Published
One thing I've learned running bots: everything sounds plausible before you test it. Volume spikes reverse. Support levels hold. Small frequent wins compound. These all feel true until you actually run them across enough data. Then you get a number, and the number is usually boring in a way that ends the conversation immediately.
Here are four strategies I built and then retired — with the specific number that closed each case.
1. Volume reversal: 7% win rate
The idea: when volume spikes sharply, price often reverses. You buy the spike-reversal and take a small profit on the snap-back. It sounds like it has a mechanical logic. It doesn't.
My bot ran this live and in paper testing. Win rate across all sessions: 7%. Not 70%. Seven. I was right about one in fourteen trades. The losses were normal size; the wins were also small. There's no scenario where 7% works unless your average win is fourteen times your average loss — and this strategy definitely wasn't structured that way.
Killed it in version 18 of the bot. Moved on. Never looked back.
2. Range-reject shorts: profit factor 0.34
Range-reject shorting means: price hits a resistance level, shows weakness, you short the rejection. It has a certain textbook logic — resistance is resistance, right?
Profit factor across backtests: 0.34. For every dollar won, I lost about $2.94. That's not a strategy with a slight issue I could tune out. A PF below 1.0 means the strategy destroys capital over time. At 0.34 it destroys it quickly.
The problem is structural: crypto has a persistent upward bias across most of the pairs I was testing. Shorting resistance in an asset class that spends most of its time in an uptrend is fighting the market's default motion. You can win individual trades — the setup can look clean — but the distribution across many trades goes against you. The math doesn't care how good the chart looked.
3. Naked DCA: 61% rescue rate hiding in plain sight
This one's subtler. "DCA" sounds disciplined — you're buying more of what you believe in at better prices. The issue is when the strategy structurally depends on averaging down to get any wins at all.
I ran one coin (LIGHT) through 1,600 simulated trades using a range-reclaim setup that had a DCA layer built in. Result: 61.3% of all winning trades required the bot to add a second buy at a lower price first. Not "sometimes used DCA." More than six in ten wins only happened because the strategy bought again into a losing position.
Reported win rate was 36.8%. Looked usable until you realized that the 36.8% was partly manufactured by doubling down. In simulation that works fine — unlimited patience, no gaps, no funding rate bleeding. Live, it means you're sitting on a red position more than half the time, waiting for a recovery that usually comes but occasionally doesn't. When it doesn't, you're not down one position size. You're down two.
I keep the strategy in simulation. I don't run it live.
4. Fee-blind scalping
This one doesn't even need a backtest to understand. It just needs arithmetic that most people don't do before they start.
On a small account, a typical taker fee is roughly 0.05–0.1% per side. That's 0.1–0.2% round-trip per trade. For a strategy that makes 1,600 trades over a period of time, the friction alone is 160–320% of your starting position in cumulative fees. The strategy has to return more than its own friction before it makes a cent. High-frequency crypto scalping in the sub-10-minute range almost never clears this bar on small capital.
I know this because I ran five different automated systems trying to scalp, totaling 81 live trades in one session alone, net result: −$0.567 after fees on a 15% win rate. The strategy "worked" on some trades. Fees consumed the margin.
Before your next scalp, run the fee calculator. Figure out what your break-even win rate actually is after round-trip fees. It's usually higher than you think.
What I run instead
I still have one strategy in active simulation: range bottom reclaim on a handful of specific crypto pairs, tuned across 220,000 parameter combinations per coin. It survives fees, has reasonable DCA dependency (under 45% on the better coins), and doesn't need to short anything. It's not exciting. I don't have live proof yet that it works in the real market.
That's the honest answer. Kill what the data kills. Run what survives. Check your position size and liquidation distance before every entry — the math that seems like a formality is usually the math that saves you.