AI Crowding Switch Shows Discipline but Limited Alpha

The strategy produced a 173.96% total return from 2019-07-01 to 2026-07-01, with a 15.52% annualised return, 9.96% volatility and a 1.00 Sharpe ratio. Risk control was credible, with a -13.28% max drawdown, -1.42% CVaR 95% and 53.7% daily win rate. However, the portfolio lagged SPY by 5.95 percentage points over the test window, making the relative-return thesis mixed rather than fully supported. The skeptic case is material: if AI earnings, hyperscaler capex and price leadership remain intact, the switch risks staying underexposed to the same assets driving index returns.

• Strategy underperformed SPY by -6.0% over the backtest window (portfolio 174.0% vs benchmark 179.9%).
• Absolute performance was solid: the strategy returned 173.96% cumulatively and 15.52% annualised over the 2019-07-01 to 2026-07-01 backtest.
• Risk-adjusted results were acceptable but not exceptional, with a Sharpe ratio of 1.00 and Sortino ratio of 1.07.
• Drawdown control was a strength of the GLD/BIL safety-switch design, with maximum drawdown limited to -13.28% and CVaR 95% at -1.42%.
• Relative performance was weaker: the strategy trailed SPY’s 179.92% total return by 5.95 percentage points, creating negative benchmark alpha.
• The regime logic is plausible for crowded AI momentum, but the 63-day momentum and 21-day volatility switch may lag fast reversals or earnings-driven upside gaps.
• Post-skeptic conviction in the thesis was only 3/10, reflecting a meaningful risk that AI leadership remains fundamentally validated rather than merely crowded.

• Thesis frames NVDA risk, but SMH (VanEck Semiconductor) is ~20% NVDA — effective book exposure ~3.0%.
• Skeptic counter-argument: The strongest counter-argument is that short/underweight NVDA/SMH/QQQ is not a valuation or crowding trade so much as a short on a still-accelerating primary-data earnings cycle: if NVDA’s SEC filings and earnings releases continue to show data-center revenue growth, high gross margins, and forward supply visibility while hyperscaler 10-Qs from MSFT/GOOGL/AMZN/META keep confirming elevated AI capex, then price momentum is being validated by fundamentals rather than speculative positioning. In that regime, a 63-day momentum / 21-day vol safety switch is structurally disadvantaged: it exits only after realized vol or trend damage appears, but AI leaders can gap higher on earnings, supplier commentary, and capex revisions before the model re-enters, leaving the strategy underexposed to the exact stocks driving SPY/QQQ returns. If AI capex remains funded by cash-rich hyperscalers and NVDA’s revenue revisions keep rising, the bear thesis becomes a repeated short-vol/short-growth bet against positive estimate revisions, not a prudent de-crowding trade.. Kill trigger to watch — NVDA reports its next quarter with data-center revenue growth of at least +15% QoQ and at least +50% YoY, while guiding next-quarter total revenue at least 5% above sell-side consensus; NVDA gross margin guidance remains at or above 72% non-GAAP, indicating no material price/mix collapse despite competition or supply normalization; Combined quarterly capex from MSFT, GOOGL, AMZN and META is up at least +20% YoY and at least 3 of the 4 management teams explicitly guide AI infrastructure capex higher or sustained on earnings calls; SMH closes at a new 52-week high and outperforms SPY by at least +500 bps over 20 trading days; QQQ 63-day total return is positive while 21-day annualized realized volatility remains below 30%, keeping the proposed switch risk-on rather than defensive.
• Concentration risk remains material because the strategy clusters exposure around NVDA, SMH, QQQ, GLD and BIL, leaving single-name and sector shocks capable of dominating outcomes.
• Regime fragility is unresolved: the backtest may not fully capture a sustained low-volatility, mean-reverting environment in which momentum-based switching underperforms.
• Liquidity and implementation slippage could reduce realised returns versus the backtest, particularly if ETF spreads on smaller venues degrade fill quality.
• The skeptic’s counter-argument is significant: if NVDA data-center growth, gross margins, hyperscaler AI capex and SMH/QQQ price leadership remain strong, the underweight becomes a recurring bet against positive estimate revisions; the stated kill scenario assigns this a 38% probability over the window.