[{"data":1,"prerenderedAt":866},["ShallowReactive",2],{"content-query-9gOpOfbVsc":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"heading":8,"prompt":10,"tags":14,"files":16,"nav":6,"presets":17,"gallery":34,"body":36,"_type":859,"_id":860,"_source":861,"_file":862,"_stem":863,"_extension":864,"sitemap":865},"/tools/exponential-smoothing","tools",false,"","Exponential Smoothing Forecast Calculator","Forecast time series online from Excel or CSV data with exponential smoothing. Model trend, seasonality, and prediction intervals with AI.",{"prefix":11,"label":12,"placeholder":13},"Apply exponential smoothing","Describe the exponential smoothing analysis you want to run","e.g. Holt-Winters smoothing on monthly sales, forecast 12 months, tune alpha beta gamma, plot prediction intervals",[15],"statistics",true,[18,24,29],{"label":19,"prompt":20,"dataset_url":21,"dataset_title":22,"dataset_citation":23},"Temperature Holt trend forecast","fit Holt double exponential smoothing (trend, no seasonality) to annual global temperature anomaly; optimize alpha and beta; plot smoothed series and 10-year forecast with 80% and 95% prediction intervals; report the fitted trend slope and one-step-ahead RMSE","https://ourworldindata.org/grapher/temperature-anomaly.csv","Global average temperature anomaly","Our World in Data",{"label":25,"prompt":26,"dataset_url":27,"dataset_title":28,"dataset_citation":23},"CO₂ Holt-Winters forecast","fit Holt-Winters additive exponential smoothing to annual CO2 concentration; optimize alpha, beta, gamma; report fitted parameters; plot smoothed fit and 15-year forecast with 95% confidence band; annotate when CO2 is projected to reach 450 ppm under this trend","https://ourworldindata.org/grapher/co2-concentration-long-term.csv","CO₂ concentration in the atmosphere",{"label":30,"prompt":31,"dataset_url":32,"dataset_title":33,"dataset_citation":23},"Energy use SES comparison","compare simple exponential smoothing (SES) with alpha values 0.1, 0.3, 0.7 on per-capita energy use; overlay all three on raw data; report RMSE for each alpha; which alpha best tracks sudden shifts? also fit Holt (with trend) and compare RMSE","https://ourworldindata.org/grapher/per-capita-energy-use.csv","Per capita energy use",[35],"/img/tools/exponential-smoothing.png",{"type":37,"children":38,"toc":848},"root",[39,48,88,133,145,151,223,229,344,349,355,505,511,634,640,707,713,750,756,785,814,824],{"type":40,"tag":41,"props":42,"children":44},"element","h2",{"id":43},"what-is-exponential-smoothing",[45],{"type":46,"value":47},"text","What Is Exponential Smoothing?",{"type":40,"tag":49,"props":50,"children":51},"p",{},[52,58,60,65,67,72,74,79,81,86],{"type":40,"tag":53,"props":54,"children":55},"strong",{},[56],{"type":46,"value":57},"Exponential smoothing",{"type":46,"value":59}," is a family of time series smoothing and forecasting methods that compute a weighted average of all past observations, with weights decreasing geometrically as observations get older. The smoothing equation is ",{"type":40,"tag":53,"props":61,"children":62},{},[63],{"type":46,"value":64},"Sₜ = α · yₜ + (1 − α) · Sₜ₋₁",{"type":46,"value":66},", where ",{"type":40,"tag":53,"props":68,"children":69},{},[70],{"type":46,"value":71},"α",{"type":46,"value":73}," (alpha) is the smoothing parameter between 0 and 1: α close to 1 gives almost all weight to the most recent observation (fast-adapting, noisy); α close to 0 spreads weight broadly across history (slow-adapting, smooth). The resulting level Sₜ is both the smoothed representation of the