Rolling correlation

Also seen as: moving-window correlation, rolling-window correlation, time-varying correlation

How is a rolling correlation calculated?

A rolling correlation applies the ordinary correlation calculation repeatedly, each time to the most recent window of observations only.

where: $r^A$ and $r^B$ are the two series' per-period returns, $w$ is the window length in periods, and $\rho_t$ is the correlation of the two series across the window ending at date $t$. Each step forward drops the oldest observation and adds the newest, so the window "rolls" and the output is a series in the range −1 to +1.

Two properties follow from the shape of that calculation. The series cannot begin until the window is full, so the first $w$ periods of history produce no value at all. And the window length is a trade-off: a short window reacts quickly to a change in the relationship but jumps around on noise, while a long window is smooth but slow, still reporting a crisis-era correlation months after the crisis ended. Windows from a few months to a few years are standard practice.

Fincanva's window is one year. Wherever Fincanva computes a rolling correlation it rolls a one-year window, so every value you read is the correlation of the twelve months ending on that date — not the whole history, and not the last quarter. The window is part of the reading, not a technical footnote: a correlation of 0.8 on a one-year window and a correlation of 0.8 on a one-month window are different statements about the same pair. It also fixes where the series can start: the first year of any pair's shared history produces no value, because until then there is no full window to compute.

Why does correlation spike in a crash?

Correlation between risk assets rises sharply in a sell-off because in that regime a single force — investors reducing exposure across the board — dominates the individual reasons the assets would otherwise move apart. Positions are sold because they can be sold, not because of what they are, so shares, credit, and often assets held precisely as diversifiers all fall in the same weeks.

The practical consequence is uncomfortable: diversification measured on calm data overstates the protection it provides in exactly the periods it is meant to protect against. A pair of holdings that spent a decade at 0.2 and combined into a genuinely smoother portfolio can behave like one holding for the three months that produce most of the max drawdown. A rolling correlation is how that becomes visible; a single full-period figure hides it by construction.

Worked example: two sleeves that decouple, then do not

Take two sleeves of a strategy — a broad equity sleeve and a corporate-credit sleeve — and compute their correlation on a rolling window across a ten-year backtest.

The full-period 0.42 is not wrong — it is the correct average — but it is the answer to a question nobody asked. The diversification benefit the 0.25–0.40 stretch implies is real for nine years and largely absent for the quarter that produced the worst of the drawdown, and only the rolling series shows that the two facts are about the same pair of sleeves.

What counts as a good value?

There is no good value, only a reading. A rolling correlation near +1 means the two series have been moving almost identically over that window, near 0 means recent moves have been largely unrelated, and near −1 means they have been moving in opposite directions. What a rolling series adds over a single figure is stability: a pair that holds a steady 0.3 through several market regimes has behaved differently from a pair averaging 0.3 by alternating between 0.0 and 0.9, even though both report the same full-period number. Correlation also says nothing about magnitude — two series can be perfectly correlated while one moves five times as far, which is what beta measures instead. Fincanva does not tell you which correlations to seek or avoid in a portfolio — see Is this financial advice?.

Does Fincanva show a rolling correlation?

Fincanva does not currently display a rolling-correlation chart, or any correlation view, on a screen in the app. Analysis today covers metrics, the monthly-returns matrix, capital, allocations, positions, Strategy analytics, and start-date sensitivity; correlation is not among them. The term is documented here because it is standard vocabulary for reading a diversified strategy, and because factors are the series such a comparison is usually made against.

Backtests show what would have happened — not what will. Fincanva provides no financial advice — see Is this financial advice?.