Navigating the AI Bubble
A slew of recent ETF launches hints at a cyclical peak for AI and underscores the riskiness of the trade
Since the beginning of the year there have been two notable calls for a market top in US stocks. In ‘On Bubble Watch’, Howard Marks compares the current market climate to that of the 2000 tech bubble and the Nifty Fifty Bubble. And in ‘2035: An Allocator Looks Back Over the Last 10 Years’, Cliff Asness notes that U.S. Equities trade at levels that are very likely to make the next ten years disappointing. The key question for investors navigating bubble risk today is whether AI expectations will continue to inflate. I use data on thematic ETFs to contextualize how AI expectations have evolved and highlight how the risk embedded in the AI trade has failed to adequately compensate investors.
A recent empirical study by Ben-David, Franzoni, Kim, and Moussawi (2022) finds that the average thematic ETF underperforms the market on a relative basis. The most compelling explanation for this underperformance is that by the time an ETF provider identifies a new market theme and goes through the steps of putting together a product to take advantage of it, the eventual target basket of stocks has become too expensive. The ETF winds up launching at exactly the worst time, just as investor interest in the subject has peaked.
I build a smaller ETF dataset and compare the performance of 300 thematic ETFs launched since 2005 to 4 recently launched AI thematic ETFs. Figure 1 shows the average excess return path of this thematic ETF universe and the average excess return path to the subset of more-recently launched AI thematic ETFs. In both instances, excess returns are calculated against the S&P 500. The data show that, on average, thematic ETFs lose about 750bps to the S&P 500 over 18 months following launch. Notably, the average returns for the AI subset have remained within the bounds of what are typical for a thematic ETF.
Figure 1: Post-Launch Excess Returns for Thematic ETFs
Source: S&P Capital IQ, Countervail Analysis
Note: AI ETFs include "AIFD", "ALAI", "AGIX", and "BOTT"
While excess returns to this AI basket remain positive relative to the S&P 500 by about 500bps, investors should consider the riskiness of the trade. AI stocks tend to have very high market betas – when the market is up these stocks tend to be up even more, but the reverse is also true. I pulled down the 236 individual stocks owned by these 4 AI ETFs and estimated their CAPM alphas using trading data since December 2022. While there are some extreme positive outliers, the median monthly alpha delivered by these 236 securities is 0.0007, or about 1/10th of 1 percent. The data show that AI stocks on average have failed to compensate investors for the riskiness of the trade: a win for market efficiency.
Figure 2: Distribution of CAPM Alphas for AI ETF Underlying Holdings, December 2022 – Present
Source: S&P Capital IQ, Countervail Analysis
Perhaps AI stocks will continue to defy gravity, but with 4 fresh AI ETF launches since April 2024 it seems more likely that we are near a cyclical peak in enthusiasm. As Howard Marks notes in his memo, “Good investing doesn’t come from buying good things, but from buying things well.” AI stocks may be ‘good things’ and investor sentiment could continue to inflate, but even so the data suggests investors aren’t receiving any outsized compensation for the risk of the trade.