Your rolling correlation depends on the window you chose

Rolling correlation between two assets is computed over a chosen trailing window — 30, 60, 250 days — and the chosen window can flip the sign of the reported relationship: the same pair of series regularly shows positive 60-day correlation and negative 250-day correlation at the same date. During March 2020, equity-treasury correlations measured on short windows turned sharply positive while long-window estimates stayed negative for weeks longer, a divergence visible in any chart of the period. UZU NEWS publishes information, not investment advice, and treats correlation as a measurement with stated conditions.

Correlation is probably the single most quoted statistic in portfolio commentary and the one most often quoted without its window, frequency or estimator. That omission is not cosmetic. The window is part of the definition of the statistic, not a detail attached to it.

What goes into a rolling correlation?

The Pearson correlation between two return series over a window is the covariance divided by the product of the standard deviations, both computed on that window only. Rolling the window forward one day drops the oldest observation and adds the newest, producing the familiar time series. Every output point is therefore an average over the window's history, weighted equally, and dated — strictly speaking — to the whole window rather than to its endpoint.

Two conventions hide inside even this simple statistic: the return frequency (daily, weekly) and the treatment of overlapping data. Daily returns give roughly 21 observations per month; weekly gives about four. The same window length in months is a different statistical object at each frequency, with different variance and different sensitivity to single observations.

Why does the window change the answer so much?

Because correlations between financial assets are not constant, and a rolling estimate is an average over whatever regimes the window happens to span. A 250-day window that includes one crisis month and eleven calm months reports a blend of the two regimes' correlations, dominated by whichever occupies more days. A 60-day window placed entirely inside the crisis reports the crisis correlation alone. Both are correctly computed; neither is "the" correlation, because no single number is.

Short windows have the opposite failure: with 30 daily observations, the estimator's sampling noise is large enough that swings of 0.2 in reported correlation can come from ordinary variability rather than any change in the relationship. The confidence interval around a 30-day correlation is wide, a fact routinely ignored in commentary that narrates every wiggle.

How can you test window sensitivity yourself?

The procedure is mechanical and worth running on any correlation claim before repeating it.

1. Fix the pair of assets and the return frequency — daily closes are the default.
2. Compute rolling correlations at several windows: 30, 60, 120, 250 days.
3. Plot all series on one axis and note the dates where signs disagree or magnitudes diverge materially.
4. Recompute at a second frequency (weekly) to check which conclusions survive the change.
5. Record the window that produced any number you intend to publish.

A correlation claim that cannot survive step 3 — sign agreement across windows — is not a claim about the assets; it is a claim about one window's history.

What happened to diversification measures in 2020?

The 2020 stress episode is the canonical demonstration. Long-window estimates of equity-bond correlation remained negative well into the crisis, implying diversification held, while short-window estimates turned positive during the March selloff as Treasuries and stocks fell together over specific days. Portfolio constructions calibrated to the long-window number behaved differently from ones calibrated to short windows, and both were surprised in opposite directions. The episode did not prove one window right; it proved that the underlying relationship was regime-dependent in a way no single window could represent.

Where does window analysis itself mislead?

Three cautions complete the picture. First, correlation is not stability: two assets can have correlation near one in calm weeks and near zero in turbulent ones — the conditional correlation question — and rolling estimates average over the condition. Second, outliers dominate short windows: a handful of extreme days can drive a 30-day estimate almost entirely, which is why rank-based estimators such as Spearman or Kendall's tau are worth computing alongside Pearson. Third, correlation measured on returns says nothing about tail dependence — the tendency to crash together — which is the quantity diversification actually depends on in the worst weeks. Copula-based tail measures exist precisely because linear correlation underdescribes crashes.

Return series for running these checks yourself are available from freely accessible public data services, including the FRED database maintained by the Federal Reserve Bank of St. Louis, where Treasury yields and reference rates can be pulled directly for replication; the Securities and Exchange Commission's investor-education desk at investor.gov covers the underlying basics.

Correlation is a tool with exact requirements: state the assets, the frequency, the window, the estimator and the date. A quoted correlation missing any of these five is not a measurement yet — it is a rumor with decimals.