Build Your Own Crystal Ball

I recently gained two important insights about the Fed and markets. Taken separately, either one would have been significant. Then I realized the separate insights were actually connected at a deep level.

This analysis could lead to a major breakthrough in my already sophisticated predictive analytic algorithms I use to forecast markets. What follows is a first-of-its-kind explanation.

The first insight started as a casual conversation during a ceremony honoring a benefactor of one of the top research universities in the world.

Such occasions are always interesting because the turnout includes trustees, professors, the benefactor himself, and his close and often wealthy friends. While waiting for formalities to begin, I spoke with a top economics professor whom I had known for years.

We talked about a mutual friend (also an economist) best described as the ultimate Federal Reserve insider. While this conversation took place exactly as described below, I have used pseudonyms to preserve confidentiality. Let’s call the economics professor “The Professor,” and the Fed insider “Big X.”

The Professor and Big X

Big X has been a top economic advisor to Fed Chairs Ben Bernanke, Janet Yellen, and Jay Powell over the past fifteen years. He’s not on the Fed Board of Governors, but he’s more powerful than any Fed governor except the Chair.

He’s in every FOMC policy meeting and his name is found in the minutes along with about 40 other attendees. But he’s also in his office near the Fed Chair’s office when the hallway is quiet, the policy crowd is not around, and important decisions are actually being made.

I’ve known the Professor for over 50 years. I’ve known Big X for over 15 years. I was responsible for recruiting the Professor to an economics research center led by Big X. The difficulty was that Big X was lured to the Fed by Ben Bernanke around the time of the global financial crisis in 2008, something I discussed with Bernanke when we met in South Korea.

I told him I was mildly annoyed at the pick-off but, of course, we were all pleased that our director was so highly regarded.

Big X returned briefly to academia but was recruited back to the Fed by Janet Yellen who had the same need as Ben Bernanke. As a result, my direct contact with X had been intermittent, but the Professor has been a reliable alternate channel.

Why a Lawyer Should Head the Fed — Not an Economist

When I spoke with the Professor last week and asked about X and the Fed, he said something that caught me by surprise. Jay Powell is a lawyer and a good one. I worked with Powell years ago when he was at Treasury and I was counsel to one of the Treasury market underwriters, the so-called “primary dealers.”

I have a lot of respect for his intelligence. Maybe that’s because I’m a lawyer too and my bias is showing. I always thought a lawyer was a good choice for Fed Chair because they are more analytical and see both sides, whereas economists tend to be one-trick-ponies who see everything through a neo-Keynesian lens.

When I asked whether X might be returning to the university soon, the Professor said, “No, he’s needed more than ever at the Fed. Powell’s not an economist, so he relies entirely on [X] for economic and policy advice.”

There it was. Powell is the face of the Fed and nominal leader, but X is calling the policy shots and writing the explanatory notes. Fed policy is just an extension of hard-shell neo-Keynesianism.

That explains why rate hikes have continued longer than markets expected, and why the Fed is ignoring clear signs of a severe recession. Rate hikes won’t stop until inflation is back in its box (it’s not right now). Recession and higher unemployment are just collateral damage.

AI Outworks Any Wall Street Analyst

The second insight I gained was in reading a newly published paper titled, Can ChatGPT Forecast Stock Price Movements? Return Predictability and Large Language Models (2023) by Alejandro Lopez-Lira and Yuehua Tang of the University of Florida.

By coincidence, I’m involved in artificial intelligence at UF through my role at the Florida Institute of National Security (FINS). Both the authors and my FINS colleagues have access to the HiPerGator AI computer (the third-fastest non-government computer in the world) for purposes of AI applications. ChatGPT is a form of generalized pre-trained transformer (GPT) technology that is the latest advance in AI.

The paper describes an experiment in which ChatGPT reads massive amounts of market literature using natural language processing (NLP) and large language models (LLMs) to predict stock prices.

