I Gave Four AI Trading Bots $15,000 Each. The Smartest One Lost the Most Money.
I built four AI trading bots and gave them $15,000 each. The smartest one lost the most money.
Four bots. Four markets. Built in Claude Code. Grok scanning Twitter for sentiment. GDELT feeding real-time news intelligence. All running off Mia, my AI agent on a Mac Mini sitting in my living room.
No human intervention. Just signals and execution. Six weeks. Let them fight.
When the world tested the bots for me
Two days into the experiment, a war broke out. The US and Israel struck Iran. Oil hit $100. Markets dropped. Volatility spiked across every asset class.
I couldn't have designed a better stress test if I'd tried. My bots kept trading through all of it.
The scoreboard
Here's how they finished after six weeks:
- Crypto bot: +$2,557 (+17.05%)
- Polymarket bot: +$488 (+5.14%)
- US stocks bot: +$148 (+0.99%)
- ASX bot: -$1,136 (-7.58%)
Why the dumb bot won
The crypto bot watched one thing: market sentiment. Fear goes up, it buys. Greed goes up, it sells. That's it. No analysis. No conviction scores. No cross-referencing multiple data sources. One signal, one action.
It returned 17%.
The simplicity was the advantage. When Iran was getting bombed and every other bot was trying to process what that meant for seventeen different variables, the crypto bot looked at sentiment, saw fear, and bought. It didn't need to understand geopolitics. It just needed to read the room.
Why the smart bot lost
The ASX bot had everything. Detailed fundamental analysis. Technical indicators. Conviction scores. Cross-market intelligence feeds. News sentiment. Sector rotation models. It was, by any reasonable measure, the most sophisticated of the four.
It generated 48 trade signals over six weeks. It executed zero of them.
Every time it identified an opportunity, it found a reason to wait. The conviction score wasn't quite high enough. The cross-market signals were mixed. The news feed introduced uncertainty. The risk-reward ratio was borderline.
Perfect analysis. Zero action. Down 7.58%.
The Polymarket surprise
The Polymarket bot was the most interesting performer. It spotted the Iran strike two days before the other bots fully processed it, the best call across all four. Prediction markets move on information before traditional markets do, and the bot was positioned to capture that edge.
It wasn't the biggest winner, but it was the most insightful. It saw something the others missed, acted on it, and captured real value.
The design lesson hiding in the data
All paper capital. The signals and failures were real. The money wasn't. But the lesson translates directly to product design.
The bot that knew the most never pulled the trigger. The bot that knew the least never hesitated. This is the same pattern I see in design teams, product committees, and roadmap meetings every week. It's why founders hire designers for the wrong reason: they optimise for analysis instead of action.
More information doesn't produce better decisions. It produces more reasons to delay. More data points create more edge cases. More analysis creates more uncertainty. The team that studies the market for six months ships nothing. The team that picks a direction and moves learns in weeks what analysis can't teach in months. AI has made this gap even more dangerous: if a competitor can rebuild your product in a weekend, hesitation is fatal.
Overthinking is expensive. Shipping is free.
The best products aren't built by the teams with the most information. They're built by the ones who knew enough to act and had the conviction to do it.
---
