{
  "id": "seasonal-adjustment",
  "title": "Seasonal Adjustment",
  "lang": "en",
  "source_lang": "en",
  "summary": "Seasonal adjustment estimates and removes recurring within-year patterns from time series to make underlying changes easier to interpret.",
  "category": "mathematics",
  "subcategory": "probability-statistics",
  "type": "concept",
  "aliases": [
    "Deseasonalization",
    "Seasonal correction"
  ],
  "generated_at": "2026-11-05",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/seasonal-adjustment/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/seasonal-adjustment/",
      "json": "https://aiwiki.page/zh/seasonal-adjustment.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Seasonal adjustment is a statistical procedure that estimates and removes recurring within-year patterns from a time series. These patterns may reflect weather, holidays, production schedules, school calendars, or other activities associated with particular times of year. By separating such movements from other changes, seasonal adjustment makes consecutive months or quarters more comparable. It is widely used in economic statistics, including employment, retail sales, industrial production, and gross domestic product. A seasonally adjusted series still contains irregular fluctuations and is not the same as an estimate of its underlying trend."
    },
    {
      "heading": "Purpose and scope",
      "text": "A series exhibits seasonality when its behavior varies systematically with the time of year. Retail sales may rise before major holidays, construction activity may weaken during winter, and tourist arrivals may peak during vacation periods. These patterns can obscure changes in economic conditions: an increase from November to December, for example, may be smaller than the increase normally expected at that time of year.\n\nSeasonal adjustment asks how an observation compares with the level expected after allowing for its usual seasonal influences. Consequently, a seasonally adjusted series may decline even when the unadjusted series rises, if the observed increase is smaller than the estimated seasonal increase.\n\nSeveral distinctions are important:\n\n- Seasonal adjustment is not trend estimation. Removing a seasonal component does not remove short-term shocks or measurement noise.\n- Seasonal adjustment is not inflation adjustment. Correcting prices or monetary values for inflation addresses a different source of change.\n- Seasonal adjustment is not annualization. Expressing a quarterly flow at an annual rate changes its reporting scale rather than removing seasonal effects.\n- Year-over-year comparison is not a complete substitute. Comparing the same month in successive years reduces some seasonal influences, but can still be affected by moving holidays, changing seasonal patterns, and unusual conditions in the comparison period.\n\nThe procedure ordinarily targets identifiable recurring patterns, not every effect associated with a season. Exceptionally cold weather, for example, can remain an irregular influence rather than being removed as normal winter seasonality."
    },
    {
      "heading": "Decomposition models",
      "text": "Seasonal adjustment usually rests on a decomposition of observed values into components. These components are not separately observed; they are estimated under assumptions about their behavior.\n\nAdditive decomposition\n\nAn additive model is\n\n\\[Y_t=T_t+S_t+I_t,\\]\n\nwhere \\(Y_t\\) is the observed series, \\(T_t\\) is the trend-cycle component, \\(S_t\\) is the seasonal component, and \\(I_t\\) is the irregular component. The trend-cycle combines relatively persistent movements, including longer-term growth and business-cycle fluctuations.\n\nThe seasonally adjusted series is\n\n\\[A_t=Y_t-\\widehat S_t.\\]\n\nAdditive decomposition is suitable when the size of seasonal fluctuations is approximately independent of the level of the series. For a fixed seasonal pattern with \\(m\\) observations per year, a common normalization is\n\n\\[\\sum_{j=1}^{m}S_j=0.\\]\n\nThis normalization helps distinguish the seasonal pattern from the series level.\n\nMultiplicative decomposition\n\nA multiplicative model is\n\n\\[Y_t=T_tS_tI_t,\\]\n\nwith adjustment\n\n\\[A_t=\\frac{Y_t}{\\widehat S_t}.\\]\n\nIt is useful when seasonal fluctuations tend to grow proportionally with the level of the series. A seasonal factor of \\(1.20\\), for example, indicates a level about 20 percent above its nonseasonal counterpart.\n\nFor strictly positive observations, the natural logarithm converts a multiplicative decomposition into an additive one:\n\n\\[\\log Y_t=\\log T_t+\\log S_t+\\log I_t.\\]\n\nLogarithmic models require special handling when observations are zero or negative. Transforming estimates back from logarithms also requires care: exponentiating a conditional mean on the logarithmic scale does not generally produce the conditional mean on the original scale.\n\nIdentification\n\nDecomposition is not unique without restrictions. A slowly changing pattern could be allocated partly to the trend and partly to the seasonal component, depending on the method. Seasonal-adjustment procedures therefore impose constraints through smoothing filters, statistical models, normalization rules, or combinations of these.\n\nModern methods commonly permit seasonality to evolve gradually. A rigid set of twelve monthly factors can be inadequate when shopping habits, production schedules, or other seasonal influences change."
