As I noted in a previous post, measurement of correlation varies wildly (even in sign) when measured at different time scales (day, week, month). Everyone knows it varies over market cycles, with correlations often increasing toward 1 in times of market volatility.
Over longer time periods, though, there seem to be stronger patterns. William J. Coaker II publishes on this topic. Here’s a good one. He looks not at average correlations over time, but ranges of correlations over time. When examined in this way, the choices for low correlated assets become more clear, and the case for diversifying using assets identified this way is even stronger than using average correlations. He also notes that behavior of assets can be categorized in unexpected ways when analyzed in this way. That is, growth stocks as a class behave essentially equivalently to blend, but value is different from them both.
Liu Xinyi and Hua Fan have also looked at how correlations change over time and in relation to one another, specifically for US stock and treasury asset classes. Larry Swedroe has summarized their findings, which focus on predictive uses for correlations.