It's October 12, 2021. Bitcoin is at $56,000. The chart looks like a rocket. Twitter is screaming "SUPER CYCLE." My best bot — the one I trust with real money — finally flips green and buys.
53 days later, the bot sells. Bitcoin is at $49,202.
That trade just lost 12.1%. The biggest single-trade loss in 4.8 years of Watchdog history.
And I'm not going to fix it.
If you've ever looked at a backtest curve and thought, "if I could just shave off the bad trades, this would be perfect," you're one of those people. One of those who optimize themselves into losing strategies. After this article, not anymore.
Why I'm telling you this
Last week a reader scrolled the new DM+LD signal-history heatmap and spotted exactly the trade I just described. Bot enters near the top. Bot exits after the crash starts. Loss locked in.
His instinct was the same instinct most traders have: we should add a tighter stop. The bot should sell faster.
I tested it. The numbers don't agree. They say something more interesting.
In this article, I'll show you:
- The exact mechanics of the worst trade Der Wachter (The Watchdog) ever made
- The full per-trade record across 4.8 years (25 trades, 13 wins, 12 losses)
- Why the win/loss ratio of 3.3:1 matters more than the win rate
- Why a tighter stop would have made the bot worse, not better
Buckle up.
The trade I'd love to delete
Here's the chronology, straight from the bot's signal log:
| Date | Event | BTC Price |
|---|---|---|
| 2021-07-20 | Local bottom (after summer crash) | $29,783 |
| 2021-10-12 | Watchdog flips green → BUY | $56,000 |
| 2021-11-08 | Cycle Top | $67,465 |
| 2021-12-04 | Watchdog flips red → SELL | $49,202 |
Read that twice. The bot waited 84 days after the bottom — by which point Bitcoin had already nearly doubled — to enter at $56k. Then it held through the cycle top. Then it kept holding while the price slid 27% from the peak. Finally, on December 4, it sold at $49,202.
Net result on that single trade: −12.1%. Or, in real money on a $10,000 paper position: a $1,210 hit.
That's the Watchdog at its worst.
See it for yourself
Words aren't enough here. The full picture is in the heatmap below. Hover over October 2021 — you'll see the green band where the bot was holding BTC, then the red band starting in early December where it went back to cash. The white price line shows what happened in between.
What you'd see on the heatmap
If you zoom into October-December 2021, you'll see two clean facts:
- The bot bought late (the green band starts well after the July $30k bottom)
- The bot sold late (the red band starts well after the November $67k top)
Worst of both worlds. Late on entry, late on exit.
This is also what the data quantifies if you average all 25 trades:
Entries lag the local trough by an average of 17 days, missing 12.3% of the rebound. Exits lag the local peak by an average of 11 days, giving back 7.0%.
So technically, entries are worse than exits. But losing realized money on the way down feels worse than missing unrealized money on the way up. Loss aversion in a chart.
Still with me? Good. Because this is where it gets interesting.
The math that saved me anyway
Out of 25 trades over 4.8 years, the Watchdog won on 13 (52%) and lost on 12 (48%). That's almost a coin flip. Most traders would call that a broken system.
It's not. Look at the size of the wins versus the losses:
| Metric | Value |
|---|---|
| Average winning trade | +15.9% |
| Average losing trade | −4.8% |
| Win/Loss ratio | 3.3 : 1 |
| Largest single win | +50.5% (2024-01-16 → 2024-04-21) |
| Largest single loss | −12.1% (the 2021 trade) |
Read that win/loss ratio again: 3.3 to 1. For every dollar lost on the losers, the winners returned $3.30. The bot doesn't have to be right to make money. It has to be RIGHT BIG when it is right.
The 2021 trade that hurt? It was followed in 2024 by a +50.5% trade that paid back the loss four times over. Then a +21%, +21%, +18%, +20%, +20% sequence. The losers fund the winners; the winners pay them back with interest.
Remember this: in trend-following systems, the win rate is decoration. The win/loss ratio is the engine.
Why I tested fixing it (and why I won't)
The honest move was to actually try the fix instead of just defending the system. So I overlaid a trailing stop — exit BTC if price drops more than X% from a rolling high — on top of the existing signal, then re-ran the entire 4.8-year backtest. Eight variants. Stop at 8%, 10%, 12%, 15%, 20% × lookback windows of 20 and 30 days.
