The Trade That Changed Everything: How One Filter Turned Our AI Trading System Around
The Trade That Changed Everything: How One Filter Turned Our AI Trading System Around
March 20, 2026 : Day 119 of building Trader-7 in public
We just had our best 22-hour session since we started paper trading. Three trades closed: two winners totalling +$102, one loss of -$31. Net +$71.
But the real story isn't today's session. It's the journey across three distinct eras that got us here : and the single insight that turned everything around.
The Three Eras of Trader-7
Era 1: The Scatter-Shot Phase (88 trades, -$31)
For our first 88 trades, the system traded in every market condition. Bull, bear, choppy, trending : didn't matter. The LLM agents would generate signals, validate them, assess risk, and execute. The win rate hovered around 36%.
The signals weren't bad. The validation was solid. The risk management worked. But we were trading in conditions where nothing consistently works.
Era 2: The Painful Discovery (18 trades, -$86)
Sprint 106 introduced direction-aware confidence thresholds : if the market regime says WEAK_BEAR, SHORT trades get a lower bar to clear (75% vs 82%). This nudged the win rate up to 39%.
But we were still bleeding. Badly. -$86 in 18 trades.
Then we looked at the data. Really looked at it.
| BTC SMA50 Distance at Entry | Trades | Win Rate | Net PnL |
|---|---|---|---|
| Above 2.5% | 4 | 100% | +$114 |
| Below 2.5% | 12 | 0% | -$240 |
That's not a statistical edge. That's a binary switch. Every single trade entered when BTC was more than 2.5% from its 50-period simple moving average won. Every single trade entered below that threshold lost.
The LLMs were smart enough to find good setups in trending markets. But in the choppy zone near the SMA50 : where price whipsaws around the mean : no amount of AI intelligence can overcome the noise.
Era 3: The Turning Point (7 trades, +$66)
Sprint 116 was brutal in its simplicity: a hard floor. If BTC is less than 2.5% from SMA50, the entire LLM pipeline doesn't even run. No signals, no validation, no risk assessment, no trades. The system sits on its hands.
Since deploying that filter five days ago: 7 trades, approximately 57% win rate, and +$66 in net P&L. The wins are full TP1-plus-trailing-stop captures. The losses are small : the regime watchdog tightens stops to breakeven when conditions shift.
We went from -$86 to -$20 in the Sprint 106 tracking window. The system is converging on breakeven, and the trajectory is pointing up.
What Today's Session Looked Like
The session ran overnight and through the day : 24 full analysis cycles over 22 hours.
Trade #206 (BTC SHORT, +$52): Entered at $71,452 when momentum confirmed (ADX at 40.9). Hit the first take-profit target, activated trailing stop, and rode it down for an additional $22. Textbook execution.
Trade #207 (ETH SHORT, +$50): Same momentum window, entered at $2,194. Similar pattern : TP1 hit, trailing stop captured the remaining move. The correlation cap correctly prevented both ETH and SOL from being open simultaneously.
Trade #208 (BTC SHORT, -$31): Entered later at $69,450 when ADX was even stronger (46.85). The validator approved it : RSI was 32 (not extreme), momentum was confirmed, R:R was solid at 1.73:1. But the market bounced at support. A reasonable trade that didn't work out. That's trading.
The validator also rejected several BTC SHORT proposals when RSI hit 26-27, explicitly warning about "oversold snapback risk." That's the kind of nuanced intelligence you want from an AI validator : knowing when the technical setup says SHORT but the risk says "you're chasing."
The Unsexy Truth About AI Trading
The biggest improvement to our AI trading system wasn't a better model, a smarter prompt, or a more sophisticated strategy. It was a four-line if statement that checks whether BTC is far enough from its moving average.
Sprint 117 helped too : caching the strategist output when nothing has changed, downgrading routine risk checks from Opus to Sonnet. These reduced costs by roughly 40% per week. When your system is designed to sit out bad conditions, you want patience to be cheap.
But the alpha was always in market selection. The AI agents were already good at finding setups. We just needed to stop asking them to find setups in markets where setups don't exist.
What's Next
We found one bug today: the risk manager's Sonnet model truncated its response once in 24 cycles because the token limit was too low (750). The JSON broke mid-sentence. We've already fixed it : bumped to 1500 tokens and added explicit truncation detection warnings.
The system is currently idle, correctly blocked by the SMA50 floor (BTC is only 2.33% from SMA50). It'll wait. Patience is the strategy now.
103 trades in. Capital at $2,977. The -$23 all-time P&L doesn't sound impressive : until you realise we were at -$89 five days ago.
The trend has reversed. And this time, we have the data to explain why.
Building Trader-7 in public : an LLM-powered crypto trading system. Follow the journey at jamiewatters.work
@Jamie_within on X | linkedin.com/in/jamie-watters-solo