THE ALLOCATION

Every statistic on this page is a description of one window of history. None of it is a forecast. Change the window in the bar above and watch how much the answers move; that movement is the honest part.

• The asset classes are ETF proxies, not indices. SPY, GLD, IEF, GSG, EFA, VNQ and BIL carry fees, tracking error and, for GSG, the roll cost of futures. Gold here is a fund that holds bars, not the spot fix.
• Adjusted closes assume every distribution was reinvested on the day it was paid, with no tax. A taxable holder of IEF or VNQ did worse than the line drawn here.
• Bitcoin trades seven days a week and is sampled here on exchange days only, so a weekend move lands on Monday. Daily correlations understate its co-movement with stocks; monthly ones are the ones to trust.
• Bitcoin's early history dominates any window that includes it. The default window starts in 2014-01, the first full year after the 2013 top, and the chip above says so when a window starts at a cycle top. Over this window bitcoin's deepest drawdown was ….
• Optimisers maximise estimation error. A max-Sharpe portfolio puts the most weight where the sample mean was luckiest, which is why the Frontier tab shows a resampled cloud and the Optimised tab an out-of-sample column. A frontier is a statement about the past sample, not a plan.
• Correlations rise in crises. The average correlation in the matrix is not the correlation that matters on the worst day; read the conditional-correlation chart and the episode table before the matrix.
• Past Sharpe is not a forecast. Bitcoin's Sharpe over this window is …; the next asset's is …. Both numbers would have looked different one year earlier or later.
• Fees, taxes, custody, slippage and the bid-ask spread are excluded everywhere. Rebalancing a sleeve monthly costs more than the page shows.
• "60/40" means SPY and IEF, 7 to 10 year Treasuries, unless AGG is switched on in the universe, in which case the 40 is the aggregate bond market.
• Survivorship: these are the asset classes that exist and have an ETF today. An investor in 2011 could not have bought a bitcoin ETF, and several things that looked like asset classes then no longer do.
• The risk-free rate is BIL's own return, so Sharpe ratios are versus cash as actually earned, not versus zero. Set it to zero in the bar and every Sharpe rises.
• Monthly windows are short. … monthly observations is enough for a Ledoit-Wolf covariance of eight assets and not enough to estimate a mean return with any precision, which is the reason the frontier has a conservative preset.

Methods are the textbook ones: Markowitz mean-variance with long-only bounds, Sharpe ratios on excess returns, Ledoit-Wolf shrinkage, Black-Litterman with one view and Idzorek confidence, Michaud resampled frontiers. Every formula and default is written up for the integrator's methods note.