My crypto trading bot started with a simple idea: connect it to Binance, trade on one-second candles, and catch every small move from low to high. Over time the project drifted to slower strategies, and I let it, without ever testing the original idea myself.
So I did. This post is the test, the numbers, and why they came out the way they did. It's an engineering exercise, not trading advice.
The setup
The data is Binance's official archive of spot one-second candles for BTC, ETH and SOL, seven days from 31 October to 6 November 2026. That's 604,800 candles per coin, about 1.8 million in total, each file checked against its published SHA-256 checksum.
I tested four kinds of rules, both long and short:
- Momentum: the last 3 (or 5) seconds all closed higher, so buy (or the reverse, so sell).
- Breakout: the close broke above the previous 60 seconds' high, or below the low.
- Mean reversion: the close sat 2.5 standard deviations away from its 60-second average, so bet on it coming back.
Each rule ran with four target/stop pairs (5/5, 10/10, 20/20 and 20/10 basis points) and two maximum hold times (30 and 120 seconds), on all three coins. That's 96 combinations.
A backtest is only as honest as its fill rules, and these were deliberately strict:
entry = cand + 1 # a signal at second t fills at the open of second t+1
p = o[entry]
# f_tp / f_sl: the first second the target / stop is touched
stop = (f_sl <= f_tp) & (f_sl < hold) # both in the same second: the stop winsThey were generous elsewhere, because they assume no spread and no slippage. Real fills would only make the results worse.
Fees came from Binance's regular tiers, per side: 0.10% on spot, 0.075% with the BNB discount, 0.05% as a futures taker and 0.02% as a futures maker. A round trip pays twice.
Even perfect foresight can't pay the fees
Before testing any rule, I asked a simpler question: if you knew exactly where the price would be a few seconds from now, how often would the move be big enough to cover a round trip of fees?
| Seeing this far ahead | BTC, spot fees | BTC, cheapest futures fees |
|---|---|---|
| 1 second | 0% of the time | 0.2% |
| 10 seconds | 0% | 5.7% |
| 1 minute | 1% | 29% |
| 5 minutes | 7% | 58% |
The median one-second move for BTC is zero, because most seconds the price doesn't change at all. A crystal ball doesn't help when there's no move to predict.
The rules: a real edge, far too small
Before fees, 71 of the 96 combinations made money. There's a genuine pattern: short bursts of movement tend to continue for a moment. To check that this wasn't just a flaw in the simulator, I fed it random buy and sell signals. They averaged about 0.05 basis points per trade, essentially zero. The edge the rules found was real.
It was also tiny. The median rule earned 0.36 basis points per trade before fees, and the best earned 0.83. The cheapest realistic round trip, futures maker on both sides, costs 4 basis points. Spot costs 20.
After fees, 0 of the 96 combinations made money, at every fee level.
The best rule was the same on every coin: breakout, with a 20 bp target, a 10 bp stop and a 30-second hold. I picked it using the first four days, then ran it on the three days it hadn't seen. On BTC it traded 1,400 to 1,700 times a day. On the unseen days it earned +0.77 bp per trade before fees, which is −3.23 bp after the cheapest fees and −19.23 bp at spot fees. With 1,000 USDT per trade, the four tuning days alone came to roughly −1,900 USDT at maker fees and around −11,000 USDT at spot fees.
What the numbers say
The edge was six to thirty times smaller than the cost of trading it. That gap doesn't close with a cleverer rule. It's the structure of the market at that timescale: price moves within a second are smaller than what the exchange charges to trade, and the people who do profit there have servers next to the exchange and fees near zero.
The foresight table also points the other way. The longer the holding period, the more often normal price moves are bigger than costs. That's why the version of the bot that survived its own backtests trades on four-hour candles, not seconds.
What I took from it
Test the idea you're attached to before you argue about it. A short, honest backtest settled in an afternoon something I had been taking on trust. And when a backtest comes back clearly negative, check the simulator with random signals before you believe it. Here, that check is what showed the small edge was real, and too small to matter.
Seven days is one stretch of market, and a different week would change the details. It wouldn't change the arithmetic: an edge under one basis point can't pay a fee of four.