The Setup
I had just killed my 4th trading bot. Grid bot, RSI scalper, lead-lag arb, momentum bot — all dead for different reasons.
I was frustrated. Not because the code didn’t work — it worked perfectly. The strategies were the problem. I kept building bots that looked great in backtests and bled money in production.
So I tried something different. Instead of coming up with another strategy myself, I asked Claude Code to analyze my failures and suggest what to build next.
What I Gave It
I didn’t just say “give me a trading strategy.” That would have gotten me a generic RSI crossover tutorial.
Instead, I gave it context:
- My dead bots — what each one did and why it failed
- My constraints — limited starting capital, Binance Futures, 3x leverage max
- My backtest results — the actual numbers, including the ones that looked good but failed live
- My principles — risk-reward > win rate, no overfitting, must survive trending AND sideways markets
Then I asked: “Given everything that failed, what should I build?”
What It Suggested
Claude didn’t give me one strategy. It analyzed the failure patterns and pointed out something I’d missed:
Every dead bot shared the same flaw — they all tried to predict direction.
- Grid bot: predicted a range
- RSI scalper: predicted reversals
- Momentum bot: predicted continuations
- Lead-lag: predicted correlation persistence
Claude’s suggestion: Stop predicting. Start reacting.
Instead of “I think price will go up, so buy” → “Price IS going up strongly right now, confirmed by volume, so ride it.”
This became my Trend Following bot.
The Strategy Claude Helped Design
The core logic was surprisingly simple:
Entry Signal
1. 5-minute candle has a large body (≥ 0.7%)
→ Price moved decisively, not just noise
2. Volume Ratio ≥ 1.5x average
→ The move has real participation behind it
3. Choppiness Index < 50
→ Market is actually trending, not chopping sideways
That’s it. No prediction. No “RSI is oversold so it should bounce.” Just: is something happening right now, with conviction?
What Claude Got Right
The volume filter was the key insight I kept missing.
In my previous bots, I entered on price signals alone. Claude pointed out that price moves without volume are noise — they reverse immediately. Volume confirms that real money is behind the move.
Adding a volume ratio filter significantly improved the win rate and overall profitability.
What I Had to Override
Not everything Claude suggested worked. Some ideas sounded logical but performed poorly in backtests. The key lesson: AI gives you good frameworks, but the specific implementation needs human judgment and backtesting.
Every suggestion got tested. Some survived. Many didn’t. The ones that survived became the final strategy.
The Iteration Loop
The collaboration followed a consistent pattern:
- Run the bot or backtest
- Show Claude the results — actual trade logs, PnL numbers, losing trades
- Claude analyzes patterns: “most losses happen during choppy markets” or “this SL level is too tight”
- I implement the suggestion, backtest again
- Repeat
This cycle happened dozens of times over weeks. Each round, the strategy got slightly better.
Claude never wrote the final strategy alone. It analyzed data faster than I could, spotted patterns I missed, and challenged my assumptions. But the decisions — which suggestions to keep, which to reject — were mine.
The Parameter Optimization Trap
Grid searching parameters is where AI becomes dangerous. Claude can find the “optimal” combination faster than any human — but “optimal on training data” usually means “overfitted.”
I learned this the hard way with my momentum bot. Amazing backtest. Terrible out-of-sample. The parameters had memorized the past instead of finding a real pattern.
The fix was to optimize fewer parameters at a time (2-3 max), use longer data periods, and always validate on out-of-sample data. When training results and OOS results diverge, the training results are lying.
AI can find the overfitted peak faster than any human. That doesn’t make it useful. The discipline to reject “optimal” parameters in favor of robust ones is a human job.
What AI Is Actually Good At (In Trading)
After months of this, here’s my honest assessment:
AI Excels At:
| Task | Why |
|---|---|
| Bug detection | “Your SL is checking best_price but should check worst_price” |
| Pattern analysis | “Most losses cluster in specific market conditions — here’s where” |
| Code generation | Building the actual bot, fast and (mostly) correct |
| Framework design | “You need trend detection + volume confirmation + regime filter” |
| Failure analysis | “This strategy fails because X, not because of bad luck” |
AI Is Bad At:
| Task | Why |
|---|---|
| Predicting markets | No AI can consistently predict price direction |
| Parameter optimization | It finds the overfitted answer too easily |
| Knowing when to stop | It’ll keep adding complexity if you let it |
| Risk management | It doesn’t feel the pain of losing money. You do. |
The Results
The Trend Following bot (designed with Claude, not by Claude) has been running live since late February 2026:
- 6 weeks of live trading
- Win rate: ~57%
- Risk-reward: ~1:1.2
- Still running. Still profitable.
It’s not a moonshot. It’s not “quit your job” money. But it’s a strategy that survived the gap between backtest and reality — which is more than I can say for bots #1 through #4.
The Second Bot
After the trend follower stabilized, I asked Claude to help me build something complementary — a strategy that would perform well when trend following struggled.
The answer: mean reversion using Fair Value Gaps.
Different entry logic, different market conditions, different edge. When the trend bot sits idle in choppy markets, the FVG bot trades. When the FVG bot struggles in strong trends, the trend bot prints.
Two bots, designed together, covering each other’s weaknesses. That portfolio idea? Claude’s suggestion. And it was one of the best ones.
What I’d Tell Someone Starting Today
-
Don’t ask AI for a strategy. Ask it to analyze your failures. The strategy emerges from understanding what doesn’t work.
-
AI is a collaborator, not an oracle. It accelerates your learning by 10x, but the judgment calls are yours.
-
Be suspicious of AI’s “optimal” answers. The best answer in a backtest is usually the most overfit one.
-
Feed it real data, not hypotheticals. “What strategy should I use?” gets a generic answer. “Why did these 47 trades lose money?” gets a useful one.
-
The AI doesn’t care if you lose money. It’ll build whatever you ask for, beautifully, even if what you asked for is garbage. You need to be the skeptic.
The Honest Truth
Claude Code didn’t give me a winning strategy. It gave me a faster way to find one.
Without AI, I’d still be on bot #2, manually scanning trade logs, guessing at what went wrong. With AI, I iterated through 6 strategies in a few months, identified exactly why each one failed, and converged on something that works.
The AI wrote most of the code. The AI analyzed most of the data. But the AI didn’t make the decisions that mattered — when to cut a strategy, how much risk to take, when to stop optimizing and start trading.
That part is still human. And honestly? I think it should stay that way.
The best use of AI in trading isn’t to replace your judgment. It’s to give your judgment better data to work with.