Finding the Rhythm: Seasonality Detection in Business Data
Sales dip every Sunday. Traffic peaks on Wednesdays. If you do not model the rhythm, your forecasts and anomaly alerts will keep crying wolf — here is how to detect cycles properly.
Most business data has a heartbeat. Web traffic peaks on weekdays and collapses on weekends. Retail revenue swells every December. Support tickets spike on Monday morning. If you ignore that rhythm, two things break: forecasts systematically miss, and anomaly detection flags the same "unusual" Wednesday every single week.
Detecting seasonality is not exotic — with daily or weekly data it is a few well-defined calculations. Here is the practical sequence, the same one a client-side analysis tool like KPI Master runs when it inspects a dated column.
Step 1: Remove the Trend First
Seasonality and trend are entangled. A growing revenue line looks like it has a yearly cycle even when it does not, because every month is higher than the last. Before measuring any cycle, isolate it from the trend with a centered moving average whose window equals the candidate cycle length. For a weekly rhythm on daily data, that is a 7-point window; for monthly data with an annual rhythm, a 12-point window.
Subtract (or divide by) that smooth line and what remains is the detrended series — the part that repeats.
Step 2: Measure the Lag-k Autocorrelation
The correlation between a series and itself shifted by k periods — the lag-k autocorrelation — is the standard detector. For daily data, compute it at k = 7; for hourly data, k = 24; for weekly data, k = 52.
r_k = cov(x_t, x_(t-k)) / var(x_t)An autocorrelation above roughly 0.4–0.5 at the candidate lag, with near-zero values at other lags, is a strong signature of a real cycle. Two useful sanity checks:
- Check adjacent lags too. A weekly cycle shows up at lag 7 — but also, more weakly, at lag 14. If only lag 1 is high, you have momentum (autocorrelation), not seasonality.
- Check a second season. Many datasets have both a weekly and an annual pattern. Test the shorter cycle first, remove it, and test the longer one on the remainder.
Step 3: Use the Rhythm, Don't Just Admire It
A detected cycle changes three things in practice:
- Forecasting. Add the seasonal shape on top of your trend model (seasonal naive, seasonal decomposition, or a model with seasonal dummies). A flat forecast of a series that visibly cycles is a forecast nobody should trust.
- Anomaly detection. Compare a point against the same point in previous cycles — Monday against Mondays, January against Januaries — rather than against the overall mean. This is what turns "Wednesday spike" from an alert into a non-event.
- Baselines and targets. If you set a weekly target on a daily-series sum, make sure the expectation accounts for which days fall inside the week. Holiday weeks are not weak weeks; they are different-shaped weeks.
When the Rhythm Changes
The most valuable finding is not the cycle itself but a change in it. Recompute the lag-k autocorrelation on a rolling window: if a steady weekly rhythm weakens or shifts, something structural changed — a new audience, a new pricing page, a new campaign cadence. That is the moment to investigate, because it is exactly when your old seasonal adjustments start producing quietly wrong answers.