Two volatile, weakly correlated assets, rebalanced to fixed weights, compound faster than either alone: trim what popped, add to what dropped, and the geometric mean of the mix exceeds the weighted geometric means of the parts. The gain is roughly half the variance you diversify away — it comes from the swings, not the drift. Shannon showed the toy version (cash + a coin-flip stock, 50/50, rebalanced daily); Thomas Cover’s 1991 “Universal Portfolios” paper carries the famous specimen: Iroquois Brands and Kin Ark over 22 years, ~8.9x and ~4.1x on their own, ~73x as the best constant-rebalanced pair (memory figures). Cover’s actual result is the universal portfolio — a weighting scheme that, with no forecast, asymptotically tracks whichever constant rebalance turns out best in hindsight; the same math as optimal data compression.
What the viral retelling (velesxbt, 09/2026) gets wrong, in order of damage: the 73x is the hindsight-optimal mix, not a track record — nobody turned $100k into $7M; “two stocks that went nowhere” rose 4x and 8x; and “Wall Street buried his math” is fiction, universal portfolios are a standard topic in online-learning theory. The real limits are quieter: the bonus needs assets that oscillate rather than trend (a persistent loser keeps getting bought — see Now show Japan), it is eaten by transaction costs and taxes at daily frequency, and at realistic vols it is a fraction of a percent a year — the “rebalancing bonus” a 60/40 already collects, not a strategy. The right takeaway is the log-wealth criterion: compounding runs on the geometric mean, so volatility is a cost the arithmetic mean hides — the same logic as the kelly criterion.