Last year, I used GPT o1 to create a trading strategy. It is the WORST strategy I’ve ever seen
No Exaggeration. This strategy did worse than penny stocks!
Ten months ago, OpenAI shocked the world with their O1 model.
O1 was the first of its kind Large Reasoning Model, a special class of language models that are capable of thinking before it responded. It transformed the world.
To demonstrate its abilities in finance, I had it generate an algorithmic trading strategy. The backtests showed 277% returns over 4 years. I was shocked!
I deployed this portfolio for real-world paper-trading, hoping to prove to everybody the insane power of these language models. And for a while, I was correct. It was DOMINATING the market.
The portfolio was up 270%, more than SPY’s 94%But after the April downturn, everything changed. And now, after 10 months of live trading, this “market-beating” strategy is down a WHOPPING 40%.
While the GPT O1 Portfolio had some early gains, it is now down an astounding 40% since last year and at complete all-time lowsThe Sharpe ratio? Negative. The max drawdown? 52%. It literally performed worse than randomly picking penny stocks.
Here’s what happened in this 10-month long, real-world algorithmic trading experiment.
Recapping How I Got GPT o1 to Create an Automated Trading Strategy
First, let me explain how I was able to use GPT o1 to create a trading strategy.
I built NexusTrade, a no-code platform to help savvy investors create their own algorithmic trading strategies.
I integrated large language models to reduce manual configuration effort and allow the creation of strategies using natural language. When creating a portfolio, the process is as follows:
- Send Request: The message from the user is sent to the server
- Classify Request: From a list of prompts, such as the “AI Stock Screener Prompt” or “Create Portfolio”, the model determines the most relevant prompt to process the request
- Forward to Prompt: We send the message to the most relevant prompt and get a response, often containing initial details and configurations.
- Post-Process Response: Depending on the prompt, we’ll perform additional actions. With the “Create Portfolio” prompt, it goes through more steps.
A Diagram Depicting the Process of Creating a Portfolio of Trading StrategiesThe system first detects that we need to create strategies and so generates a portfolio outline. This includes a description of the trading strategies.
Then each strategy is processed further with LLMs. We generate an action (such as buying or selling an asset) and a condition (such as a stock’s price falling below its 30-day average price).
We then take all of the strategies, combine it with the portfolio outline, and send it back to the user.
In this case, I chose to create a portfolio of trading strategies using GPT o1. And wow, it didn’t disappoint.
A Screenshot of the Original Conversation and the “Omni” Portfolio Generated by the O1 modelUsing the prompt:
“I want a SMA crossover strategy on TQQQ. I want a take profit, but no stop losses – I’m bullish on tech long-term and won’t want to be stop lossed out. I also want to space out my buys and not go all-in at once”
The AI generated a portfolio with a 277% 4-year gain. It has the following rules:
- Buy 2500 dollars in TQQQ Stock when 50 Day TQQQ SMA > 200 Day TQQQ SMA and # of Days Since the Last Filled Buy Order of TQQQ ≥ 7
- Sell 100 percent of current positions in TQQQ Stock when 14 Day TQQQ Rate of Change > 15
These simple rules had insane results. I decided to launch and paper-trade it, to show the world that AI can be used to fully automate algorithmic trading.
Spoiler alert: it cannot. At least not now.
Discovering the Failed Promise of the Market Beating AI Strategy
The way I re-stumbled on this portfolio is by pure coincidence.
I recently redesigned a large portion of my algorithmic trading platform, including the Public Portfolios (now called Community Portfolios page).
The Community Portfolios is a public library of portfolios shared by the NexusTrade CommunityWith this redesign:
- I added performance metrics, graphs, and other indicators onto each card
- I created a button to allow users to clone these strategies with the click of a button
- Users can sort through the best (and the worst) portfolios
Going through my UAT testing for this work, I noticed that the worst portfolio by far was the OpenAI o1 one.
A sorted list of all public strategies, from best to worse. GPT o1-mini has more than 2x the percent loss as the 2nd worse strategyBut it’s not the only strategy that failed. Many of the worse trading strategies deployed on the platform were fully generated by AI.
But why?
Tracing the AI-Generated Portfolio’s Downfall Step-by-Step
Let’s trace the portfolio’s history to see where it went wrong.
1. Initially soars 22% by December
At the very start of the experiment, the portfolio soars, gaining 22% of its portfolio value in 3 short months.
December 15th 2024, the O1-Generated Portfolio reached its all-time highIt went downhill after this.
2. Holding during a dramatic downturn
After that, the market became choppy, remaining roughly flat after the election of President Trump. It stayed this way until March.
Then, an incredible downturn happened. Trump announced tariffs on most of the world, which shocked the global economy.
From Feb 14th to April 7th, the S&P500 fell roughly over 18%. The portfolio was fully allocated throughout the worse of the bear market and loss over 44% from its starting value.
The trough of the portfolio was down 44% on April 6th.But it gets even worse.
On April 9th, Trump announced he was delaying the tariffs for 90 days. This caused the entire market to soar. TQQQ did exceptionally well, rising more than 38%.
TQQQ Percent Change from April 8th to April 25th is nearly 40%This caused our portfolio’s selling conditions to trigger right before the massive relief rally, solidifying its monumental loss permanently as the rest of the market soared to all-time highs.