In a nutshell, the application scans news headlines about particular companies. The AI scientists then ask Chat GPT if the headline is good news (answer: “YES”) or bad news (answer: “NO”) for the company in question over a defined time horizon.

Based on multiple headlines and answers, a ChatGPT sentiment score is compiled where “YES” is mapped to 1, “UNKNOWN” is mapped to 0, and “NO” is mapped to -1. Regressions are then run on the next day’s actual stock returns.

The paper concludes, “ChatGPT sentiment scores exhibit a statistically significant predictive power on daily stock market returns.” In other words, AI can beat Wall Street analysts when it comes to predicting the market.

A Real-world Example

The paper offered a real-world example of how AI can outperform other forms of headline readers used on Wall Street. In an intellectual property case involving Rimini Street versus Oracle, it was reported that Rimini Street had been fined by the court.

Ravenpack, a typical Wall Street news analytic tool, scanned the headlines and rated the news as negative, giving it a -0.52-sentiment score. This was likely based on a simple keyword reader that saw “Rimini Street Fined $630,000 in Case Against Oracle” and rated words like “fine” and “against Oracle” as negative.

Common sense knows the headline is positive for Oracle because their opponent just had a setback in court. That tends to support Oracle’s effort to protect its IP. ChatGPT reached the same positive conclusion because it can read billions of documents (or a targeted subset in this case) and put language in context with nuance. The simple news scanner can’t do that and stops at first approximations when it sees the world “fine.”

Why not just use common sense then? The answer is that even the most diligent human analyst can read maybe twenty or so stories per day. ChatGPT can read millions (as targeted) backed up by billions of documents (again, as targeted) and combine its sentiment scores with automated execution to front-run the analysts. It appears that ChatGPT really can beat the market.

FedSpeak Manipulates the AI Sentiment Score

How does the ChatGPT study relate to my friend Big X? The Fed is influential in markets, but it’s less powerful than many believe. It has some impact on the short end of the yield curve and can affect mortgage rates (which are often linked to short-term Treasury rates), but that’s about it. The idea that the Fed can actually steer a $25 trillion economy or eliminate unemployment is mostly a myth.

Still, it’s a myth the Fed cultivates. That’s where X’s power derives.

By coming up with the right mix of policy rationales and Fedspeak and putting that into the hands of Jay Powell, Big X can move markets and influence investment in ways desired by the Fed.

As long as people believe the Fed is all-powerful and act accordingly, then the Fed can exert real power. It’s a confidence game, but it works.

Now, think about ChatGPT. It reads headlines. What if you can create your own headlines? What if you’re powerful enough to shape the text that ChatGPT is reading? This means you can manipulate the sentiment score and catalyze market activity in any desired direction. ChatGPT is just doing what it’s designed to do. It won’t know it’s being spoon-fed Fed pablum.

Other market participants are already on to this. They’ll be creating stories to flood the natural language processors with words and phrases that even sophisticated LLMs won’t be able to distinguish from objective truth.

You can call this fake news or market manipulation of whatever you like. In the intelligence community, we call this the wilderness of mirrors. The point is, it works. Even with machine learning, it will take years (if ever) for machines to sort out the real from the planted.

Even humans struggle with this. That’s why governments lie.

How to See the News Before It’s News

Where does this leave the human analyst, including myself? With the right tools, we’re in a remarkably strong position. The key is to understand how the systems work, reverse engineer them, and legally front run the front-runners.

Here’s an oversimplified algorithm:

<Thanks to Big X, we know what “news” he’ll promote>
<Because of ChatGPT algos, we know what conclusion AI will reach>
<The market will trade in line with ChatGPT>
<The momentum won’t last because the “news” is invented>
<Fade the market>

There’s more to it than that, of course. Some news really is the news and should be followed. ChatGPT might get better over time at sorting out the fakers (then again, the fakers might get better using their own ChatGPT).

The most powerful skill of all is seeing the news before it’s news using proprietary predictive analytic tools. That’s what I bring to the table.