    },
    {
      "heading": "Calendar effects and preliminary adjustment",
      "text": "Calendar effects are often treated alongside seasonality, although they are conceptually distinct.\n\nTrading-day effects arise because different weekdays can have different levels of activity, while their numbers vary from month to month. A month containing five Saturdays may have a different retail-sales total from one containing four.\n\nMoving-holiday effects arise when holidays shift between months or affect different portions of reporting periods. Easter is a common example. Leap-year effects and variations in month length can also matter, depending on whether a series measures totals, averages, or values at a particular date.\n\nA common preliminary model is\n\n\\[g(Y_t)=\\mathbf{x}_t^\\top\\boldsymbol{\\beta}+Z_t,\\]\n\nwhere \\(g\\) is a transformation, \\(\\mathbf{x}_t\\) contains calendar and intervention variables, and \\(Z_t\\) is the series subsequently decomposed. This combines regression with a model for serially dependent errors.\n\nPublished terminology varies: a series described as seasonally adjusted may also be calendar-adjusted, but this should not be assumed without checking its definition.\n\nOutliers and abrupt level changes may be represented explicitly so that they do not distort estimated seasonal factors. Such preliminary treatment does not necessarily mean that the event disappears from the final adjusted series. A nonseasonal shock can be removed temporarily for estimation and then restored."
    },
    {
      "heading": "Main methods",
      "text": "Moving-average methods\n\nA moving average smooths a series using neighboring observations. Classical decomposition estimates a trend, removes it, estimates seasonal patterns from the remaining values, and then calculates an irregular component.\n\nThe X-11 family develops this approach through iterative filtering, seasonal-factor estimation, and treatment of extreme observations. Its historical successors added modeling and diagnostic capabilities.\n\nThe U.S. Census Bureau’s X-13ARIMA-SEATS software combines regression and ARIMA modeling with a choice of X-11 or SEATS decomposition. ARIMA forecasts and backcasts can extend the observed series, reducing some difficulties associated with applying centered filters near its endpoints.\n\nModel-based signal extraction\n\nModel-based methods specify a stochastic model for the observed series and derive component estimates from that model. SEATS, whose name refers to signal extraction in ARIMA time series, decomposes a fitted ARIMA model into components such as trend-cycle, seasonality, and irregular variation.\n\nSuch methods make assumptions about component dynamics explicit. Their estimates nevertheless depend on model adequacy and on the restrictions used to distinguish components.\n\nA related approach uses structural time-series models, in which trend and seasonal states evolve according to specified equations. These models can be estimated using a Kalman filter and associated smoothing procedures.\n\nLocally fitted decomposition\n\nSTL decomposition, or seasonal-trend decomposition using LOESS, estimates components through local regression and iterative smoothing. It allows seasonal patterns to evolve, and robust variants reduce the influence of extreme observations.\n\nBasic STL does not automatically account for trading-day or moving-holiday effects. These may require separate modeling. Extensions address multiple seasonal periods, as can occur in hourly data with daily and weekly patterns.\n\nRegression and harmonic methods\n\nSeasonal patterns can also be represented using month or quarter indicators, or sine and cosine terms based on Fourier series:\n\n\\[S_t=\\sum_{k=1}^{K} \\left[ a_k\\cos\\left(\\frac{2\\pi kt}{m}\\right) +b_k\\sin\\left(\\frac{2\\pi kt}{m}\\right) \\right].\\]\n\nHere \\(m\\) is the seasonal period and \\(K\\) controls the complexity of the pattern. Fixed coefficients imply fixed seasonality; allowing coefficients to change over time produces a more flexible model. Serial dependence in the remaining errors may require additional modeling."
