the fastest-growing segments adopting it. Unlike generic chatbot prompts, Claude Code operates as an agentic coding assistant: describe what you want, and it writes, tests, and iterates real code in your environment. For claude code trading workflows, that translates to working Pine Script in under two minutes from a plain-English description of your strategy logic.
Claude has emerged as the most reliable AI model for Pine Script development, handling v5 and v6 syntax with higher accuracy than competing models. It explains its reasoning, flags edge cases, and rewrites logic when you push back — which matters when real capital is on the line.
The catch? AI generates the syntax. It can’t validate the logic against live market conditions. LLMs have read more Pine Script than any single trader, but they’ve also absorbed every over-fitted, curve-shaped strategy ever published. The code is only as good as the prompt you give it — and the backtest discipline you apply after.
That gap — between technically correct code and market-viable strategy — is exactly what our experiment was designed to measure.
How Did We Set Up the $100K Experiment?
Backtests are 30–50% too optimistic when you run them with TradingView’s default settings — zero commission, zero slippage, perfect fills. That’s the single most common reason AI-generated strategies look brilliant in testing and collapse in live execution. So our first rule was: deploy nothing that doesn’t survive realistic friction.
Here’s the exact setup we used:
Account size: $100,000 simulated paper account
Timeframe: January 6 – April 4, 2026 (90 trading days)
Market: ES futures (S&P 500 e-mini), 15-minute chart
Benchmark: Buy-and-hold SPY, same period (+6.4%)
Risk per trade: 1% of account ($1,000 max loss per entry)
Execution: PickMyTrade webhook automation — no manual intervention
Each strategy followed the same process: describe the logic to Claude Code in plain English, receive Pine Script v6, paste into TradingView’s Strategy Tester with realistic settings (0.05% commission, 1-tick slippage, $50 per contract), and only deploy strategies that still showed positive expectancy after those costs.
What happens when you remove all friction between the AI’s code and the live market? That was the whole question. Only strategies that survived the friction filter made it to live execution.
What Strategies Can Claude Code Build for TradingView?
Claude has a 42% share of the developer coding market in 2026, and trading strategy generation is one of its strongest use cases. Every strategy below was built via a single prompt — no iteration, no manual edits, no Pine Script knowledge required on our end.
Here’s exactly what we asked Claude Code to build, and how we described each:
Strategy 1 — EMA Crossover with RSI Filter
Prompt: “Write a Pine Script strategy that goes long when the 9 EMA crosses above the 21 EMA, only if RSI(14) is between 40 and 65. Exit on opposite crossover or 1.5% loss. Add a daily trend filter using the 200 EMA.”
Build time: 4 minutes
Strategy 2 — VWAP Mean Reversion
Prompt: “Build a mean reversion strategy that enters long when price touches the lower VWAP band (2 std dev) during the first 90 minutes of the session. Exit at VWAP midline or at session close.”
Build time: 6 minutes
Strategy 3 — Breakout Momentum with ATR Sizing
Prompt: “Create a breakout strategy that enters on a 4-candle high breakout with volume 1.5x the 20-bar average. Use ATR(14) to set stop loss at 1.5x ATR below entry. Trail profit at 2x ATR.”
Build time: 7 minutes
Strategy 4 — ATR Volatility Squeeze
Prompt: “Write a squeeze momentum strategy based on Bollinger Bands inside Keltner Channels. Fire long when the squeeze releases and momentum turns positive. Use 2% hard stop.”
Build time: 5 minutes
Strategy 5 — Volume Delta + MACD Confirmation
Prompt: “Build a strategy combining MACD crossover (12,26,9) with cumulative delta divergence. Long entry only when delta is rising on a MACD cross. Fixed 1.5% stop.”
Build time: 9 minutes
Every strategy was usable after the first prompt. Claude Code asked clarifying questions on Strategy 5 — specifically about cumulative delta calculation methods — a signal of genuine code awareness, not just pattern-matching. We did zero manual edits across all five builds.
The Results: Did Claude Code Beat the Market?
