The Session Where Everything Clicked: +$118 in 36 Hours
The Session Where Everything Clicked: +$118 in 36 Hours
Date: February 23, 2026 Author: Jamie Watters Project: Trader-7 — An LLM-Powered Crypto Trading System
The Setup
Two days ago, my AI trading system was sitting at +$90.73 — a modest 3% return after 24 days of paper trading. The win rate was a mediocre 34%. The Sharpe ratio was 0.57. And the system had just opened what I described in my notes as a "marginal" ETH SHORT: borderline ADX, minimum confidence, mediocre score.
Thirty-six hours later, the P&L reads +$208.70 (+7.0%). The Sharpe ratio is 1.83. And that "marginal" trade turned out to be the system's biggest winner to date.
The Market Shifted — And the System Noticed
When the session started on February 21, the market was indecisive. BTC was oscillating around its 50-day moving average — a fraction of a percent above, then below, then above again. The ADX (a measure of trend strength) was stuck between 10 and 15: textbook trendless conditions.
The system responded correctly: mostly HOLD signals, defensive stance, high confidence thresholds. It closed two inherited positions early — a SOL LONG captured by trailing stop (+$26.32) and an XRP LONG stopped out (-$33.09). Then it just... sat there.
But around cycle 15 (mid-afternoon UTC on Feb 22), something changed. BTC slipped below the SMA50 and stayed there. The regime classification shifted to WEAK_BEAR. The system opened that ETH SHORT I'd flagged as marginal.
Then the sell-off accelerated.
Over the next 12 hours, BTC dropped from $68,400 to $65,300. The SMA50 distance went from -0.6% to -4.75%. The ADX surged from 15 to 42 — from "no trend" to "strong trend" in half a day.
And the system adapted at every step:
- Stance shifted to defensive — raising the confidence threshold from 72% to 82%, blocking marginal proposals
- Bias locked to bearish — all new signals evaluated through a SHORT lens
- Regime-adaptive take-profits activated — when ADX crossed 35, the system switched to wider targets (2.0R and 3.5R instead of the default 1.5R and 2.5R), letting winners run further in the strong trend
- New positions aligned with regime — SOL SHORT opened at $83.29, perfectly directional
None of this was hard-coded. The system was reading conditions and adjusting its own behavior cycle by cycle — which is kind of the whole point of the project, but it's one thing to design that and another to watch it actually work.
Two Trades That Made the Session
ETH SHORT #144: From "Marginal" to +$104
This trade was the one I almost dismissed. Entry at $1,954.52 with ADX at 15 and a confidence score of 72% (the floor). The only thing going for it was regime alignment — SHORT in WEAK_BEAR.
But the system held. As the sell-off deepened, ETH dropped through $1,931 — the first take-profit target. The system:
- Closed 50% of the position at $1,931 — locking in $22.95
- Activated the trailing stop below the current price
- Ratcheted the trail as ETH continued falling — through $1,900, $1,890, $1,880
Finally, at 06:25 UTC on Feb 23, the trailing stop triggered at $1,872.12. The remaining 50% closed for +$81.71.
Total realized: approximately $104.66. From a trade I called "marginal."
SOL SHORT #145: The Clean One (+$130)
If ETH was the surprise, SOL was the clean one. Opened at $83.29 in the WEAK_BEAR regime with 85% confidence and strong regime alignment.
SOL fell hard — $82, $80, $79. TP1 hit at $81.70, partial close banked $30.61. The trailing stop activated and tracked the price down. It captured the remaining position at $78.12 for +$99.52.
Total realized: approximately $130.13. The system's biggest single-trade profit.
The Session Numbers
| Metric | Start | End |
|---|---|---|
| Cumulative P&L | +$90.73 (+3.0%) | +$208.70 (+7.0%) |
| Sharpe Ratio | 0.57 | 1.83 |
| Win Rate | 34% | 38.5% |
| Fee Drag | 321% | 42.7% |
Session-specific stats (36 hours only):
- 6 trades opened, 4 closed
- 2 wins, 2 losses — 50% session win rate
- Session P&L: +$117.97 (+3.9% of capital)
- Fees: $60.74 total accumulated vs $142.15 realized profit
The fee drag number is worth calling out. It was 321% two days ago — meaning fees exceeded realized profits by 3x. After this session's trail captures, it's now 42.7%. Still elevated, but the trajectory is exactly right. Profitable trades are the cure for fee drag, and the system just produced two large ones.