current value and the one-step-ahead forecast. This simple form is called ",{"type":40,"tag":53,"props":75,"children":76},{},[77],{"type":46,"value":78},"simple exponential smoothing (SES)",{"type":46,"value":80}," or ",{"type":40,"tag":53,"props":82,"children":83},{},[84],{"type":46,"value":85},"single exponential smoothing",{"type":46,"value":87},", and it is appropriate for stationary series with no systematic trend or seasonal pattern.",{"type":40,"tag":49,"props":89,"children":90},{},[91,96,98,103,105,110,112,117,119,124,126,131],{"type":40,"tag":53,"props":92,"children":93},{},[94],{"type":46,"value":95},"Holt's double exponential smoothing",{"type":46,"value":97}," extends SES by adding a second equation that tracks the ",{"type":40,"tag":53,"props":99,"children":100},{},[101],{"type":46,"value":102},"trend",{"type":46,"value":104}," component separately, controlled by a second parameter ",{"type":40,"tag":53,"props":106,"children":107},{},[108],{"type":46,"value":109},"β",{"type":46,"value":111}," (beta). The level equation updates the current baseline; the trend equation updates the current slope. The forecast for h steps ahead is level + h × trend, projecting the current trend linearly into the future. ",{"type":40,"tag":53,"props":113,"children":114},{},[115],{"type":46,"value":116},"Holt-Winters triple exponential smoothing",{"type":46,"value":118}," adds a third component and parameter ",{"type":40,"tag":53,"props":120,"children":121},{},[122],{"type":46,"value":123},"γ",{"type":46,"value":125}," (gamma) for ",{"type":40,"tag":53,"props":127,"children":128},{},[129],{"type":46,"value":130},"seasonality",{"type":46,"value":132},", producing a complete decomposition into level, trend, and seasonal components that can be either additive (seasonal amplitude is constant) or multiplicative (amplitude grows with the level). Holt-Winters is the workhorse of operational forecasting in retail, demand planning, and energy management — it handles trending, seasonal data with interpretable parameters and fast computation.",{"type":40,"tag":49,"props":134,"children":135},{},[136,138,143],{"type":46,"value":137},"Parameter optimization is typically done by ",{"type":40,"tag":53,"props":139,"children":140},{},[141],{"type":46,"value":142},"minimizing the sum of squared one-step-ahead forecast errors",{"type":46,"value":144}," (SSE) over the historical data. The optimal α, β, and γ are the values that best balance smoothness against responsiveness in the specific dataset. A small α (0.05–0.2) indicates the series is smooth and level; a large α (0.6–0.9) indicates the series is noisy or changes rapidly. The fitted parameters are diagnostically meaningful: a large β means the trend changes quickly (non-linear); a large γ means the seasonal pattern shifts year to year.",{"type":40,"tag":41,"props":146,"children":148},{"id":147},"how-it-works",[149],{"type":46,"value":150},"How It Works",{"type":40,"tag":152,"props":153,"children":154},"ol",{},[155,180,196],{"type":40,"tag":156,"props":157,"children":158},"li",{},[159,164,166,171,173,178],{"type":40,"tag":53,"props":160,"children":161},{},[162],{"type":46,"value":163},"Upload your data",{"type":46,"value":165}," — provide a CSV or Excel file with a ",{"type":40,"tag":53,"props":167,"children":168},{},[169],{"type":46,"value":170},"date",{"type":46,"value":172}," column and a ",{"type":40,"tag":53,"props":174,"children":175},{},[176],{"type":46,"value":177},"value",{"type":46,"value":179}," column. Monthly, quarterly, or annual data all work. Seasonal data needs