Honest receipts · real Binance data · after fees

We built two trading bots to beat crypto.
They couldn't.

Two AI-assisted bots ran on real market data for months — one on Ethereum, one on Bitcoin — trying every popular way to outsmart the market. Here's exactly what happened, with nothing hidden.

Both bots are switched off. Everything below is a frozen snapshot of real backtests — nothing is trading live.
Meet the bots

Two coins, one honest verdict

Same engine, same AI, pointed at the two biggest coins. Click either to explore its results below.

Ξ
Aether
Ethereum · ETH
Bought every dip and sold every bounce on Ethereum — its settings tuned in real time by a reinforcement-learning optimiser and a Claude AI advisor.
market conditions where it lost to simply holding
See Aether's last run, history & settings →
BitGain
Bitcoin · BTC
The identical strategy and AI pointed at Bitcoin instead. Different coin, same outcome — a clean test that the result wasn't just one unlucky market.
market conditions where it lost to simply holding
See BitGain's last run, history & settings →

Nothing here is hidden. The full source for both bots — the trading engine, every strategy, the reinforcement-learning optimiser, the Claude AI advisor and the backtests behind these verdicts — is open on GitHub: https://github.com/joubja/PaperTradingBot. Read it, run it, or check our numbers yourself.

The receipts

strategy vs. just holding

We didn't just test one idea. Pick any popular strategy and see how it did across real crashes, slow bear markets, flat chop and bull runs — every trade charged a realistic fee and slippage.

Each card shows the money you'd actually have after the run. Beating buy-and-hold in a downturn just means losing less — not making money.

The part nobody admits

"But it has AI" — why that didn't help

These weren't dumb bots. Each one carried a self-learning optimiser (a multi-armed bandit that rewarded settings that made money) and a Claude AI advisor reading the market. People assume that's the missing ingredient. It isn't — and here's the honest reason why.

1
You can't optimise an edge that was never there. The AI tuned how the bot bought dips — how deep, when to sell. But buying dips on spot has roughly zero expected profit before costs, and a negative one after fees and slippage. Optimisation just finds the settings that lost the least on past data. That's curve-fitting to history, not a prediction that holds up next month.
2
The bandit assumes a stable game. Markets aren't. A bandit only converges when each choice has a roughly fixed payoff. Here the best setting in a calm month is the worst one in a crash — and which regime comes next is exactly the thing nobody can predict. So it kept confidently chasing the last regime's noise into the next one.
3
An LLM can't see tomorrow's price. The Claude advisor reads news and indicators, but it has no privileged information about where ETH or BTC goes next — nobody does. In practice it added confident-sounding, costly noise, and sometimes overrode the optimiser for the worse.
4
We fixed the bugs. It didn't matter. We found and corrected real wiring faults in the reward loop. The ceiling didn't move: a perfectly-tuned optimiser on top of a no-edge strategy still nets to about break-even before costs — and just holding wins after them.
Before you @-reply us

Is it all trading bots?

No — and we won't pretend otherwise. What loses is the popular kind: bots that try to predict short-term price direction. On crypto, and by the same logic on forex and stocks, those lose to simply holding once you pay real fees — which is why the large majority of retail algo and day traders end up underwater.

The strategies that genuinely survive the data don't predict at all. They're market-neutral — long-short and pairs trades that harvest the price gap between related instruments, carry that collects a structural yield instead of betting on direction, and true arbitrage and market-making. Those edges are real. They're also small, fiercely competed by funds with scale and near-zero costs, and at retail fees usually get eaten before you see them. So: not "bots can't work" — rather, the bots being sold to you, the ones that promise to call the market, are the ones that don't.