What the Validator Got Right
The most impressive aspect wasn't the winning trades. It was the trades the system didn't take.
During the sell-off, as RSI dropped to extreme oversold levels (24-28), the signal generator kept proposing new SHORT positions. But the validator blocked them cold:
"RSI at 25.1 is deeply oversold without reversal confirmation — shorting into extreme oversold conditions carries high risk of a snap-back rally, even in a bearish trend."
This turned out to be exactly right. SOL bounced from $78 to $80.54, which stopped out Trade 146 (the one entry that slipped through after RSI had partially recovered). The validator's oversold rejections prevented what could have been multiple additional losses during that bounce.
The defensive stance (82% confidence threshold) also blocked 6+ low-conviction proposals. That's the kind of discipline I was hoping to see — capital preservation when the setup isn't clean.
The One Loss Worth Discussing
Trade 146 was a SOL SHORT opened at $78.99 — right after the trailing stop had just captured profit at $78.12. The system was essentially re-entering the same trade immediately after taking profits.
SOL bounced to $80.54, hitting the stop for -$33.28.
The lesson: don't chase extended moves. After a deep sell-off with RSI in the 20s, a mean-reversion bounce is likely. The validator had been blocking exactly these setups for hours, but by the time this cycle ran, RSI had recovered enough (above 30) to pass the check.
This is a pattern worth tracking. The system's pre-entry validation is strong, but timing re-entries after major moves needs refinement.
Trailing Stops: The Profit Engine
The trailing stop mechanism (fixed in Sprint 99.1 just three days ago) has now completed its lifecycle successfully on 3 out of 3 trades:
- SOL #142: Trail captured +$26.32
- ETH #144: Trail captured +$81.71 (after TP1 took +$22.95)
- SOL #145: Trail captured +$99.52 (after TP1 took +$30.61)
Combined trail captures: $207.55 out of $254.04 total session profit (81.7%).
Without trailing stops, the system would have captured only TP1 partials — roughly $77 instead of $254. The trailing stop mechanism is the single most impactful feature for profitability.
Three days ago, this feature was completely broken because of five missing fields in a database model. Silent failures. No errors in the logs. Just a feature that never triggered.
What I'm Taking Away
The big one: entry quality matters less than I thought. That ETH SHORT was borderline — ADX at 15, barely passing confidence checks. But because the system held the position and let the trailing stop do its job, it became the best trade we've had. I've been obsessing over entry signals when the real edge is in position management.
Trailing stops are doing the heavy lifting. 81.7% of session profit came from trail captures, not the initial TP1 partials. Without them, this session would've been maybe +$77 instead of +$254. Three days ago that feature was completely broken (five missing database fields, no errors, just silence). Terrifying in hindsight.
The regime-adaptive targets are working too — wider TPs when ADX crossed 35 let the system ride the strong trend instead of taking profit too early. Fixed targets would have left a lot on the table.
And honestly, the validator rejections impressed me more than the entries. Blocking shorts into RSI 25 while the sell-off was screaming "go short!" takes a kind of mechanical discipline that I'm not sure I'd have had manually.
The one clear miss: don't re-enter a trade you just took profit on. Trade 146 cost us $33 learning that lesson. The system needs a cooldown timer or a "just captured trail" flag. That's Sprint 100 territory.
What's Next
The system currently has two open positions:
- BTC SHORT #147 at $65,838 — very close to its $66,494 stop, BTC hovering around $66,400. This one's at risk.
- SOL SHORT #148 at $80.38 — near breakeven, needs SOL to drop to $77.43 for TP1.
The regime is recovering — BTC SMA50 distance went from -4.75% back to -1.41%. If that recovery continues, both positions face headwinds.
The next development sprint (Sprint 100) will likely add a regime-direction check before opening new trades. Three previous trades (XRP LONG into WEAK_BEAR, ETH SHORT into WEAK_BULL) demonstrated the gap. The system should never open a trade that opposes the global regime direction.
But for now, the system just had its best 36 hours. From a 3% return with a 0.57 Sharpe to a 7% return with a 1.83 Sharpe. The trailing stop fix from three days ago has generated more than $200 in captured profits. And the multi-LLM pipeline — signal generator, strategist, validator — is starting to behave like the collaborative intelligence layer it was designed to be.
Trader-7 is a paper trading system. No real money is at risk. All figures represent simulated results. Past simulated performance does not guarantee future results.
Follow the build: @Jamie_within | jamiewatters.work