This part is critical to the failure. Even if the condition was altered to say “and the TQQQ positions are not down”, the strategy would have done monumentally better. But instead, it sold everything at the very bottom.
Absolutely devastating.
A Graveyard of the Empty Promise of AI-Generated Trading Strategy
Scrolling through the Community Portfolios from worse to best, we see several trading strategies that were nearly 100% generated by AI, including:
But what happened?
Static and Unchanging Trading Strategies
These trading strategies aren’t actively being managed by me, an AI, or anybody.
They are “fire and forget”.
Because of this, they lack larger scale contexts on the market. When stocks turned previously, the specific rules did well on past data, but couldn’t translate to the real-world (nearly Black Swan-like) crash due to the tariffs.
That’s the biggest flaw.
Strategies That Aren’t Grounded in Evidence or Theory
But in addition to being fire and forget, these strategies don’t have real evidence that they will translate to outsized returns.
No journaling or self-reflection. No hypothesis for why the rules worked. Nothing.
Thus, there was nothing we could really blame… we don’t know why these strategies worked!
Incredibly High Risk Leveraged Trading Strategies
This goes without saying, but these strategies used leveraged instruments. They were designed from the get-go to be incredibly high risk! This high risk demonstrated itself clearly, as this strategy fell drastically as the overall market soared!
Evidence of the riskiness was evident (but not emphasized) in the original article, where we saw the sharpe ratio and max drawdown. This combined with the lack of a real risk management plan is a recipe for disaster.
The original backtest performance for the O1 portfolioWith many of these AI-Generated trading strategies being some of the worse in the platform, you would think that would be a massive blow to using AI for algorithmic trading.
You would be incredibly wrong.
The Great Deal of Silver Linings
This experiment marks one of the most successful experiments I’ve launched with the NexusTrade platform. I collected data to learn from, extracted real-world results, and gathered a large list of silver linings and future opportunities.
Discovering What Works and What Doesn’t
This experiment showed me exactly what AI is good for and what it’s not.
Out of the box, a language model cannot just one-shot a trading strategy.
However, AI can do other things which helps with algorithmic trading, including:
- Translating my ideas from natural language into insights.
- Help me test my ideas 2000x faster than any sort of manual process
- Filter out the noise to focus on the things that truly matter.
For example, if we sort the Community Portfolios from greatest to worse, we see several highly profitable trading strategies with insane sharpe and sortino ratios.
Many of the top strategies shows a clear pattern (with 5/6 of them were created by me). Of these, most of these capitalized on emerging trends (for example, the rise of artificial intelligence) or had robust backtest results more based on stock fundamentals.
For example, in this article, I talked about how NVIDIA and Google were some of the best AI stocks in the market.
Since writing that article, this simple Google/NVIDIA rebalancing strategy is over over 30%.
This simple NVIDIA/Google rebalance strategy capitalizes on the rise of artificial intelligenceWhile we can’t say for sure what will happen in the future, we can continue to gather data, gain insights, and become more knowledgeable on what works and what doesn’t.
Massively Increased Confidence In Backtesting Accuracy
Another HUGE silver lining is that I gained MASSIVE confidence in my system’s backtesting accuracy.
I built NexusTrade to share code between backtesting and live-trading with an event-driven architecture. The idea is the system is agnostic to whether its performing a backtest or whether its running in the real world. It just runs the same for everything.
Being a software engineer (and not a financial engineer), I always wondered if I built the system correctly to accurately backtest strategies.
This experiment gave me massive confidence I did.
The backtest peformance from 10/15/2024 to 07/12/2025Both the backtest and live-trade portfolios are down 40%. The final portfolio value for the backtest is $5,986.51. And the value for the live-trade is $5,997.88. This discrepancy is tiny!
The performance of the live-trading strategyWhile some metrics such as the sharpe ratio and sortino ratio are slightly off, they are computed in slightly different ways (using different historical history values). With these results, we can conclude that the backtests are a reasonable approximation of the live-trading performance.
For a no-code trading platform, this is everything.
The Real Lesson: AI Amplifies, It Doesn’t Replace
After 10 months and a monumental loss, this experiment taught me something crucial: Language models are powerful tools, but they’re terrible fortune tellers.
The GPT o1 strategy failed spectacularly because it lacked three essential elements:
- Active management — Markets change; static strategies don’t adapt
- Theoretical grounding — Without understanding why a strategy should work, we’re just gambling with fancy math
- Risk awareness — AI optimizes for backtests, not real-world volatility
But here’s what actually worked: When I used AI to accelerate my own ideas — like the NVIDIA/Google AI play that’s up 30% — the results were extraordinary. The difference? Human insight augmented by AI, not replaced by it.
This experiment answered one question definitively: Can AI create a one-shot winning trading strategy? No. Not yet.
But it raised better questions:
- What if AI continuously adapts strategies based on market conditions?
- What if we train models on why the best strategies succeed, not just their rules?
- What if we combine human expertise with AI’s processing power?
The graveyard of failed AI strategies on my platform isn’t a cautionary tale — it’s a roadmap. Every failure teaches us what doesn’t work, bringing us closer to what does.
Ready to build something better? Join me in exploring the intersection of human insight and artificial intelligence. Share your strategies, learn from others, and let’s discover what’s possible when we use AI as a tool, not a prophet.
Because the future of trading isn’t about replacing human judgment — it’s about amplifying it.