at least 2–3 full seasonal cycles. One row per time point.",{"type":40,"tag":156,"props":181,"children":182},{},[183,188,190],{"type":40,"tag":53,"props":184,"children":185},{},[186],{"type":46,"value":187},"Describe the analysis",{"type":46,"value":189}," — e.g. ",{"type":40,"tag":191,"props":192,"children":193},"em",{},[194],{"type":46,"value":195},"\"Holt-Winters additive smoothing; optimize alpha, beta, gamma; 12-month forecast; 80% and 95% prediction intervals\"",{"type":40,"tag":156,"props":197,"children":198},{},[199,204,206,213,215,221],{"type":40,"tag":53,"props":200,"children":201},{},[202],{"type":46,"value":203},"Get full results",{"type":46,"value":205}," — the AI writes Python code using ",{"type":40,"tag":207,"props":208,"children":210},"a",{"href":209},"https://www.statsmodels.org/stable/generated/statsmodels.tsa.holtwinters.ExponentialSmoothing.html",[211],{"type":46,"value":212},"statsmodels ExponentialSmoothing",{"type":46,"value":214}," and ",{"type":40,"tag":207,"props":216,"children":218},{"href":217},"https://plotly.com/python/",[219],{"type":46,"value":220},"Plotly",{"type":46,"value":222}," to fit the model, produce in-sample fitted values, forecast with shaded prediction intervals, and report the smoothing parameters and error metrics",{"type":40,"tag":41,"props":224,"children":226},{"id":225},"required-data-format",[227],{"type":46,"value":228},"Required Data Format",{"type":40,"tag":230,"props":231,"children":232},"table",{},[233,257],{"type":40,"tag":234,"props":235,"children":236},"thead",{},[237],{"type":40,"tag":238,"props":239,"children":240},"tr",{},[241,247,252],{"type":40,"tag":242,"props":243,"children":244},"th",{},[245],{"type":46,"value":246},"Column",{"type":40,"tag":242,"props":248,"children":249},{},[250],{"type":46,"value":251},"Description",{"type":40,"tag":242,"props":253,"children":254},{},[255],{"type":46,"value":256},"Example",{"type":40,"tag":258,"props":259,"children":260},"tbody",{},[261,303],{"type":40,"tag":238,"props":262,"children":263},{},[264,274,279],{"type":40,"tag":265,"props":266,"children":267},"td",{},[268],{"type":40,"tag":269,"props":270,"children":272},"code",{"className":271},[],[273],{"type":46,"value":170},{"type":40,"tag":265,"props":275,"children":276},{},[277],{"type":46,"value":278},"Date or timestamp",{"type":40,"tag":265,"props":280,"children":281},{},[282,288,290,296,297],{"type":40,"tag":269,"props":283,"children":285},{"className":284},[],[286],{"type":46,"value":287},"2020-01",{"type":46,"value":289},", ",{"type":40,"tag":269,"props":291,"children":293},{"className":292},[],[294],{"type":46,"value":295},"Jan 2020",{"type":46,"value":289},{"type":40,"tag":269,"props":298,"children":300},{"className":299},[],[301],{"type":46,"value":302},"2020-01-31",{"type":40,"tag":238,"props":304,"children":305},{},[306,314,319],{"type":40,"tag":265,"props":307,"children":308},{},[309],{"type":40,"tag":269,"props":310,"children":312},{"className":311},[],[313],{"type":46,"value":177},{"type":40,"tag":265,"props":315,"children":316},{},[317],{"type":46,"value":318},"Numeric time series",{"type":40,"tag":265,"props":320,"children":321},{},[322,328,329,335,336,342],{"type":40,"tag":269,"props":323,"children":325},{"className":324},[],[326],{"type":46,"value":327},"245.3",{"type":46,"value":289},{"type":40,"tag":269,"props":330,"children":332},{"className":331},[],[333],{"type":46,"value":334},"312.1",{"type":46,"value":289},{"type":40,"tag":269,"props":337,"children":339},{"className":338},[],[340],{"type":46,"value":341},"198.8",{"type":46,"value":343}," (sales, temperature, etc.)",{"type":40,"tag":49,"props":345,"children":346},{},[347],{"type":46,"value":348},"Any