Seasonal Adjustment

Removing the calendar pattern so the underlying movement is visible.

Also known as: seasonal adjustment, deseasonalising, seasonally adjusted

Seasonal adjustment estimates a series’ repeating calendar pattern and divides (or subtracts) it out, leaving trend plus noise. The adjusted series answers “how are we doing apart from the time of year?” — the number behind every “seasonally adjusted” jobs report, retail figure and growth chart.

observed:   Dec spike every year
seasonal:   +25% each December (estimated from history)
adjusted:   Dec spike removed → trend readable month to month

It is an estimate, not a fact: the “season” is inferred from past repetition, and the adjustment revises history as new data arrives — last quarter’s adjusted growth can change next quarter. Treat adjusted numbers as modelled, and keep the raw series beside them.

The classic mistakes:

Adjusting with too little history. A season estimated from one repetition is a guess wearing a method’s clothes. Two to three full cycles minimum before trusting it.
Hiding the raw data. Adjusted series look authoritative and smooth; stakeholders forget the adjustment exists. Always show or link the unadjusted series.
Adjusting away real breaks. A level shift (new pricing, a pandemic) gets smeared across seasons by the estimator, distorting years of history. Detect breaks first; adjust around them.
Over-adjustment of noisy series. On thin, volatile data the seasonal estimate is mostly noise, and “adjusting” adds error rather than removing pattern. Check the season is stable before removing it.
Comparing adjusted to unadjusted. Mixing the two in one chart or sentence produces nonsense growth rates. Label every number.

When not to adjust: when the audience needs the raw truth (cash planning needs actual December, not adjusted December), when history is short, or when a simple period-over-period comparison already answers the question. Adjustment is for trend-reading, not for every chart.