Coffee and algos

In some ways, 21st-century New York resembles a 19th-century metropolis more than it does the 20th. These days the city sidewalks are dotted with fruit stands and coffee carts just like on the Lower East Side in the late 1800’s. Entrepreneurial immigrants are willing to withstand sub-zero conditions to run these things 24/7 in order to save paying rent. In NY everyone has a fruit guy and a coffee guy.

I’ve known my coffee guy for nearly three decades. We’ve both seen our hair grow gray together. Since I am up for the London open everyday we often chat in the wee hours of the morning about the World Cup or his family back in Egypt, but when I return for my second cup after dropping my youngest off at school it's all business as the line is long and he is trying to fill more orders in a minute than Starbucks does in an hour. Often he will see me in line and just fill my order (large, milk only) without us exchanging a word. In that sense he is like the human embodiment of an algo. Reading the proper pattern and fulfilling the order in the most efficient way possible.

Except

This week  I bought a new espresso maker and wanted a small rather than a large cup because I wanted to make a red eye at home.

Too late. My big cup coffee order was ready by the time I reached the cart and I just shrugged and accepted it, not wanting to disrupt his rhythm. But that 2 second interaction made me realize why algos in our business always fail in the end.

Trading algos of course are always trained on past data patterns while financial markets are always adapting to new information that could radically transform the future price action. A classic example of that is the orange juice market which had a  very repeatable and predictable pattern of falling prices in winter and rising moves in summer because the supply and demand would coincide with the Florida growing season. But sometime in the 1980’s Brazil decided to get into the business and soon became a major supplier of oranges on the world stage. Since Southern hemisphere seasons are inverted orange juice patterns changed as well and the seasonal buy summer/sell winter trade completely reversed causing more than minor bit of damage to the specs that happily traded it for years.

Trading is always a binary decision. Price either continues in the prior direction ( trends) or reverses ( mean reverts), so the whole game  is relatively simple but hard because the flip in the pattern - which is the hardest part to predict - is always where all the profit is.

I use the exact same algo to trade both continuation and reversal setups. It is a great tool that manages everything for me - size, risk, structure - except one thing: determining when a regime change is about to occur.  For now that part, small as it is, is still a matter of art rather than science. 

It may not be so for much longer. The last few weeks everyone  has been enamored with ChatGPT, the latest AI tool that  has taken the Internet by storm. ChatGPT uses something called transformer technology to create completely new ideas and concepts with just a few simple prompts. 

Twitter is full of startling examples of its work from the prosaic (write me a rental contract for a North Carolina property) to the absurd (rewrite Jack and Jill in the form of a Shakespearean sonnet). 

ChatGPT is the first “non-robotic” robot but as of now it has only been trained on data up to 2021. In the  future it may be powerful enough to be trained in real time at which point it may be able to discern changes in supply and demand and therefore price regime shifts faster than any human alive. That day will bring its own set of challenges and problems for the markets, but for now the trading  algo remains a tool rather than a replacement because at least for now the future is still unknowable to such a degree that humans can still compete in the game.

---