column names work — describe them in your prompt. For seasonal Holt-Winters, specify the seasonal period (12 for monthly data, 4 for quarterly).",{"type":40,"tag":41,"props":350,"children":352},{"id":351},"interpreting-the-results",[353],{"type":46,"value":354},"Interpreting the Results",{"type":40,"tag":230,"props":356,"children":357},{},[358,374],{"type":40,"tag":234,"props":359,"children":360},{},[361],{"type":40,"tag":238,"props":362,"children":363},{},[364,369],{"type":40,"tag":242,"props":365,"children":366},{},[367],{"type":46,"value":368},"Parameter / Output",{"type":40,"tag":242,"props":370,"children":371},{},[372],{"type":46,"value":373},"What it means",{"type":40,"tag":258,"props":375,"children":376},{},[377,393,409,425,441,457,473,489],{"type":40,"tag":238,"props":378,"children":379},{},[380,388],{"type":40,"tag":265,"props":381,"children":382},{},[383],{"type":40,"tag":53,"props":384,"children":385},{},[386],{"type":46,"value":387},"α (alpha)",{"type":40,"tag":265,"props":389,"children":390},{},[391],{"type":46,"value":392},"Level smoothing — near 1 = reacts fast to new observations; near 0 = heavy historical averaging",{"type":40,"tag":238,"props":394,"children":395},{},[396,404],{"type":40,"tag":265,"props":397,"children":398},{},[399],{"type":40,"tag":53,"props":400,"children":401},{},[402],{"type":46,"value":403},"β (beta)",{"type":40,"tag":265,"props":405,"children":406},{},[407],{"type":46,"value":408},"Trend smoothing — near 1 = trend changes quickly; near 0 = slow-changing trend",{"type":40,"tag":238,"props":410,"children":411},{},[412,420],{"type":40,"tag":265,"props":413,"children":414},{},[415],{"type":40,"tag":53,"props":416,"children":417},{},[418],{"type":46,"value":419},"γ (gamma)",{"type":40,"tag":265,"props":421,"children":422},{},[423],{"type":46,"value":424},"Seasonal smoothing — near 1 = seasonal pattern updates rapidly year to year",{"type":40,"tag":238,"props":426,"children":427},{},[428,436],{"type":40,"tag":265,"props":429,"children":430},{},[431],{"type":40,"tag":53,"props":432,"children":433},{},[434],{"type":46,"value":435},"In-sample RMSE / MAE",{"type":40,"tag":265,"props":437,"children":438},{},[439],{"type":46,"value":440},"Average forecast error on training data — lower is better",{"type":40,"tag":238,"props":442,"children":443},{},[444,452],{"type":40,"tag":265,"props":445,"children":446},{},[447],{"type":40,"tag":53,"props":448,"children":449},{},[450],{"type":46,"value":451},"AIC / BIC",{"type":40,"tag":265,"props":453,"children":454},{},[455],{"type":46,"value":456},"Penalized goodness of fit — use to compare additive vs multiplicative models",{"type":40,"tag":238,"props":458,"children":459},{},[460,468],{"type":40,"tag":265,"props":461,"children":462},{},[463],{"type":40,"tag":53,"props":464,"children":465},{},[466],{"type":46,"value":467},"Forecast",{"type":40,"tag":265,"props":469,"children":470},{},[471],{"type":46,"value":472},"Point estimate for each future period based on projected level + trend + seasonal",{"type":40,"tag":238,"props":474,"children":475},{},[476,484],{"type":40,"tag":265,"props":477,"children":478},{},[479],{"type":40,"tag":53,"props":480,"children":481},{},[482],{"type":46,"value":483},"Prediction interval",{"type":40,"tag":265,"props":485,"children":486},{},[487],{"type":46,"value":488},"Range where future observations are expected to fall — widens with