dplyr cummean() in R: Running Mean Across a Vector
The cummean() function in dplyr returns the cumulative (running) mean of a numeric vector. It is the "mean so far" companion of cumsum().
cummean(1:5) # 1, 1.5, 2, 2.5, 3 cumsum(x) / seq_along(x) # base R equivalent df |> arrange(date) |> mutate(running_avg = cummean(value)) df |> group_by(g) |> mutate(running_avg = cummean(value)) cummean(c(10, 20, NA, 30)) # NA propagates: 10, 15, NA, NA zoo::rollmean(x, k = 3) # fixed-width rolling mean (different)
Need explanation? Read on for examples and pitfalls.
What cummean() does in one sentence
cummean(x) returns a numeric vector where position i is the mean of x[1:i]. It accumulates the average from the start through every position.
Useful for "performance to date", "session average", and other "growing window" metrics.
Syntax
cummean(x). x is a numeric vector. Returns a numeric vector of the same length.
cummean(x) is cumsum(x) / seq_along(x): same result, more readable. Use cummean for the intent; the cumsum version is the manual fallback.Five common patterns
1. Running average
Position 1 is just 10. Position 2 is mean(10, 20) = 15. Position 3 is mean(10, 20, 30) = 20. And so on.
2. Per-game running average (sports analytics)
For pre-divided values, cummean(hits / at_bats) would be wrong (mean of ratios, not ratio of cumulative sums).
3. Running mean inside a pipeline
4. Per-group running average
Group_by makes cummean restart for each team.
5. Lagged running mean (avoid current row)
lag() shifts the cummean down by one, so each row sees the average of all PRECEDING rows (excluding self). Useful for forward-looking analysis without leakage.
cummean(x) and cumsum(x) / seq_along(x) always produce the same result. The cumsum version is what cummean does internally. Use cummean for clarity; reach for cumsum/seq_along if you need to handle NAs explicitly or weight values.cummean() vs cumsum() vs zoo::rollmean()
Three "running average" computations in R, with different windows.
| Function | Window | Best for |
|---|---|---|
cummean(x) |
Growing (1..i) | All-time average to date |
cumsum(x) / seq_along(x) |
Growing (manual) | Same as cummean; explicit form |
zoo::rollmean(x, k) |
Fixed width k | "Last 7 days average" |
slider::slide_dbl(x, mean, .before = k) |
Configurable | Modern rolling-window |
When to use which:
cummeanfor growing-window means.zoo::rollmeanorsliderfor fixed-width rolling means (last N observations).
A practical workflow
The "growing-window average" pattern is the canonical cummean use case.
For each user in chronological order, the lifetime average through each row. Common in cohort and customer-LTV analysis.
For "average of last N", use slider::slide_mean(x, .before = N - 1). The two solve different problems: cummean grows; slide_mean is a fixed window.
Common pitfalls
Pitfall 1: NA propagation. cummean(c(10, 20, NA, 30)) returns c(10, 15, NA, NA). Once NA appears, every later position is NA. Filter NAs first or use cumsum(x[!is.na(x)]) / seq_along(...).
Pitfall 2: order dependence. cummean reads the vector left-to-right. Always arrange() first if the order is meaningful.
cummean(x) is NOT a rolling-window mean. It is "mean of EVERYTHING from start to here". For "mean of last 7 days", use zoo::rollmean() or slider::slide_mean(). The two are very different and easy to confuse.Cumulative vs rolling: a common confusion
Cumulative means a growing window: position i averages everything from start through i. Rolling means a fixed-width window: position i averages the previous k values (or the centered window). The two solve different problems and produce very different results. cummean(1:100) climbs slowly toward 50; rollmean(1:100, k=7) increases by 1 each step. For "season-to-date stats", cummean is right. For "last 7 days moving average", you need slider or zoo. The dplyr cumulative family (cummean, cumsum, cumprod, cummax, cummin) handles only the growing-window case; reach for slider::slide_* or zoo::roll* whenever the window has a fixed width.
Try it yourself
Try it: Compute a running monthly average of revenue, sorted by month. Save to ex_running.
Click to reveal solution
Explanation: Sort by month, then cummean computes the running average. Each row shows the average through that month.
Related dplyr functions
After mastering cummean, look at:
cumsum(),cumprod(),cummax(),cummin(): other cumulativescumall()/cumany(): cumulative logicalslag()/lead(): shift to compare across rowsslider::slide_mean(): rolling-window mean (fixed width)zoo::rollmean(): classic rolling meanRcppRoll::roll_mean(): fast rolling mean for big data
For fixed-width rolling means, the slider and RcppRoll packages are the modern tools.
FAQ
What is the difference between cummean and cumsum in R?
cumsum(x) returns the running SUM. cummean(x) returns the running MEAN, which is cumsum(x) / seq_along(x). cummean is the per-position average up to that point.
What is the difference between cummean and rollmean?
cummean is a growing window: each position averages all values from start to here. rollmean (zoo) is a fixed window: each position averages the previous k values. Different semantics, different use cases.
How do I do a per-group running mean?
df |> group_by(g) |> mutate(running = cummean(value)). group_by makes cummean restart for each group.
Does cummean handle NA?
NAs propagate: once NA appears, every later position is NA. Filter NAs before cummean, or use cumsum(x[!is.na(x)]) / cumsum(!is.na(x)) for an NA-skipping version.
Is cummean a rolling-window mean?
No. cummean is a growing window (1..i). For fixed-width rolling means (last 7 days, last 30 minutes), use slider::slide_mean() or zoo::rollmean().