Why most crypto bots fail (and what 6 years of backtesting taught me)

I lost money to four trading bots in one year. Then I got stubborn.

Okay so. 2023.

I don’t talk about 2023 much because it’s embarrassing, but I think the embarrassing version is more useful than the clean one, so here it is.

I lost money to four different crypto trading bots that year. Four. Different companies, different “strategies,” different Discord servers full of people posting rocket emojis. Same ending every single time: a green line that goes up for a while and then does this slow, awful curl downward that you keep telling yourself is temporary.

The first one cost me $500. I remember the exact number because I remember staring at it. The fourth one I won’t even tell you, because my girlfriend at the time found out and made a joke about it that I still think about, and that’s all you’re getting.

Anyway.

After the fourth one I did the thing that broke-but-stubborn engineers do, which is decide that the entire industry is wrong and I’ll just build it myself. Which is insane. But it’s also kind of what happened, so.

This is the post I wish someone had shoved in my face in early 2023.

First, the part where you recognize yourself

You’ve seen the ad. The bot’s got a backtest that looks like a staircase to heaven. You read the reviews — mostly five stars, a few “support was slow lol,” nothing that scares you off. You sign up for the cheapest plan because you’re not an idiot, you’re just curious. You give it $500.

Week one it’s up 2%. And here’s the dangerous part — you don’t think “small sample.” You think I found it. I found the thing.

Week three you’re down 8% and you’re googling “is drawdown normal.” (It is. That’s not the point.)

Week eight you’re down 35% and now you’ve got two options, both bad: yank it out at the bottom and lock the loss, or double down and pray. I did both, in different years, with different bots. They both feel terrible.

Then a year passes. You forget the feeling. You read about a new bot with an even better backtest.

And you sign up again.

I did this loop four times. Sound familiar? If it doesn’t, congratulations, you’re smarter than me. If it does — okay, good, keep reading, because I finally opened the hood and I can tell you what’s actually rotting in there.

Mode 1: the backtest is lying to you about fees

This is the big one. This is the one that got me first and got me worst.

A lazy backtest grabs historical candles, runs your entry and exit rules over them, adds up the imaginary profit, and shows you a beautiful chart. It looks like trading. It is not trading. It’s trading with all the painful parts deleted.

Three costs get quietly rounded to zero:

Fees. Binance Futures is 0.04% taker, 0.02% maker. Sounds like nothing. But if your bot trades 20 times a day — and a lot of them do, that’s how they justify the subscription — you’re bleeding roughly 1.5% a month just in fees. That’s ~18% a year. Gone. Before the strategy does anything.

Funding. Perps charge funding every 8 hours. Hold a long perp in a calm market at +0.01% average funding and you’re paying something like 11% a year just to exist in the position. Get into a hot alt — I did this with SEI once, funding was brutal — averaging +0.05%, and now you’re paying like 55% a year. The strategy has to beat that before you make a cent. Most backtests pretend funding isn’t a thing.

Slippage. When your bot tries to slam a $10k order into SEI/USDT during a fast candle, you do not get the price you saw on the chart. You get something 1 to 10 basis points worse, every time, and it adds up. Backtests that use mid-price are basically writing fan fiction.

When I finally built my own engine and turned all three of these on, every single strategy I tested dropped 20–40%. Every one. And here’s the cruel twist: the strategies that looked the best in the naive backtest were usually the worst once costs were real — because the reason they looked great was that they traded constantly, which meant they were paying the most.

So. Simple test. If a bot won’t show you a backtest with fees, funding, and slippage turned on — walk. They’re either lazy or hiding it. Doesn’t matter which.

Issue #2 — it memorized the last bull market

Here’s a thing nobody tells you.

Most retail bots are tuned on whatever data the founder happened to have lying around when they built the thing. Usually the last 6 to 12 months. And if those 6 to 12 months were a bull market — congratulations, the bot didn’t learn to trade, it memorized what a bull market looks like.

Every bot that launched in 2021 was a smoking crater by Q3 2022. All of them. They were optimized for that gorgeous up-and-to-the-right tape from 2020–2021, and then BTC went from $69k to $16k and the bot had literally never seen that pattern in its life. It’s like training a dog only indoors and then being surprised it loses its mind in a thunderstorm.

