Too many idiots are using OpenClaw to trade. Here’s how to trade with AI the right way
How to ACTUALLY trade using AI agents
When I first saw Clawdbot (now called OpenClaw) appear on my LinkedIn feed, I dismissed it as the latest AI-hyped product with no real use-cases. One week later, I learned I was wrong.
It’s so much worse than that.
This wrapper is no more useful than using Claude Code with different tools. Yet somehow, it soared to over 150,000 stars. People are hooking it up to their Twitter accounts and Gmail APIs, and are relearning with blood that the decades of hard-fought cybersecurity lessons aren’t just suggestions.
But that’s not the purpose of this article.
I was unfortunate enough to stumble on this tweet.
Press enter or click to view image in full size Some moron giving their “AI” $2,000 to tradeAssuming it’s real (which I unfortunately think it is), this idiot gave OpenClaw $2,000 to trade.
Based on what you might ask? Nothing but vibes.
No backtesting. No analysis. Just “does this trade sound like a good idea to a statistical model that can predict language?”
How embarrassing.
The crazy part is that this isn’t the first time someone has done this. In fact, researchers Tsinghua and Beijing University have created their own benchmark to test how well LLMs trade.
The inputs to the LLM are news articles, price data, and fundamentals, so at least it’s based on something. But it’s still incomplete. No parameter optimization, no ability to learn from the past or adapt based on market conditions.
This is ALL WRONG.
I’ve been building AI solutions for algorithmic trading for over 5 years now. I’ve built Deep Reinforcement Learning bots at Carnegie Mellon University, launched no-code trading platforms, and built production-grade AI query systems.
Real traders, at least those that are successful, don’t trade based on vibes or gut feel. They don’t care if a trade idea “sounds” good. They care if it IS good.
They know from performing backtests, statistical analysis, and developing automated trading plans. Here’s how you can trade with AI the right way.
The Wrong Way of Applying AI Agents to Trading
When most people apply AI agents to algorithmic trading, they are applying the agents as if they are discretionary traders.
Take StockBench as an example.
With StockBench, they test a plethora of different LLM agents to see how well they can trade. Each day the LLM receives:
- Opening price for each stock
- Up to 5 news articles from the previous 48 hours
- 7 days of historical trade actions (what the agent previously did with held stocks)
- Fundamental data for stocks selected for deeper analysis: market cap, P/E ratio, dividend yield, 52-week high/low, recent quarterly dividends
If you’re thinking “this sounds pretty idiotic; LLMs can’t trade”, then you’re right! Of the dozen LLMs that were tested, the best performing one earned 2.5% and had a risk-adjusted return of 0.031.
Press enter or click to view image in full size Performance benchmark of StockBench: The best model was Kimi-K2, and it earned a sortino ratio of 0.042, which is mediocreIf you’re even a little-bit quant-minded, you know that this is terrible! And it’s not surprising. We’re trying to test how well an LLM can trade, but we’re just throwing a bunch of stock data at it and saying “trade this money, don’t make mistakes!”
We’re going about it all wrong.
Instead of telling the LLM to make decisions like a discretionary retail trader, we need to frame it like an institutional quant. Instead of the tools being “what is SPY’s price today”, the tools need to be “under these market conditions, this set of rules returned on average 8% in the next 6 months”.
In other words, instead of telling the LLM to trade, we need to teach it to engineer trading strategies. This is how.
Using AI to Build Battle-tested Trading Strategies
Instead of asking Claude to make trades, I ask Aurora (a Claude-Powered AI agent) to make a trading strategy.
Press enter or click to view image in full size I typed one sentence. Aurora responded with a full research plan.You can think of it as Clawbot attached to an institutional trading test. Aurora does a process of creating hypotheses, backtesting strategies, and identifying strengths and weaknesses.
I literally typed one sentence, and Aurora asks pointed follow-up questions, generates trading strategies, and even launches genetic optimizations. The entire process takes 15 minutes or less.
