Taking the calendar out

Retailers hire for a holiday rush and shed jobs afterwards. Schools operate to a calendar. A monthly employment report that treated each predictable swing as a new development would announce the same boom and bust every year. Seasonal adjustment asks what happened beyond the usual pattern.

America's Bureau of Labor Statistics uses retail hiring before Christmas, followed by a decline, as an example in its state and area employment guidance. The jobs are real in both months. Removing their recurring seasonal effect is a way of comparing underlying conditions, not a claim that those workers never existed.

Better than an ordinary January

Imagine an invented industry that usually loses 100,000 jobs in January after temporary holiday contracts end. This January it loses only 60,000. In a deliberately simplified additive calculation, the seasonally adjusted change would be a gain of 40,000. The actual number of jobs fell, but it fell less than the normal January pattern would suggest.

The published payroll figures require more than that subtraction. Real seasonal factors are estimated from time series, can vary between industries and can use additive or multiplicative models. The example nevertheless shows why an adjusted increase and an unadjusted decline can coexist without a contradiction. BLS publishes the methods used in its seasonal-adjustment documentation.

The distinction changes the question a statistic answers. A business deciding how many uniforms it needs cares about the actual workforce. An analyst assessing whether hiring is accelerating across the economy may prefer a comparison that removes predictable holiday and school effects. BLS's introductory explanation describes the adjustment as a tool for identifying nonseasonal movements between months.

An annual comparison with the same month a year earlier can reduce some seasonal problems, but it answers a different question. It can remain positive after a recent slowdown because earlier months were strong. Looking at both a recent adjusted trend and a same-month annual comparison helps separate momentum from the accumulated result of the past year.

The calendar has awkward details

Even the distance between two surveys can matter. The American establishment survey refers to the pay period containing the 12th of each month. Consecutive reference weeks can be four or five weeks apart. In highly seasonal industries, the extra interval can change how many hires or departures have occurred before the survey catches them. BLS explicitly models this effect in its national employment methodology.

Some adjustments are more specialised. The same methodology describes treatment of temporary poll workers in election Novembers and moving holidays that can affect reported hours. The relevant calendar includes election rules and pay periods as well as the seasons. The precise adjustment depends on the series; a rule used for employment should not casually be applied to retail sales or industrial output.

Because seasonal patterns are estimated, unusual events can make them harder to identify. A sudden closure, a changed shopping habit or an unusually timed holiday may complicate the division between recurring and exceptional movement. An adjusted figure should therefore be read as the statistical agency's best estimate under its published method, with uncertainty still attached.

Why yesterday's figure moves

Revisions have several causes. First, more survey responses arrive after an initial release. BLS revises its preliminary national payroll estimates in each of the next two months to incorporate additional information. Second, annual benchmarking aligns the survey's employment level with more comprehensive administrative counts, largely derived from unemployment-insurance records. These counts take longer to assemble.

Seasonal re-estimation is another source. BLS uses each monthly national release to update seasonal factors for the current month and the previous two months. Its annual benchmark process revises five years of seasonally adjusted series. A change to a historical adjusted estimate can therefore occur even when the underlying historical observation itself has not changed.

A revision caused by late responses says something different from one caused by a new employment benchmark or a revised seasonal pattern. All deserve explanation; none establishes manipulation merely by changing the previous number. Publishing a timely sample estimate and improving it later is a statistical trade-off.

Return to the invented January loss of 60,000 jobs. It answers how many jobs disappeared; the adjusted gain of 40,000 answers how that month compared with its expected seasonal pattern. Neither should be substituted for the other. A claim about a turning point needs a consistent series, a stated release vintage and more than one month's evidence.