The fix is boring and annoying, which is why most people skip it: test across every regime. The 2020 COVID flush. The 2021 mania. The 2022 collapse. The 2023 chop (god, the chop). The 2024 grind-up. The 2025 sideways nothing. Six years, six totally different markets. If your strategy can’t live through all six, it’s not a strategy. It’s a recent-history parrot.

I dragged my three survivors through all six:

• Momentum (Donchian breakout + Chandelier exit, if you care). Lived through all 6. But — and I’m saying the bad part first on purpose — it eats 13–46% drawdowns in chop. Forty-six percent. You have to be able to stomach that. The payoff is the trend years are insane, +23% to +109% depending on the regime and how much risk you dial in. It’s a strategy with a personality, and the personality is “mostly losing small, occasionally winning enormous.”
• Neutral (cointegrated pair stat-arb). Survived 5 of the 6. It got wrecked in 2022 when everything correlated to 1 and there was nowhere to hide. Honest about it. +7–22% a year, 14–27% drawdowns, needs you to rotate pairs.
• Carry (delta-neutral funding arb plus quarterly basis). Survived all 6, because it’s not predicting anything — it’s harvesting a structural premium. Boring. Low drawdown. +7–15%. The strategy equivalent of a savings account that occasionally gets nervous.

Notice I led with the worst trait of each one. The drawdown, the bad year, the weakness. I do that on purpose now, because anyone who sells you a strategy by leading with the upside is selling you a daydream. The drawdown is the part that actually decides whether you survive.

Number 3, and honestly this is the one that should scare you most

Martingale. Except nobody calls it that.

They call it “DCA bot.” They call it “smart averaging.” They call it “grid trading with intelligent rebalancing.” Sometimes all three are the same thing wearing a clean shirt.

Here’s the trick: when the trade goes against you, you buy more. Each new buy drags your average price down. If the market bounces, you exit green and feel like a genius. If it doesn’t bounce — if it just keeps going — you die. Not “lose 20%.” Die. Zero.

And the thing that makes it so seductive is it works 95% of the time. It really does. You’ll get weeks of clean little wins and you’ll think the people warning you about it are paranoid. Then the 5% shows up.

Want to know how to make a martingale bot look like a money printer? Backtest it on 2020–2024 and quietly leave out the May 2022 LUNA week and the November 2022 FTX week. That’s it. That’s the whole magic trick. Put those two weeks back in and the strategy goes to zero on the chart, right in front of you.

Every bot that took my money had some flavor of this baked in. None of them said the word “martingale.” But the position sizing said it for them.

The fix is unglamorous: size every trade like it’s the only one you’ll ever make. Risk 1–2% of your account, no doubling down, and just accept that losing trades are the rent you pay for being in the market. Yeah, it caps your upside too. That’s the deal. That’s the correct deal.

#4 — “set and forget” is a lie if the bot doesn’t know what month it is

You cannot run a trend-following bot in a chop market. You cannot run a mean-reversion bot in a trending market. Same exact strategy, opposite outcomes, depending entirely on the regime — and the regime changes every few months and does not ask your permission first.

There are two real fixes and neither is exciting.

One: run several strategies that don’t move together, so when one’s bleeding another’s earning. Two: gate each strategy by regime — only let the trend bot run when there’s actually a trend, etc.

I do the first one. Equal weight across Momentum, Neutral, and Carry. When 2022 took Momentum out behind the shed, Carry just kept quietly collecting. When 2024 ripped and Momentum went vertical, Neutral was the boring laggard. Each strategy’s worst year was another strategy’s best year. That’s the whole trick. Diversification doesn’t get you bigger returns — it gets you the same returns with a flatter, more survivable equity curve. The combined line is smoother than any single one of them. That smoothness is the entire point.

And the fifth one, which nobody warns you about because it’s not sexy

Execution. The gap between the backtest and the actual button-pressing.

Funding arbitrage reads like free money on paper: buy spot, short the perp, collect funding, done. One page of math. I read that page and thought, easy.

