Is AI Sophisticated Enough To Be Useful Yet? That's The Wrong Question To Ask.

Summary

• AI is often viewed as not quite ready for investment purposes.
• This view may be rooted in a misperception of the value of AI.
• AI doesn't need to be sophisticated to be useful, much of its usefulness comes from its ability to dynamically assess quantitative rules we already use.

There’s a common sentiment that AI isn’t yet sophisticated enough to beat the stock market. This is a curious view, since people have historically been pretty successful coming up with quantitative rules that beat the market, and AI isn't all that different from that approach.

There’s a fairly long history of quantitative rules or formulas designed to beat the market. AI doesn’t need to be complex - at all - to join this party. If all AI ever did was what we already do in a static way - but did it dynamically – that would represent a major innovation. The problem with quantitative rules has never been that they’re too simple to work, it has always been that they’re static and only work for a little while.

One famous example of a really successful quantitative rule is Robert Shiller’s claim that price to earnings ratio over the long run predict stock returns. Shiller is a pretty impressive guy. He’s an Economics Professor at Yale; he co-founded the financial analytics company Case Shiller Weiss, which was acquired by Fiserv in 1991; he won the Nobel Prize in Economics in 2013; he was elected President of the American Economics Association in 2016; and he writes a regular column for the New York Times.

He developed a quarterly index based on the price to earnings ratio called CAPE (Cyclically-Adjusted Price to Earnings), and famously showed that over the course of 20 years a buy-and-hold strategy based on CAPE beat the market.

So, the problem with these quantitative rules clearly isn’t that they can’t work. Very basic, unsophisticated rules have worked. The problem has always been that once people know they work, they stop working for regular people. For instance, right now, CAPE doesn't seem to be a very good investment strategy at all for someone who was rebalancing quarterly. 

Consider a simulation where an investor put $100,000 in the market in mid-2009. They keep the best 15 stocks, updating their opinion every quarter as new financial information is released. So, each quarter they rebalance using a Shiller-type index (figure 1). The details of the simulation don’t matter much, the point holds regardless of simluation inputs. That strategy does worse than cash.

Figure 1: Shiller's CAPE v. Market

In fact, most famous fundamentals based formulas have done badly recently. Consider the Benjamin Graham valuation method. Graham is considered the father of value investing, and was an finance professor at Columbia and UCLA. He wrote a famous book called The Intelligent Investor, which Warren Buffet has cited as a big influence.

In that book, Graham claims that an appropriate valuation of a company - based on financial fundamentals - is:

Value = Earnings x 1.5 x (8 + 2 x Growth) 

This formula actually does a good job of predicting market capitalization, but Figure 2 shows that buying stocks that are relatively underpriced based on the gap between the fundamental price and market price doesn't do better than the market (but it does do better than Shiller).

Figure 2: Graham's value-method v. Market

Another popular fundamentals-based valuation method is a discounted cash flow valuation. Discounted cash flow is similar to the Graham formula, but takes free cash flow, growth and opportunity cost as inputs: 

Value = Free Cash Flow x (1+Growth)^25 / (1+MarketReturn)^25

Still a no on that one. See Figure 3.

Figure 3: DCF v. Market

Dividend based strategies are very popular. One might want to value a stock based on the present value of expected future dividend payments. Assuming dividends are always rolled back into purchasing more stock, and not used as income, this method does no better than cash. See figure 4.

Figure 4: Dividend strategy v. Market

Finally, I had a finance professor in undergrad that told us you'll never beat a monkey throwing a dart at the NYTimes finance page. I realized later he was stealing a famous quote from Princeton professor Burton Malkiel who wrote that "A blindfolded monkey throwing darts at a newspaper's financial pages could select a portfolio that would do just as well as one carefully selected by experts." Anyway, that strategy still does surprisingly well - it does just as well as Buffett, who consistently beats the market.

Figure 5: Monkey darts strategy v. Market

I expected this to do better than cash, but the idea behind the quote is that the market is a random walk, so a random pick should do as well as the market. I didn't expect the random picks to beat the market! Looking into it a bit, it seems this strategy does well by not focussing so much on recent results, particularly recent profits, which are highly variable and therefore a noisy signal. The Shiller method tries to do this too, but ends up weighting heavily on companies that did well a generation ago, that the market thinks are becoming obsolete. The random number picks up enough undervalued companies that may have been unlucky in recent earnings, but are otherwise financially healthy, so it beats the market. It's possible to do better than this, for example, by removing companies that are obviously in financial distress by whichever metric you like.

But if the random number works because it picks up under the radar companies (because of earnings or for any other reason) it must be possible to focus on that ans beat this already decent bar by a considerable margin. For example, just looking at the smallest companies may do better than a random number if small companies are no more likely to be in financial distress, but are less likely to be on investors' radars. Indeed, these companies do beat the market (and Buffett and the dart-throwing monkey) by a pretty substantial margin.

This kind of logic is all that AI does. It’s not that complicated. The intuition is that it comes up with a bunch of random formulas similar to CAPE, Graham, dart-throwing monkeys, etc., looks for what works and what doesn't, and then drills-down to refine. In doing so, it develops pretty simple quantitative rules, as we’ve always done, and have historically had considerable success with. It’s therefore odd that on the one hand we’d be really willing to accept something like CAPE, but on the other hand dismiss the potential of AI. The big advantage of AI is it can come up with new rules really quickly and often. If the CAPE signal stops working, or the Graham signal becomes less effective, it’ll learn whatever new rules do work. It can easily do this very thoroughly, every day, with moderate computing resources.

So, the current advantage of AI isn’t that it's an entirely new approach, and it’s not even that it’s super sophisticated, insightful or smart. It’s that even though AI basically does the same thing that we’ve been doing decades, it has the capability to do it in a dynamic rather than a static way. In other words, we don't need to wait for the Ben Graham of today to write a new book, or for Bob Shiller to publish a new paper. The AI looks for rules that work on an ongoing basis, and adjusts in real time as things start/stop working. It does the same basic analysis, just way, way faster. That's a massive advantage even if it's pretty crude.

If you look that the assets that AI likes (at least right now), many are smaller, not heavily traded, under the radar companies that are not always super profitable. That’s not surprising, it’s kind of easy to see how an algorithm got there. It used the same kinds of quantitative rules and logic that we’ve all been using for decades. The only difference is it assessed the viability of millions of potential new, simple quantitative rules while I slept last night.