The AI bubble signal is currently concentrated in semiconductors.SOXX signals. Cloud does not. Software does not. Cybersecurity does not. Robotics does not. That is what makes the framework interesting. It does not force us to classify the entire AI trade as either a bubble or not a bubble. It lets us ask where the bubble is.
What does the model call a bubble?
The starting point is simple. Part of an industry’s return can normally be explained by movements in the overall stock market. The remaining portion — the return that cannot be explained by normal market exposure — is commonly called alpha. The bubble methodology asks whether that abnormal return has undergone a persistent upward shift. In practical terms, the model looks for three things.1. A significant upward break
The industry’s returns suddenly become much stronger than its normal relationship with the market would suggest.2. Persistent abnormal performance
The change must persist long enough to look like a genuine change in regime rather than a few unusually strong months.3. No evidence that the bubble has already broken
A sufficiently severe negative abnormal-return shock cancels the signal. A useful shorthand is:Bubble = structural upward break + persistent abnormal return + no recent crashThe baseline methodology uses ten years of monthly observations and searches for a structural break that occurred roughly one to five years earlier. That makes this different from ordinary momentum. The model is not simply asking what has gone up recently. It is asking:
Has the return-generating process itself changed?
Why apply this to AI?
AI is not one industry. The investment chain includes several distinct economic layers:- semiconductors;
- cloud computing;
- software;
- cybersecurity;
- robotics; and
- broader AI companies.
An ETF becomes eligible only after it has accumulated 120 monthly total-return observations.
That ten-year requirement is important because the underlying statistical test needs enough history to distinguish an unusual regime from normal variation.
At present, AIQ has not yet accumulated enough history to participate.
What does the model say today?
Using factor information through June 2026, the current classification is:
The signal is therefore not “AI is in a bubble.” It is:
Semiconductors currently exhibit the statistical characteristics of an active bubble regime.That distinction matters. Cloud does not currently qualify. Software does not. Cybersecurity does not. Robotics does not. The strongest abnormal acceleration appears to be concentrated in the infrastructure layer supplying the AI buildout. If the AI investment cycle eventually broadens into software, cloud, robotics, or another part of the stack, the monitor can identify that transition using the same rules.
What happens when an ETF signals?
The allocation rule is deliberately simple. If only one eligible ETF is classified as being in a bubble, it receives the full bubble allocation. So today:SOXX signals → 100% SOXXIf two ETFs signal, each receives 50%. If three signal, each receives one-third. And so on. The model is recalculated using monthly data. This is not a high-frequency strategy and it does not react to every daily move in AI stocks.
Does riding bubbles actually work?
Before applying the idea to AI, we ran the methodology across a broad universe of Fama–French industry portfolios from July 1936 through June 2026. Using the conventional academic treatment of factor data, the CAPM version produced:
That is broadly consistent with the intuition behind the original research.
But there is an important implementation issue.
Historical academic datasets are clean in retrospect. A real investor can only trade on data once that information has actually become available.
So we also ran a deliberately conservative version that delays the factor information before allowing it to affect the strategy.
The result weakens:
That difference is important.
It suggests that some of the apparent historical advantage is sensitive to assumptions about when factor data could actually have been used.
For us, the lesson is straightforward:
The interesting question is not only whether the strategy works in historical data. It is whether it survives when run prospectively using information actually available at the time.That is why we built the live monitor.
How did the AI version perform historically?
We also applied the same framework to the fixed AI universe. When one or more AI segments signal, the portfolio moves into those ETFs. When none signal, the production version simply holds SPY. The historical result is modest:
Using contemporaneous factor data, the AI rotation adds only about 0.3 percentage points of annual return and still has a lower Sharpe ratio than SPY.
With the more conservative publication timing, it underperforms SPY by roughly 1.9 percentage points annually.
So we do not view the historical AI results as evidence of established trading alpha.
That is not the purpose of the live strategy. The interesting part is what happens from here.
Bubble signals are rare
Historically, the AI strategy spends very little time in an active bubble allocation. Under the conservative CAPM specification, an AI bubble is detected during only about 1.1% of monthly observations. In other words, this is not really an “always invested in AI” portfolio. It is better understood as:A broad-market portfolio that occasionally rotates into an AI segment when that segment enters an unusually strong statistical regime.That makes the benchmark straightforward. The strategy does not merely need to make money. It needs to outperform the market exposure it replaces.
The live experiment
We are now running the strategy prospectively through a non-trading Scalar Field monitoring agent. Follow the live AI Bubble Rotation monitor → The official prospective inception date is:September 1, 2026No earlier returns count toward the live record. The agent runs every weekday at 9:35 a.m. ET, but the underlying strategy remains monthly. Each weekday, the agent checks whether a new eligible monthly factor dataset has become available. If nothing new has been published, nothing changes. When new data become available, the signal is recalculated once.
Two shadow portfolios
The monitor maintains two independent $10,000 shadow portfolios. They answer two slightly different questions.Research Book
When one or more AI ETFs signal, equal-weight the signaling ETFs. When no ETF signals, earn the published risk-free return. This book asks:Does the bubble signal itself identify periods of unusually strong subsequent returns?
Production Book
When one or more AI ETFs signal, equal-weight the signaling ETFs. When none signal, hold SPY. This version includes a 10 basis-point one-way turnover assumption. A separate SPY portfolio is tracked alongside it. This book asks the practical question:Does temporarily replacing SPY with an AI industry in a detected bubble improve the investor’s outcome after costs?That is the harder — and ultimately more useful — test.
The current portfolio
At the time of writing, the primary CAPM model identifies one eligible bubble:SOXXThe Fama–French three-factor robustness check independently reaches the same conclusion. The current preview allocation is therefore:
100% SOXXBut the prospective performance record has not started yet. Both shadow portfolios begin at $10,000 on September 1, 2026. From that point forward, readers can see the experiment unfold directly: Follow AI Bubble Rotation on Scalar Field →
What are we actually testing?
The experiment has two possible forms of success.Does the bubble signal contain information?
If industries identified as bubbles subsequently continue generating unusually strong returns, that would support the central intuition behind Riding Bubbles.Can that information improve a portfolio?
The harder test is whether the Production Book can outperform SPY after costs. Those are not the same thing. A signal can be statistically interesting without being strong enough to create an economically better portfolio. And the AI version gives us an additional question:Does the bubble migrate through the AI stack?Today it is semiconductors. Later it could appear in cloud. Software. Cybersecurity. Robotics. Or nowhere at all. The advantage of the framework is that we do not need to decide the answer narratively. We can let the same statistical process evaluate each segment over time.
Where we go from here
There is no claim here that we have discovered a historically proven AI trading edge. The historical evidence is not strong enough for that. What we do have is a clear hypothesis:Some parts of the AI investment stack may periodically enter bubble regimes in which abnormal returns continue long enough to be worth riding.Today, the model says that regime exists in semiconductors. Now we get to observe what happens next. The universe is fixed. The signal rules are fixed. The portfolio construction is fixed. And the prospective record begins on September 1. Track the live experiment →
Further reading
- Original strategy: Quantpedia — Riding Industry Bubbles
- Live AI experiment: Scalar Field — AI Bubble Rotation: Riding Industry Bubbles