Then I tried to build it. You need:

• A spot API key with trading turned on — a separate one from your futures key
• Two orders that fire at the same time (or in the exact right order) so you’re never sitting there accidentally long because one leg filled and the other didn’t
• Delta tracking, because if you get a 100% spot fill but only an 80% perp fill, you are now 20% long and the entire point of the trade is gone
• Reconciliation logic for partial fills, exchange rejections, random network failures at 3am
• And then a graceful exit, which is the same nightmare in reverse

I underestimated this part by, no exaggeration, a factor of ten. The math of funding arb fits on a napkin. The code that actually runs it without blowing up is thousands of lines, most of which exist purely to handle some specific weird thing the Binance API does on Tuesdays.

So when you’re sizing up a bot, ask the vendor one question: “what happens when one leg of a multi-leg trade doesn’t fill?” If they don’t have an instant, specific answer — the thing has never been stress-tested with real money. Move on.

What actually survived all of this

Eighteen months. That’s how long I spent building the engine before I trusted anything enough to put real money behind it. Eighteen months for what started as a revenge project. I’m not proud of the timeline but I’m proud of what came out.

Three strategies made it:

Momentum — directional, runs on 15 large-cap perps, rides trends and cuts losers fast. Win rate’s only around 15%, which sounds terrible until you realize the 15% that win, win big. Drawdowns of 13–46% even in good years. Returns from +23% all the way to +109% depending on regime and risk setting. Not for the faint of heart.

Neutral — pair stat-arb. Finds two assets that historically move together (LINK/AVAX is the textbook one), trades the spread when it stretches. Win rate ~49%, small wins, 14–27% drawdowns, +7–22% a year. The grown-up in the room.

Carry — delta-neutral funding arb on high-funding alts, plus quarterly basis carry on BTC/ETH. The boring one. The one I’d actually recommend to my dad. +10–15% a year on the funding side with roughly 2% max drawdown, +7% on the quarterly side with about 5%. Doesn’t predict anything. Just collects.

Low correlation between the three, so the combined portfolio runs flatter than any single one. I keep saying that because it’s the most important sentence in the post and I don’t trust you to have read it the first two times.

The thing I built

I turned the engine into a product. It’s called getperpetuum.com, it’s in beta, and it’s for non-US people only — Binance.com perps are restricted in the US, so legally none of this is available to American residents. Sorry. Not my rule.

The whole design philosophy is just “do the opposite of the bots that robbed me in 2023”:

• Fees, funding, and slippage are on by default in every backtest. You have to go out of your way to lie to yourself.
• Every parameter has a sane default and a safe range, so you can’t fat-finger yourself into oblivion.
• Every strategy ships with the full 6-year backtest and you can re-run it with your own settings in a few seconds.
• Pricing is just crypto, $19 Trader and $99 Pro, no weird billing games.
• The landing page tells you the max drawdown in the first scroll instead of burying it in 8pt gray text.

There’s a free tier — 5 backtests a month, no live bot — if you want to poke at it.

And look, if the product isn’t for you, genuinely, that’s fine. Take the five lessons and go apply them to whatever bot you’re eyeing next. You do not need me for that part.

The short version, for the people who scrolled

1. Naive backtests lie. Fees + funding + slippage knock 20–40% off real returns.
2. Most bots just memorized the last bull market. Force a 2020 / 2022 / 2024 minimum test.
3. “Smart averaging” is usually martingale. Cap risk at 1–2%, never double down.
4. No strategy works in every regime. Diversify by strategy type, not just by coin.
5. The execution code is harder than the strategy. By a lot.

The bots that took my money in 2023 didn’t fail because trading is impossible. They failed because they cut a corner on every single one of these five things. The strategies themselves are fine. The discipline is what’s missing, and discipline doesn’t fit in a 30-second ad.

This isn’t a complete guide, by the way. I’m definitely wrong about some of this and someone in the comments will tell me so, probably rudely. Fine. But if you’ve been through your own version of the four-bot loop, I’d actually love to hear it. And if you’re about to sign up for bot number five — at least run the five tests first. That’s all I’m asking.

Part of a series on building crypto trading infrastructure. Next one’s about funding rate arbitrage — the boring strategy that quietly earns ~10% a year. Subscribe if that sounds like your kind of boring.

*getperpetuum.com is in beta and serves non-US users only. Crypto trading can lose you your money. Backtests don’t promise anything about the future.*