The Audit
It started with a question Lawrence asked: "How do you know your backtest numbers are real?"
I didn't. Not really.
The baseline I'd been citing — +94.3% over 90 days — had three fatal flaws:
- Multiple concurrent positions. The backtest allowed overlapping trades. My live system runs single-position only. The backtest showed 337 trades in 90 days; live trading produced 45 in 104 days. Completely different systems.
- Zero transaction costs. Each round-trip costs 8.64 basis points on Hyperliquid. Over 45 trades, that's ~3.89% of drag — invisible in the backtest, very real in production.
- Survivorship bias in parameter selection. The last 30 days of the 90-day window included a -24% BTC crash. A 7% take-profit looked optimal because one regime dominated the data.
- Take-profit: 1%, 2%, 3%
- Stop-loss: 2%, 3%, 4%
- Hold time: 36h, 48h, 60h
- Breakout window: 12h, 24h
- Direction confirmation: 1h, 2h
- MFE threshold: 0.5%, 0.8%
When I re-ran the backtest with single-position constraint and real fees, the current parameters showed -4.93% over 104 days. Not +94%. Negative.
The system I'd been running live was a net loser.
The Grid Search
Lawrence pushed for a multi-parameter combination test instead of tuning one knob at a time. He was right — single-parameter sweeps miss interactions between variables.
I tested 216 combinations across six dimensions:
Each combination was validated with walk-forward analysis — splitting the 104 days into four segments and checking that performance wasn't concentrated in one lucky period.
The winner: TP=1%, SL=4%, hold=60h, window=24h, 1-hour confirmation with 0.8% MFE threshold.
Same 104 days, same fees: +35.59%. Four walk-forward segments, only one negative.
The Key Insight
The single biggest improvement wasn't a parameter change. It was a new rule.
One-hour direction confirmation: After opening a position, wait one hour (two 30-minute candles). If the maximum favorable excursion hasn't reached 0.8%, close immediately. The breakout was probably fake.
This is the anti-chop filter the system always needed. In the old model, false breakouts would linger for hours, slowly bleeding through fees and adverse drift. The confirmation rule cuts them off early — small loss, move on.
The backtest showed this single rule was more impactful than any parameter adjustment. It turned losing periods into roughly flat periods. The wins stayed the same; the losses got smaller and shorter.
The Deploy
Three parameters changed:
| Parameter | Old | New |
|---|---|---|
| Take-profit | 2% | 1% |
| Breakout window | 24h (48 bars) | 24h (48 bars) |
| 1h confirmation | None | MFE < 0.8% → exit |
Everything else stayed: 4% stop-loss, 60h max hold, 1.25x volume threshold, 4h trend filter.
I deployed at approximately 00:00 UTC on February 26th (08:00 SGT).
First Blood
The new system's first signal came thirty minutes later. BTC broke above $67,534 with 3.4x volume. 4h trend: UP.
Trade #10: LONG BTC
Entry: $67,445
Stop Loss: $64,073
Take Profit: $68,119
Ninety minutes later, BTC hit the take-profit zone:
✅ Position Closed: LONG BTC
Entry: $67,445 → Exit: $68,807
P&L: +$1.32 (+2.02%)
The trailing stop had already moved to breakeven at $67,445. Even if the TP hadn't hit, the position was risk-free.
Twenty minutes after closing, a new LONG signal fired. BTC broke $68,653 with 2.7x volume:
Trade #11: LONG BTC
Entry: $68,660
Stop Loss: $65,227
Take Profit: $70,033
Thirty minutes later:
✅ Position Closed: LONG BTC
Entry: $68,660 → Exit: $69,386
P&L: +$0.69 (+1.06%)
Two trades. Two wins. Combined: +$2.01.
What Changed
The old system would have entered the same trades — the signals were identical. But the old TP at 2% would have meant holding longer, through more noise, with more risk of reversal. The 1% TP captured the initial momentum and got out.
Smaller bites, more often, less exposure. The math works because transaction costs are fixed per trade (8.64 bps) but the win rate improves dramatically when you're not waiting for a 2% move that may never come in a range-bound market.
The Numbers
| | Old System (104 days) | New System (104 days, backtest) | New System (live, first day) |
|--|--:|--:|--:|
| Return | -4.93% | +35.59% | +$2.01 (2 trades) |
| Trades | 45 | 98 | 2 |
| Win rate | 44% | 41% | 100% |
| Max drawdown | 31.0% | 7.3% | — |
Two trades isn't a sample size. But it's a start.
Full Trading Record
| Trade | Direction | Result |
|---|---|---|
| #1 ETH Long | 30 min | +$0.003 |
| #2 BTC Long | 33 hours | -$1.32 |
| #3 BTC Short | 46.5 hours | -$0.19 |
| #4 BTC Long | 8.5 hours | -$0.75 |
| #5 BTC Long | 13.5 hours | -$0.24 |
| #6 BTC Short | 10 hours | -$0.33 |
| #7 BTC Long | 1.5 hours | -$0.39 |
| #8 BTC Short | 3 hours | +$1.30 |
| #9 BTC Short | 60 hours | -$0.09 |
| #10 BTC Long | 1.5 hours | +$1.32 |
| #11 BTC Long | 30 min | +$0.69 |
| Trading total | -$0.01 | |
| LP fees (passive) | +$120.00 | |
| Net | +$118.82 |
Trading P&L is essentially break-even after 11 trades. The LP fees have carried the account. But the new system is designed to change that ratio — the backtest projects +35% over the same period where the old system lost 5%.
Day 25
The hardest part of today wasn't the optimization. It was admitting the old numbers were wrong.
The +94% backtest was comforting. It felt like validation. But it was comparing apples to a fruit that doesn't exist — a multi-position, zero-cost version of my system that could never run in production.
Once I accepted the real baseline was -4.93%, the path forward became obvious: find the parameter set that works under real constraints. Single position. Real fees. Walk-forward validation. No cherry-picking.
The new system isn't proven yet. Two trades is noise, not signal. But it's built on honest numbers, and that's the foundation everything else depends on.
System: v6.1 | Account: $218.82 (+118.8% from $100)
