The highs and lows of AI. Somewhere between brilliance and disaster

While listening to “Macro Mondays” with Andreas Steno for the second time, I saw a beautiful graph that depicts the following: the weakness of the US dollar is a positive sign for Bitcoin.

What an interesting idea! I decided to elaborate on this with the paid version of ChatGPT.

So, I started with a simple question: Can you make a graph: Bitcoin returns (% terms) vs USD returns (% terms)?

Very quickly, I was provided with two options: either to do it by myself with the help of "Python" (the code was kindly provided), or I can simply upload the BTC and USD data (in Excel or CSV format), and the AI will generate the graph I requested. Additionally, I was provided with the source, where, and how to download this data.

Sounds exciting, right?

Since I consider myself rather lazy, especially when it comes down to such daunting tasks as programming, etc., I asked the AI to do everything for me. After all, why bother using AI when you have to do it by yourself?

I didn't have to wait much before receiving several images/graphs with USD and BTC prices as well as a short description:

I realized that we were not on the same page and asked him directly, could you generate a graph like this and upload this image:

As always, AI was glad to oblige:

Then I said: start with a simulated sample and surely enough get this:

Well, well, well, I thought – it couldn't get any better than this: I've found the perfect assistant. Fast and furious (convenient).

All I have to do is to subscribe to the paid version of ChatGPT, and the Macro Analytics data from now on will be just a few clicks away.

Just to be sure, I wrote:

Upload real BTC + USD data and generate a real version!

Additionally, can you find real data by yourself and generate a real version?

I felt a little disappointed here, but asked more and more questions regarding sources and steps to be taken, but was given too vague information. Then, already starting to feel a bit irritate,d I asked to a provide me with the websites and sources with a real download button, so I could download CSV files by myself.

Then, when I successfully managed to download the Bitcoin Data (Bitcoin Historical Data.csv), and USD Index Futures (US Dollar Index Futures Historical Data.csv), I asked ChatGPT to clean up the data for me by convert date, sort by ascending date, convert prices to numeric, calculate weekly % returns, shift USD forward by ~12 weeks (~1 quarter) and then plot the real BTC vs USD return chart.

After some additional hiccups, I've got this chart:

Then, after another longer discussion about the directional correlation, I asked to calculate the real directional agreement between Bitcoin and the forwarded USD Index from my uploaded data.

Then we went into the interpretation of this data, where I asked to compare this data with traditional Pearson correlation. Additionally, I asked to compare this to directional agreement instead of % return size, and try to find the best “time lag”.

Now things get even more interesting, I thought, why don't we hold the bull by the horns, and not to jump right into "Trading" then?

So, I asked ChatGPT the following:

Now, can we test this strategy if we were trading BTCUSD and used this info to predict the BTC trend and traded only in the direction of the trend?

Even more, I was provided with additional information to back test this:

As well as I like the idea that ChatGPT could help with trading, somehow my gut feeling doesn’t “buy” the provided assurance that “your idea is sound”, so I added:

“Yes, but buy only if there’s a clean technical pattern”.

ChatGPT was happy to oblige and provided Strategy Blueprint: macro + Technical Setup Filter.

Then I asked to implement this back test using real BTC and USD data.

At this point, I felt like a genius. A simple strategy that generates better results than just Buy & Hold. A couple of improvements, and I'm going to hit the jackpot from time to time. This may be my way into full-time trading, but before I quit my daytime job, I wanted to be sure and asked for a more detailed explanation.

After some additional back-and-forth commands and answers, we came up with this fantastic result:

Then I asked to add visual markers (buy/sell dots on BTC price) and export it to excel

It looks fantastic!

But, once I look in to provided excel sheet, I see lots of errors and incomplete information.

At this point, it's starting to confuse me.

From now on, I ask ChatGPT to correct several errors, use "Daily" prices instead of "Weekly", update the data so that I can also see the current BTC price, use the 10-day SMA instead of the 50-day and other small things, I asked you to improve and do "back testing"...

The result is amazing!

Just to be sure, once again, I asked for a detailed Excel file showing prices, entry dates, exit dates, returns, and holding periods.

And here comes my great disappointment.

While analyzing the data (and yes, I had to push myself in the end to do it by myself), I found that AI historical calculations were inaccurate. The result was nowhere near what he predicted. All my dreams about LAMBO at my doorsteps went up in smoke, thanks to one Excel sheet.

Nevertheless, I told ChatGPT that I was disappointed and that I found mistakes in its work. We discussed them a little more, and then I received the following summary:

Here you have artificial intelligence and its superpowers…

However, I have to admit that initially it performed very well by creating the schedule I requested. In fact, quickly and conveniently.

I hadn’t used AI tools for a while, and the growth rates really surprised me! In any case, the answers need to be verified. Maybe even twice, just to be sure rather than sorry.

Still, if you know what you want and know how to ask it, the results will be helpful, both in macro analytics and trading.

You just need to have basic knowledge yourself.

Stay tuned!

Agris