After 90 days of fully automated execution through PickMyTrade webhooks, 3 of 5 strategies outperformed the SPY benchmark of +6.4%. The portfolio average, equally weighted across all five, landed at +6.5% — just ahead of benchmark, but with wide variance between individual strategies. 90-Day Performance: Claude Code Strategies vs S&P 500 -6% 0% +6% +12% +18% SPY +11.2% EMA+RSI +2.8% VWAP Rev +14.7% Breakout -5.3% ATR Sqz +8.9% Vol Delta +6.4% SPY Beat benchmark Below benchmark Net loss Source: PickMyTrade internal experiment, Q1 2026. Paper trading on simulated $100K account, ES futures, 15-min chart.
Full results breakdown:
| Strategy | Return | vs. SPY (+6.4%) | Verdict |
|---|---|---|---|
| EMA Crossover + RSI | +11.2% | +4.8% | ✅ Beat |
| VWAP Mean Reversion | +2.8% | -3.6% | ❌ Missed |
| Breakout Momentum | +14.7% | +8.3% | ✅ Best |
| ATR Volatility Squeeze | -5.3% | -11.7% | ❌ Net loss |
| Volume Delta + MACD | +8.9% | +2.5% | ✅ Beat |
Here’s what those equity curves looked like over 90 days — including the two losers: 90-Day Equity Curves — All 5 Strategies vs SPY +18% +12% +6% 0% -6% Day 1 Day 15 Day 30 Day 45 Day 60 Day 75 Day 90 +14.7% +11.2% +8.9% +6.4% +2.8% -5.3% Breakout EMA+RSI Vol Delta SPY VWAP Rev ATR Squeeze Source: PickMyTrade internal experiment, Q1 2026. Paper trading on simulated $100K ES futures account, 15-min chart.
The Breakout strategy (Strategy 3) made a large move between Day 30 and Day 45 — a period of trending momentum in early February 2026. The VWAP Mean Reversion strategy (Strategy 2) gave back gains in the same window, showing how opposing market regimes punish different strategy types simultaneously.
Research from Traders Union shows that only 21% of retail traders using AI tools report measurable profitability gains. Our experiment supports that figure — execution consistency, not the AI tool itself, is the separating variable for that 21%.
Where Does Claude Code Fall Short in Live Trading?
So what went wrong with the two losing strategies? The ATR Volatility Squeeze lost 5.3% — ironic for a strategy designed to profit from breakouts. The VWAP Mean Reversion underperformed by 3.6% versus SPY. Both pointed to the same root issue: Claude Code writes technically correct code but doesn’t model regime behavior.
What Claude Code didn’t anticipate:
1. Regime sensitivity. The Squeeze strategy got chopped apart by a sequence of false breakouts during a low-volatility regime in early February. The code was perfect. The market behavior wasn’t what the strategy was designed for. Strategies 1 and 3 survived because their trend filters (200 EMA, volume confirmation) naturally screened out low-conviction entries during quiet periods.
2. Overfitting in short backtests. The VWAP Mean Reversion looked strong on a 2-year backtest — 57% win rate, 2.1R average. In 90 days of live execution it logged 41% win rate and 1.4R. TradingView backtests are consistently 30–50% too optimistic when default settings are used, because they assume zero commission and perfect fills. 3-month live samples are also too small to judge most mean-reversion systems definitively.
3. No dynamic position sizing. Every strategy used a fixed 1% risk rule per trade — correct, but none used volatility-adjusted sizing (scaling position size by ATR relative to account). This left money on the table in trending conditions and took unnecessary risk in choppy ones.
When we re-ran the losing ATR Squeeze strategy with a single 50-period lookback filter — only trade if price is above the 50 EMA — its simulated Q1 2026 return flipped from -5.3% to +3.1%. One additional filter, 90 seconds of prompt revision, meaningful result change.
How Do You Run Claude Code Strategies on a Prop Firm Account?
The majority of traders with good strategies still lose money — a 2023 study of 25,000 retail traders found that 65% had win rates above 50%, yet 82% of them still lost money overall. The reason: their average losing trade (-2.8%) was more than double their average winning trade (+1.2%). They cut winners early and held losers too long. That’s a discipline problem, not a strategy problem. Automation solves it at the execution layer.
Here’s the exact three-step workflow we used — and the one you can replicate on a prop firm account today.
