We construct portfolios of stocks and bonds that are maximallypredictable with respect to a set of ex-ante observable economicvariables, and show that these levels of predictability arestatistically significant, even after controlling for data-snoopingbiases.We disaggregate the sources of predictability by usingseveral asset groups — sector portfolios, market-capitalizationportfolios, and stock/bond/utility portfolios — and find that thesources of maximal predictability shift considerably across assetclasses and sectors as the return horizon changes.Using threeout-of-sample measures of predictability — forecast errors, Merton'smarket-timing measure, and the profitability of asset-allocationstrategies based on maximizing predictability — we show that thepredictability of the maximally predictable portfolio is genuine andeconomically significant.