What Artificial Intelligence Predicts for Stocks
Yin Luo was working on artificial intelligence and machine learning long before the terms entered the investing lexicon. Today, he uses gleanings from a model he created 14 years ago to lead quantitative research, economics, and portfolio strategy at Wolfe Research. But the longtime champion of AI’s promise to transform investing says that some of the recent buzz around it borders on science fiction.
Before joining boutique research firm Wolfe in 2016, Luo was a top-ranked quantitative analyst at Deutsche Bank. Growing up in northern China, Luo learned to code at age 10. When it came time for college, he decided to try something new, majoring in economics and then finance, and digging into the work of modern portfolio theorists like Eugene Fama and Kenneth French, among others. As an analyst at CIBC, he began to experiment with early quantitative work, creating the base of his machine-learning model in 2004. It’s now in its fourth iteration.
Luo, 43, says he joined Wolfe in part because the boutique research firm was unencumbered by the legacy infrastructure and processes of a big bank. Plus, it offered him almost double the budget to pursue AI and machine learning. Luo doesn’t think that robots will replace money managers soon, but he does think that machine learning can give investors an edge. He incorporates these models to make global market calls for investors who use the firm’s research. So far this year, he says, the model spotted increasing risk from global trade battles, as well as the regulatory risks for technology companies like Facebook (ticker: FB), before they rattled markets.
We sat down with Luo at his New York office to talk about how AI differs from quantitative analysis, why adjective-happy executives can be a red flag for investors, and why trade tensions could be bad for energy stocks.
arron’s: Let’s start with the basics. How do you define AI?
Luo: Big data, AI, and machine learning are related, but not necessarily the same. AI and machine learning are often used with big data because you need machines to help analyze lots of data, but AI works with or without big data. AI trains machines to learn from human challenges and then to make decisions almost on their own.
How close are we to machines investing on their own?
That’s more sci-fi. This stage still requires very heavy human interaction, with humans designing models and the machines learning, rather than machines doing the investing.
What about the AI exchange-traded funds that have launched recently?
That’s more traditional quant research and relabeling it. It’s not even what real AI might look like. In the foreseeable future, AI is more about giving fund managers better insight, so it’s complementary rather than a replacement for them.
What’s the difference between traditional quantitative research and AI?
Machine learning marks a fundamental shift in research philosophy and opens up a completely new world for investors. The key difference is that in traditional financial research, you form a hypothesis before you do the testing, like believing that valuation or price momentum drives stock returns. Then you collect data and back-test. Essentially, if you try 8,000 different options across different countries and over various periods, you will find something that supports the hypothesis. I have never seen a back-test that doesn’t work.
In AI or machine learning, there is no human hypothesis. Instead, you just say, “I believe there may be some relationship between, for example, valuation and future stock performance.” You don’t know if it’s positive or negative. You can also condition it on other variables or include other information, like credit-card payments or tax-roll information, and let the machine decide what is relevant.
This is very important: The machine will learn on its own. As the market evolves, the machine should be able to identify new patterns and relationships. If done properly, it should show you what variables work in what types of environments. Whatever the data says is what you invest in. That’s a big psychological shift for investment firms where senior management has been trained in Fama and French and linear thinking.
How do you use AI or machine learning?
One example is with management presentations. As humans, we think we are good at seeing truth from lies, but in reality we aren’t. Even if you are a professional interrogator, the hit rate is about 50%. Plus, the amount of information [provided by companies] is overwhelming. Regulatory filings often include 150,000 words, and there are hundreds of earnings calls in a reporting season. We can train the machine to learn from human language. For example, when management uses descriptive language with a lot of modifiers, it tends to mean they are overconfident or potentially lying. Each firm has its own way of presenting, so it is important to look at relative change. The model also combs through the risk disclosures in 10-K and 10-Q filings, which tend to list hundreds of risks and are very broad. When a company changes or adds a risk, it’s a red flag.
What did you pick up from the last earnings season?
Overall, sentiment is still fairly positive, though inflation and interest rates are an economic risk. The biggest concern is political, with trade and regulatory risk. Our model picked up these red flags in Facebook, for example, before the market reacted to them. In Facebook’s fourth-quarter earnings call, we saw negative sentiment in management discussion and analysis, including management expressing concerns on its own business, and significant changes in the language used in that section, highlighting the possibility of change in the business environment.