These aren’t just simple strategies. We’re not buying VOO below a simple moving average or selling NVIDIA if its relative strength is greater than 70. We’re building institutional-grade systems.
Press enter or click to view image in full size An example strategy created by the AI; it looks at all US stocks and applies a list of operations to find target stocksTake a look at this example. One portfolio can do the following.
- Look at all US stocks
- Filter to S&P 500 members only
- Filter to stocks with an AI Report score above 3 — NexusTrade’s AI-generated stock report gives each stock a rating, and this throws out anything rated 3 or below. So you’re only keeping S&P 500 stocks that the AI considers favorably.
- Rank the survivors by their 30-day price rate of change (momentum)
- Take the top 15 — of the S&P 500 stocks the AI likes, grab the 15 with the strongest upward price movement over the last month.
And with one prompt, we can launch a dozen. And, similar to AI tools like Cursor and Claude Code, I have full autonomy. I can “vibe trade” and let it run loose and create crazy ideas, or I can run it in semi-automated mode, make pointed suggestions, or ensure it picks certain stocks.
The end result is far better than StockBench — a plethora of profitable trading strategies each with their own strengths and weaknesses.
Press enter or click to view image in full size Different portfolios generated by AuroraSome portfolios are stronger in bull markets and weaker in bear markets. Some portfolios have great defensive capabilities, but underperforms when the market soars. And each portfolio is deployable with zero-risk, running automatically in a sandbox environment so we can see how it actually performs.
Just look at the difference between this and OpenClaw. While StockBench’s best portfolio had a sortino ratio of 0.04, even the worst portfolios generated by this process have a sortino ratio of 2 or higher during the testing period.
The difference is night and day.
Press enter or click to view image in full size The strategy chosen by Aurora earned 36.94% vs SPY’s 15.97%, more than 2x the return, with a better risk-adjusted return and lower drawdown. From one prompt. At the end of the process, Aurora picks a “winner” based on the results. I decided to backtest it across the past year, from 01/31/2025 to 01/31/2026. And in this period, the best strategy objectively dominated.
The best part is that this isn’t a walled garden hidden behind a $2000/month paywall. You can get this power for free, without risking a single dollar.
How YOU can vibe trade with AI?
Traders that win aren’t giving their API credentials and brokerage passwords to unsecured bots trading based on vibes.
They trade with trading strategies.
The key difference is this. Don’t treat the LLM as a discretionary trader that can’t learn. Use it as a strategy engineer. The model doesn’t execute trades; it builds rules that can be evaluated, scrutinized, refined, and optimized.
You can build your own using open-source tools and free resources on my blog.
Or you can try it yourself with a free tool called NexusTrade.
NexusTrade is how AI trading is supposed to work. I’ve been building it for over 5 years, before ChatGPT was revealed and OpenAI’s best model was text-davinci-003. Before Cursor and Claude Code and “Agents”. It does what Twitter users wish OpenClaw can do, but just better.
Because unlike worthless tools that scrape twitter and returns “buy” or “sell” decisions, NexusTrade is designed to trade like a quant. It has tools that can create algorithmic trading strategies, analyze backtesting results, and launch genetic optimizations on promising candidates. It can also act as a researcher, analyzing historical data for correlations and searching for news. It can even perform “deep research” by connecting with APIs like Perplexity.
The AI becomes an engineer.
Because as the StockBench paper suggests, LLMs aren’t good at being discretionary traders. After all, they’re language models. But they’re incredible at designing systems, generating hypotheses, writing rules, and iterating on strategies based on real data.
That means, they can create trading strategies, launch backtests, read results, and initiate optimization processes. These are the tools that set Aurora apart.
If you’ve read this far, you probably already get it. You’re not the person giving an LLM $2,000 based on vibes. You’re the person who wants to know if a strategy actually works before risking a dime.
Now, you can.