The results, sorted by total return:
| Variant | Total Return | Max Drawdown | Trades |
|---|---|---|---|
| HODL (no bot) | +132.0% | −49.5% | 0 |
| DM+LD original (no stop) | +202.0% | −34.7% | 49 |
| 8% stop / 30d | +96.7% | −32.8% | 59 |
| 10% stop / 20d | +166.2% | −30.3% | 57 |
| 15% stop / 30d | +136.5% | −34.8% | 53 |
| 20% stop / 30d | +212.8% | −32.3% | 51 |
Look at the tight stops (8% to 15%): they cost between 30 and 100 percentage points of return over the period. Why? Because every minor correction triggers the stop, the bot exits, and then the price recovers without it. Whipsaw eats the strategy alive.
Only the 20% / 30-day stop is roughly free: +11pp better return, 3pp better drawdown, 2 extra transitions. Tiny improvement. Not worth touching a system whose worst-case is bounded at -12% per trade.
The 2021 trade was bad. Trying to prevent it would have been worse.
What this means for your own trading
Three takeaways, no jargon:
One: don't optimize away your losers. They're the price you pay for the structure that makes the winners possible. A bot with 100% win rate either has a tiny sample, an overfit, or a massive hidden drawdown waiting.
Two: track the win/loss ratio, not the win rate. A 50% win rate with a 3:1 ratio crushes a 70% win rate with a 1:1 ratio. Most retail trading content focuses on the wrong number.
Three: a single bad trade isn't a broken system. It's a sample. Out of 25 trades, you'll get one or two ugly ones. The system survives them. You survive them — if you sized correctly.
Now it's your turn
Open the full DM+LD signal heatmap and scroll to October 2021. You'll see exactly the trade I just described. Then keep scrolling. Find the +50.5% trade in early 2024. Find the +21% in November 2023. Watch the same logic that picked the bad trade pick the great ones.
Same bot. Same rules. Different outcomes, by design.
That's how trend-following works. It's not magic. It's just math that respects asymmetry.
Honest update: I tested every filter, individually
After this article first published, I went back and tested the bot harder than ever. The Watchdog is supposed to use three filters: a moving average (SMA), a dumb-money flow indicator (DM), and a liquidity-dance gauge (LD). The pitch is "three independent checks. All must agree to buy." It sounds robust.
I turned each filter off, one at a time. Here's what came back:
| Variant | Return | Drawdown |
|---|---|---|
| Full Watchdog (SMA + DM + LD) | +203.7% | -34.7% |
| SMA only | +208.4% | -34.7% |
| SMA + LD (no DM) | +208.4% | -34.7% |
| SMA + DM (no LD) | +196.5% | -34.7% |
| DM only (no SMA) | +94.7% | -76.6% |
| LD only (no SMA) | +39.3% | -76.6% |
Read that twice. SMA-only beats Full Watchdog. SMA + LD ties Full. Adding DM actually makes things slightly worse. Without the SMA filter, the bot is mostly broken (DM-only loses to HODL, LD-only loses badly).
What this means in plain English: the moving average filter is doing all the work. The two on-chain signals (DM and LD) are tracked, but the bot's decisions don't actually change because of them. They're transparent, they're public, but they're not the edge.
I'll be honest: that surprised me. I'd been telling people the Watchdog is a "three-signal confluence bot." Turns out it's a one-signal trend filter with two well-instrumented passengers.
The 3.4:1 win/loss ratio that makes the bot work? That's the SMA filter doing its job. Not three filters. One.
This doesn't break the bot. It still beats HODL. It still keeps drawdown bounded at -35% vs HODL's -77%. It's still the only one of my ten live bots with a documented per-trade asymmetry that explains why it can lose half its trades and still make money.
It just isn't what I told you it was.
If you've been holding on to the "confluence" framing in your head, swap it out. The Watchdog is a 100-day moving average with 4% hysteresis on Bitcoin. That's it. The rest is decoration.
I trust the bot the same as before — the per-trade math hasn't changed, the historical results haven't changed, the worst-case is still -12%. But I trust the marketing less. And I'd rather you know the truth than read a story I no longer believe.
— Dominic, the guy who tested 200 strategies (including a few of his own bad ones) so you don't have to. And who occasionally has to update articles when the data tells him he was wrong.
Sources
- Full DM+LD signal heatmap — interactive, all 1,753 days from 2021-07-05 to 2026-04-22
- The Watchdog bot card — current state and live signal
- Methodology page — how every BearBullRadar bot is validated
This article describes a backtest result. Past performance does not predict future results. This is not financial advice.