Is the bull market losing steam?
Yes, but it’s not heading into a bear market yet. We are still bullish but warn investors that the downside risk is significant, which is different from last year. We are at the peak point in the economic cycle, growth is showing some signs of weakening, and inflation continues to rise. As for interest rates, we care more about the change in rates. The move in the 10-year from 1% to 3% was big, though 3% is still very low. In a typical economic cycle, it is 61 months, or five years, from peak to peak, so every five years we tend to have a rate cycle. This isn’t your typical growth cycle, but we are approaching the peak.
Does that mean a recession is coming in a year or so?
Not a recession but a slowdown. The chances for a recession are fairly remote. The Federal Reserve tends to hike interest rates at the peak and continue as the economy contracts. There are two complications this time, though. Normally, at the peak stage of an economic cycle, we should see a balanced budget. This time, we have had a massive tax cut and fiscal stimulus when the economy is already running at full steam. That may force the Fed to hike more aggressively.
And the other complication?
The other big risk is a trade war. We have an algorithm that goes through every major media and social-media site, looking for key words related to trade conflict, like “trade tensions,” “tariffs,” and “quotas.” It’s gauging the wisdom of crowds. Once people are talking about it, it is more relevant to stock performance. Today, [that chatter] is at its highest point since we started tracking these things in 2003.
What else do you do you look at?
We look at insider transactions, but not in the way others do. On their own, they are useless because often buying is based on a prespecified formula, with an executive buying 100 shares each week or selling a certain amount every month. They become interesting when serving as confirmation, such as when a company beats earnings and delivers an upbeat management call. If management also bought an outsize number of shares, then that is a high-conviction idea. Or it can be useful when there is asymmetric information, like in biopharma, where there is a long development pipeline as companies file for approvals with the Food and Drug Administration. Insiders always know more about their own drug development than outside investors, so insider buying or selling is more insightful than for a consumer company.
What are some of your contrarian stock calls right now?
In our concentrated long portfolio, the most contrarian calls are Peoples Bancorp [PEBO], MBT Financial [MBTF], and medical-device maker AngioDynamics [ANGO]—all of which rank high on traditional metrics, as well as on our proprietary algorithms that use machine and natural learning to identify red flags in management presentations and pinpoint high-conviction insider transactions. In our short portfolio, some of the more contrarian calls include Golden Entertainment [GDEN], which develops and manages casinos, and water-purification company AquaVenture Holdings [WAAS].
More broadly, what parts of the market are attractive?
In the near term, our macro models favor U.S. large-cap stocks, global real estate investment trusts, and gold, which is a more balanced portfolio given heightened trade and geopolitical uncertainties. With U.S. stocks, we are bullish consumer discretionary, technology, and industrials over the medium horizon, and are negative on consumer staples and telecom services, where fundamentals remain relatively weak and momentum has been negative. We are also underweight energy, in part because of trade tensions.
What does trade have to do with energy?
Trade concerns lead investors to stay away from risky assets, like energy. Plus, trade tensions pose a risk to consumption from Europe and China; we could see a potential demand shock in oil. In terms of investment styles, we are still bullish on price momentum. We look closely at short interest. The smart money or hedge funds are still shorting low-momentum stocks and betting on high-momentum stocks. That confirms the trend. If you look at earnings revisions, earnings growth is still pointing toward momentum stocks. In terms of asset allocation, you have to put money somewhere, and we are more concerned about fixed-income and yield-oriented assets, so we like stocks over bonds.
As the industry tries to get a handle on AI, what words of advice do you have for them?
We have been asked by many of our clients to help find them AI talent. But any new computer-science or financial-engineering graduate puts AI on a résumé and claims to have done neural networks. But when you ask more questions, many have just used an off-the-shelf system and run a simple simulation. That doesn’t mean you can build an algorithm, and that’s not AI. There are probably fewer than 20 reputable AI or machine-learning university programs in the world. We hire only Ph.D. grads, and there are just three to five per college, so it’s a very small pool, and many want to stay in academia or work in Silicon Valley.
Thanks, Yin.