60% semiconductors through SMH and 40% US energy through XLE. Target weights reset monthly and drift with returns between rebalances. The ETF holdings are the same in both universe views; only the comparison benchmark changes.
Basket total returns · Annualized using 252 trading sessions · 0% risk-free rate and downside target. As of 2026-09-11 · close.
How these statistics are calculated
Risk methodology
Sharpe: average daily simple return divided by its sample standard deviation, multiplied by √252. The risk-free rate is assumed to be 0%; this is not a cash-adjusted ratio.
Sortino: average daily simple return divided by downside deviation, multiplied by √252. Downside deviation is the square root of the average squared shortfall below 0%, counting all sessions, including zeros for gains.
Historical volatility: √(252 × average squared daily log return), matching the Vol page’s realized-volatility convention. This differs from the sample standard deviation used in Sharpe.
Each period requires complete history and at least 5 sessions. MTD and YTD follow the calendar; 1y is 252 sessions. Missing data or an undefined ratio displays — with a reason. Short periods can produce unstable ratios; intraday snapshots include a partial session. History before this basket’s creation is backtested. If a new basket has no stored daily return series yet, estimates from its rounded published chart levels are marked ≈.
S&P 500 business models on the wrong side of AI: seat-based IT services and outsourcing, ad agencies, and data/research vendors whose products LLMs increasingly substitute. A monitoring basket (potential underperformers / hedge candidates), not a buy list. Small by design — most pure displacement stories live outside the index.
The compute, networking, memory, storage, and power layer of the AI buildout. Owners of the bottlenecks: accelerators, custom silicon, switching, optics, HBM, fab equipment, test, and the server/power chain that turns capex into installed compute.
Software businesses positioned to monetize AI through existing distribution: copilots attached to installed bases, agent platforms, and the EDA/design tools that themselves get an AI productivity kicker. The bet is that distribution beats model quality.