Nothing changed. The dashboard disagrees.

This company has 1,000 people and a true turnover of 10% a year that never changes. Each panel measures the same replayed years. Only how often you look is different, and a shorter window is simply a smaller sample of the same truth.

Years replayed: 0
Try a fix:
Annual1 reading a year
–of readings outside the reference band
Quarterly, annualized4 readings a year, each quarter × 4
–of readings outside the reference band
Monthly, annualized12 readings a year, each month × 12
–of readings outside the reference band

Your turn: real change, or just noise?

Part 2. Above, nothing ever changed, and a yardstick built for yearly numbers flagged … of monthly readings in expectation. Here each month is judged the way an XmR chart judges it, and in half the rounds something really does change. Can you tell which?

  1. Read the chart. The grey dots are last year's 12 months. They set the average line and the red limits: average ± 2.66 × the average month-to-month change. The 24 dark dots are this period, each one month's leavers × 12 ÷ headcount. A red dot falls outside the limits.
  2. Make the call. In half the rounds the true rate secretly moves up or down by 25% to 60% in some month between 7 and 18. In the other half it never moves. In stable, evaluable rounds, the simulated share of dark dots turning red is ….
  3. See the truth. Answering reveals the true rate and what the three XmR rules would have flagged. The scoreboard then compares your calls with the rules.

    Observed counts and annualized rates

    Did the true turnover rate change during these 24 months?

    The baseline rate is given above; any future change stays hidden until you answer. Focus this answer area or a choice, then press Y or N.

    Benchmark waiting…

    Detection timing benchmark

    Among evaluable changed rounds, signals before the change are false alarms. A pre-change-only signal counts as “changed” in classification but misses detection. Delay starts at zero in the change month; mean delay includes detected rounds only.

    What you're looking at

    Every panel measures one simulated company: 1,000 people with a true turnover of 10% a year. Leavers are replaced and the rate never changes. Each replayed year is measured several ways at once, and every dot that drops is one number a dashboard would show.

    Why the readings spread out

    A turnover rate is a count of fairly rare events divided by headcount. How precise it is depends on how many leavers stand behind it, not on the calendar. Annualizing a month multiplies the number by 12, and it multiplies the wobble by 12 too.

    Coin flips work the same way. In 12 flips, anything from 3 to 9 heads is normal. In 1,000 flips, 47% to 53% heads is normal. Judge 12-flip results against the 1,000-flip range and about 3 in 4 look abnormal. A narrow range belongs to a big sample; using it on a small one manufactures alarms.

    ViewExpected leavers per readingTypical swing (SD)Readings outside the annual 2 SD bandExpected flagged readings/yearOutside own 2 SD band

    Exact binomial figures for your current settings. The swing grows with the square root of how many times a year you measure: quarterly is 2× the annual swing, monthly about 3.5×. The share outside the range can be above or below a bell-curve approximation because a monthly rate can only move in steps of one leaver (1.2 points here).

    Why it matters for people analytics

    False alarms. Judge monthly numbers against a yearly sense of normal and about 60% of months look like a problem in a company where nothing changed. Each one can start a meeting, a root-cause hunt, or an intervention.

    Fake wins. After an unusually bad month, the next one is usually closer to normal whatever you did. That is regression to the mean, and it makes the intervention look like it worked.

    Small teams, same trap. Under this equal-risk model, noise depends on expected leavers per reading, so smaller teams also produce less precise rates. A team of roughly 83 people measured once a year is approximately as noisy as the whole company measured monthly (headcount is rounded). Small-team rankings can reflect substantial chance variation; actual differences require evidence.

    Alarm fatigue. A dashboard that is red half the time teaches leaders to ignore it, and that is when a real shift slips past.

    Looking often isn't the problem. The yardstick is.

    Frequent monitoring is how you catch a real change early. What goes wrong is judging small windows against annual-sized expectations. Better options, each with a price:

    Rolling 12-month turnover. Updates every month and is as steady as the annual number, but it reacts slowly. Adjacent windows share 11 of 12 months: their correlation is 11/12 in this independent, constant-rate model. Flagged readings therefore arrive in streaks, not independent alarms.

    2 SD reference bands sized to the window. The checkbox uses ±2 SD of each view’s own noise. The table shows exact discrete binomial probabilities at your settings, which need not be 5%. A conventional p-chart uses 3 SD control limits; this demonstration is not that chart. Wider bands reduce flags and can conceal small changes.

    XmR charts (process behavior charts). The limits come from your own recent data (average ± 2.66 × the average month-to-month change), and three simple rules tell you when to look closer, which makes them usable by non-technical colleagues. The game on this page uses them, following Stehlík (2024), and Xmrit draws them for free. They still raise false alarms: with only one year of baseline, the three rules fire on about 41% of evaluable stable two-year dashboards in the game, while giving no signal at or after the change in about … of evaluable changed rounds. These are different questions from whole-dashboard classification. A zero-moving-range baseline is unavailable and excluded from XmR denominators.

    Counts with a range beat bare rates. "13 leavers this month, and 3 to 14 is normal" is much harder to overreact to than "annualized turnover hit 15.6%".

    Pool or shrink small units before ranking them, for example with partial pooling or Bayesian small-area estimates.

    What the simulation assumes

    Constant headcount, because leavers are replaced. Every person independently has the same 0.83% chance of leaving in any month. No seasonality, no clustered exits such as reorganizations, and simple annualization. Annual turnover here means expected annual exits divided by headcount. An original employee’s probability of leaving within a year is 1 − (1 − monthly chance)¹². Real processes may be more or less variable than this model.

    Interactive take on the Monte Carlo illustration credited to Lipinski (2017).