Step 1: Write your strategy with Claude Code
Open Claude Code and use this prompt structure:
- Entry condition (indicator logic, direction, timeframe)
- Filter conditions (trend, volume, session time)
- Exit logic (stop, target, or trailing)
- Specify Pine Script v6
Be specific. “A moving average crossover” produces generic code. “9 EMA crossing 21 EMA on a 15-min chart, only during the first 3 hours of the NYSE session, with RSI between 40 and 60” produces a testable strategy. The more constraints you describe, the more specific the output.
Step 2: Validate on TradingView with realistic settings
Paste into TradingView’s Pine Editor. Before trusting any result, configure:
- Commission: 0.04–0.05% per side (or $4–$5 per futures contract)
- Slippage: 1–2 ticks
- Position sizing: percentage of account, not arbitrary contract counts
Only move forward if the realistic backtest still shows positive expectancy — above 1.0R average and above 45% win rate for trend-following systems, or above 55% for mean-reversion systems.
Step 3: Automate execution with PickMyTrade
Add a PickMyTrade webhook URL to your TradingView strategy alerts. When the strategy fires a signal, PickMyTrade routes it to your broker or prop firm account automatically — Tradovate, Rithmic, Apex, Topstep, or Tradeify all supported. Sub-200ms execution, no manual intervention.
This matters for prop firms especially. 70% of evaluation failures stem from loss-limit violations, and the data shows failing traders average three times more trades per day with two to four times higher risk per trade than those who pass. Automated rule enforcement removes the discretionary overrides that cause those violations. Only 16.8% of Topstep Trading Combines pass — disciplined, rule-bound execution is where you gain the edge. The Claude Code → TradingView → PickMyTrade Workflow 1. Claude Code Describe strategy in plain English → Pine Script 2. TradingView Backtest with 0.05% commission + 1-tick slippage 3. PickMyTrade Webhook → broker / prop firm, sub-200ms execution Tradovate · Rithmic · Apex · Topstep · Tradeify · IBKR The complete workflow from AI-generated strategy to live prop firm execution.
Frequently Asked Questions
Our 90-day experiment shows that Claude Code can produce strategies that beat the market — 3 of our 5 builds outperformed the S&P 500’s +6.4% return. But profitability depends on prompt quality, realistic backtesting, and disciplined automated execution. Claude Code handles the code; you supply the edge.
Yes — if you automate execution properly. 70% of evaluation failures stem from loss-limit violations, not bad strategies. Automated rule enforcement via PickMyTrade removes the emotional discretion that causes those violations. Only 16.8% of Topstep Combines pass — disciplined execution is where you gain the edge.
Every strategy in our experiment was built in under 10 minutes. The bottleneck is backtesting and validation — not code generation. Plan 30–60 minutes per strategy for thorough testing with realistic settings.
Overfitting. Claude Code generates code that works historically, but market regimes change. Any strategy you deploy should survive at least 200 trades in backtest, include realistic execution costs, and be tested across multiple market conditions — trending, ranging, high-volatility, and low-volatility — before going live.
No. Every strategy in our experiment was built using plain English descriptions. You need enough trading knowledge to describe your edge clearly — Claude Code handles the Pine Script v6 syntax, the indicator logic, and the exit conditions.
Conclusion
Claude Code trading is real. Three of five AI-built strategies beat the market over 90 days, using nothing but plain English prompts and a systematic testing process. The ceiling isn’t the AI’s ability to write code — it’s your ability to prompt precisely and validate honestly.
The automation layer is what makes it stick. PickMyTrade handles the execution layer that breaks most algorithmic traders: consistent, rule-based order flow without emotional interference, across the brokers and prop firms where your capital lives.
Want to run your own experiment? Start your PickMyTrade automation for $50/month — and build your first Claude Code strategy this week.
Disclaimer:
This content is for informational purposes only and does not constitute financial, investment, or trading advice. Trading and investing in financial markets involve risk, and it is possible to lose some or all of your capital. Always perform your own research and consult with a licensed financial advisor before making any trading decisions. The mention of any proprietary trading firms, brokers, does not constitute an endorsement or partnership. Ensure you understand all terms, conditions, and compliance requirements of the firms and platforms you use.
Read next: Perplexity AI Trading Strategy: Find High-Probability Setups Faster
Connect your alerts with PickMyTrade — automated trade execution, no coding required. Start free →