Would you fit a SARIMA with m=52 to three years of weekly sales, and how would you justify the call?

## What m = 52 actually asks for Setting a seasonal period of 52 on weekly data tells the model that this week's value relates to the value 52 weeks ago, and that a seasonal difference or a seasonal coefficient should encode that relationship. It sounds like a mechanical translation of "the data are weekly and the pattern is annual." It is not, and three separate problems arrive together. ### Problem one: effective sample size for a seasonal parameter A seasonal autoregressive coefficient at lag 52 is a statement about how one year relates to the next. With three years of history you have observed roughly three such transitions. The 156 rows are not 156 pieces of evidence about that coefficient; they are three, spread across 52 week-of-year positions. Any parameter estimated at that lag is fitted on approximately three effective observations, and the standard error will reflect it even when the point estimate looks plausible. The general rule to carry into the interview: **seasonal structure is informed by the number of complete cycles observed, never by the number of rows.** ### Problem two: the cost of the seasonal difference A seasonal difference with `D = 1` pairs each observation with the one 52 weeks earlier, so the first year has no partner and is lost. From 156 weeks you keep about 104. If you also take an ordinary first difference, another observation goes. Spending a third of the history to remove a pattern you have only seen three times is a poor trade, and it happens silently — nothing in the output announces that a year of data was discarded. ### Problem three: 52 is the wrong number A year is about 365.25 days, which is roughly 52.18 weeks. A fixed lag of 52 therefore drifts against the calendar by more than a day a year, and the drift is worse for anything tied to a fixed date rather than a weekday. Over a few years the peak the model expects at lag 52 has moved. Add ISO week conventions, occasional 53-week years, and moving holidays, and a rigid integer lag is being asked to represent a cycle that does not have an integer length. The seasonal block has no mechanism to accommodate that. ### The alternative that stays inside the family Dynamic harmonic regression keeps the same model type but changes how the cycle is carried. Instead of a seasonal block, you build a small set of sine and cosine pairs at the annual frequency — `sin(2*pi*k*t/52.18)` and `cos(2*pi*k*t/52.18)` for `k = 1, 2, ...` — and enter them as exogenous regressors, letting a non-seasonal ARIMA structure handle what is left over. The advantages are concrete. Each harmonic pair costs two parameters, so two or three pairs cost four to six in total instead of a seasonal block plus 52 lost observations. The period does not have to be an integer, so the alignment problem disappears. The regressors are deterministic functions of the date, so the future values needed to forecast are exactly known — none of the future-value trouble that other exogenous drivers bring. And the number of pairs is a tunable smoothness knob: one pair gives a single smooth annual wave, more pairs sharpen the peaks and troughs. Start small and add only if the shape genuinely demands it. Holiday and event effects, which a lag-52 term handles badly anyway because holidays move relative to week numbers, go in as their own indicator regressors alongside the harmonics. ### When m = 52 is defensible The answer is not never. With a long, clean history — eight or ten years of stable weekly data on a stable calendar — the effective sample size argument weakens considerably and a seasonal block becomes estimable. It is also defensible when the annual pattern is genuinely week-indexed by the business itself, for example a retail calendar whose weeks are defined so that comparable weeks align by construction. In those cases the objection about 52.18 largely evaporates, because the business calendar, not the astronomical year, is the cycle. ### How to present the judgment A principal-level answer does not stop at "no." It sets out the criterion — parameters against complete cycles observed, and data spent against pattern recovered — applies it to the case at hand, names the concrete alternative, and states what would change the decision. It also flags the reporting consequence: a model that spends a third of a short history on an under-identified seasonal block will produce prediction intervals that look confident while resting on three observations, and that is the failure mode most likely to reach a stakeholder unchallenged.