Will the stock market soon include just robots trading amongst themselves? Can AI models like Child GPT create highly profitable trading strategies? I bet you’ve asked yourself these questions. In this article, we will explore these areas as I guide you through a journey of using Child GPT’s interpreter to create a trading strategy.
A little teaser: I was able to achieve incredible returns by using Charge GPT. Later, we will find out what specifically I was able to achieve.
To start, we will fit Charge GPT’s code interpreter with two years of daily candles of 500 stocks. Ideally, a larger dataset would be preferred, but Charge GPT has its limits.
Our goals are first to develop a profitable trading strategy that will be backtested on the provided data, and second, to repeat this process for an unconventional trading strategy. I will display both of these tasks side by side, offering a comparison of how Charge GPT will approach each task.
Lastly, we’ll tap into Charge GPT’s potential to generate reusable code. We’ll create relation metrics and cluster returns of power tickers. Let’s dive in!
I’ve loaded our data into ChatGPT and set our prompt on the left. You’ll see Charge GPT crafting a conventional strategy, and on the right, an unconventional one. Both begin with basic descriptive statistics, helping to figure out what data has been passed.
The unconventional strategy immediately jumps to the first idea. Its strategy involves entering long positions on close after a stock demonstrates strength by making a significant move from the low of the day. The idea is to hold the strength overnight and exit when a certain profit objective is met.
The conventional route asks about strategy preference. I choose momentum, but Charge GPT surprisingly proceeds with a mean reversion strategy. It enters a long position when a stock is oversold and short stocks when they are overbought. Both paths then define entry and exit signals and calculate returns for each position.
It is a sensible approach to the implementation of the backtest. A conventional strategy yields a big loss over the last two years. It shouldn’t come as a surprise given that the market has been challenging and displayed mean reversion characteristics, with momentum fading fairly quickly.
In contrast, the unconventional strategy returns 72 percent. It is a good illustration of what an experienced strategy developer should immediately do in such a situation. Whenever performance looks too good to be true, typically it is. One should then review the code and try to find biases and weaknesses in the data.
The past two years had numerous mean reversion opportunities, so it can indeed be successful. However, a 72 percent return is not that impressive. Here is our more promising use case: we will ask Charge GPT to construct calculation metrics and cluster returns of our tickers.
Charge GPT was able to recognize and fix some data challenges before constructing relation metrics. Then it displays the dendrogram, which usually demonstrates different clustering options. This is a great example of how Charge GPT can create useful code snippets that we can apply to our data.
I must admit that my gains were educational, and Charge GPT has been a big time-saver. But actually, I wouldn’t recommend trading strategies that we have developed today, at least without a much better understanding of your code, market regimes, data, strategy, and a more extensive backtest.
When it comes to trading, retail robots aren’t replacing humans just yet. In conclusion, use AI like Charge GPT to enhance your work, not to replace it. Sampling and feature engineering are crucial; they are your responsibility and won’t be done by retail AI models anytime soon.
The quality of your input data will significantly impact the results. Use AI tools to challenge your strategies on specific data sets, as you might have missed certain improvement opportunities.
As of now, publicly available AI cannot automatically design great trading strategies. It’s a powerful aid but not a silver bullet. And finally, learn to distinguish signal from irrelevant random noise in the data. AI can provide valuable information, but it has a big tendency to overfit your data. Challenge all results automatically created by AI models.
Good luck on your modeling and trading journey! Watch my previous videos to learn how to download past stock data so that you could also ask Charge GPT to develop trading strategies for you. Until next time!