    },
    {
      "heading": "Revisions and endpoint uncertainty",
      "text": "Seasonally adjusted values are estimates and can change when new observations become available. A centered filter uses observations on both sides of a date, but future observations do not yet exist at the end of the available series. Endpoint estimates must therefore rely on asymmetric filters, forecasts, or other approximations.\n\nAs additional data arrive, they can alter:\n\n- estimates for recent seasonal factors;\n- the estimated trend-cycle;\n- model parameters and calendar effects;\n- the classification of unusual observations;\n- previously published adjusted values.\n\nRevisions can also result from changes to the unadjusted source data or from revised methods.\n\nA concurrent adjustment re-estimates seasonal adjustment as new observations arrive. Other publication policies hold some specifications or factors fixed between scheduled reviews. These approaches involve a trade-off between incorporating current information and limiting revisions."
    },
    {
      "heading": "Diagnostics and quality assessment",
      "text": "Assessment concerns both the remaining adjusted series and the plausibility and stability of its estimated components.\n\nImportant diagnostics include:\n\n- Residual seasonality: whether recurring within-year patterns remain after adjustment.\n- Residual calendar effects: whether weekday composition or moving holidays still explain systematic variation.\n- Model adequacy: whether residual dependence and extreme observations are adequately represented.\n- Revision stability: how much historical estimates change as the observation span expands.\n- Sensitivity: how results depend on transformations, filters, outlier treatment, or the selected estimation period.\n\nSpectral analysis can reveal remaining peaks at seasonal frequencies, while statistical tests can examine particular forms of residual seasonality. No single diagnostic establishes that an adjustment is correct: removing every detectable seasonal-frequency movement can also remove genuine nonseasonal variation.\n\nUncertainty involves more than sampling error. Estimated parameters, component extraction, model selection, and revisions all contribute. A conventional confidence interval calculated for the original observations does not automatically describe uncertainty in the adjusted series."
    },
    {
      "heading": "Aggregation and interpretation",
      "text": "When a total contains several component series, two approaches are possible:\n\n- Direct adjustment: adjust the total itself.\n- Indirect adjustment: adjust the components separately and then aggregate them.\n\nThe results generally differ because estimated models, seasonal patterns, and adjustment operations need not combine linearly. Direct adjustment may represent seasonality in the total more effectively; indirect adjustment preserves accounting consistency between adjusted components and their sum. Statistical agencies may use reconciliation procedures where both objectives matter.\n\nLikewise, seasonal adjustment does not generally commute with division. Adjusting a ratio directly can yield a different result from dividing its separately adjusted numerator and denominator. This distinction matters for productivity measures, average earnings, and some labor-market indicators.\n\nAnnual totals are not automatically preserved by every adjustment method. Where preservation is required, additional constraints or benchmarking procedures may be applied."
    },
    {
      "heading": "Limitations and applications",
      "text": "Seasonal adjustment is central to short-term analysis in econometrics, official statistics, business planning, and demand monitoring. It supports month-to-month and quarter-to-quarter comparisons, particularly when seasonal movements are large relative to the changes of interest.\n\nIts reliability is limited by the available historical record and by the stability of recurring patterns. Short series provide few repetitions from which to distinguish seasonality from chance. Structural breaks, changing reporting practices, and exceptional disruptions can make past seasonal relationships less informative. High-frequency data introduce further complications, including overlapping seasonal cycles and calendar periods of unequal length.\n\nAdjustment also does not establish causal explanations. An adjusted increase indicates a movement beyond the estimated seasonal and calendar pattern, not the reason for that movement. Interpretation therefore depends on the underlying observations, the adjustment specification, the revision history, and other relevant evidence."
    }
  ],
  "links": [
    "time-series",
    "gross-domestic-product",
    "inflation",
    "natural-logarithm",
    "linear-regression",
    "arima-model",
    "kalman-filter",
    "fourier-series",
    "statistical-hypothesis-testing",
    "confidence-interval",
    "labor-market",
    "econometrics",
    "causal-inference"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}