horizon",{"type":40,"tag":238,"props":490,"children":491},{},[492,500],{"type":40,"tag":265,"props":493,"children":494},{},[495],{"type":40,"tag":53,"props":496,"children":497},{},[498],{"type":46,"value":499},"Additive vs multiplicative",{"type":40,"tag":265,"props":501,"children":502},{},[503],{"type":46,"value":504},"Additive: seasonal amplitude is constant; multiplicative: amplitude grows with the level",{"type":40,"tag":41,"props":506,"children":508},{"id":507},"example-prompts",[509],{"type":46,"value":510},"Example Prompts",{"type":40,"tag":230,"props":512,"children":513},{},[514,530],{"type":40,"tag":234,"props":515,"children":516},{},[517],{"type":40,"tag":238,"props":518,"children":519},{},[520,525],{"type":40,"tag":242,"props":521,"children":522},{},[523],{"type":46,"value":524},"Scenario",{"type":40,"tag":242,"props":526,"children":527},{},[528],{"type":46,"value":529},"What to type",{"type":40,"tag":258,"props":531,"children":532},{},[533,550,567,584,601,617],{"type":40,"tag":238,"props":534,"children":535},{},[536,541],{"type":40,"tag":265,"props":537,"children":538},{},[539],{"type":46,"value":540},"Simple smoothing",{"type":40,"tag":265,"props":542,"children":543},{},[544],{"type":40,"tag":269,"props":545,"children":547},{"className":546},[],[548],{"type":46,"value":549},"SES with optimized alpha; overlay on raw data; what is the one-step-ahead forecast?",{"type":40,"tag":238,"props":551,"children":552},{},[553,558],{"type":40,"tag":265,"props":554,"children":555},{},[556],{"type":46,"value":557},"Holt trend",{"type":40,"tag":265,"props":559,"children":560},{},[561],{"type":40,"tag":269,"props":562,"children":564},{"className":563},[],[565],{"type":46,"value":566},"Holt double exponential smoothing; optimize alpha and beta; forecast 5 years; report trend slope",{"type":40,"tag":238,"props":568,"children":569},{},[570,575],{"type":40,"tag":265,"props":571,"children":572},{},[573],{"type":46,"value":574},"Holt-Winters forecast",{"type":40,"tag":265,"props":576,"children":577},{},[578],{"type":40,"tag":269,"props":579,"children":581},{"className":580},[],[582],{"type":46,"value":583},"Holt-Winters additive; seasonal period 12; optimize all parameters; 12-month forecast with 95% PI",{"type":40,"tag":238,"props":585,"children":586},{},[587,592],{"type":40,"tag":265,"props":588,"children":589},{},[590],{"type":46,"value":591},"Alpha sensitivity",{"type":40,"tag":265,"props":593,"children":594},{},[595],{"type":40,"tag":269,"props":596,"children":598},{"className":597},[],[599],{"type":46,"value":600},"SES with alpha = 0.1, 0.3, 0.5, 0.7; overlay all on raw data; which alpha minimizes RMSE?",{"type":40,"tag":238,"props":602,"children":603},{},[604,608],{"type":40,"tag":265,"props":605,"children":606},{},[607],{"type":46,"value":499},{"type":40,"tag":265,"props":609,"children":610},{},[611],{"type":40,"tag":269,"props":612,"children":614},{"className":613},[],[615],{"type":46,"value":616},"fit both additive and multiplicative Holt-Winters; compare AIC; which fits better?",{"type":40,"tag":238,"props":618,"children":619},{},[620,625],{"type":40,"tag":265,"props":621,"children":622},{},[623],{"type":46,"value":624},"Residual check",{"type":40,"tag":265,"props":626,"children":627},{},[628],{"type":40,"tag":269,"props":629,"children":631},{"className":630},[],[632],{"type":46,"value":633},"Holt-Winters fit; plot ACF of residuals; test if residuals are white noise (Ljung-Box test)",{"type":40,"tag":41,"props":635,"children":637},{"id":636},"assumptions-to-check",[638],{"type":46,"value":639},"Assumptions to Check",{"type":40,"tag":641,"props":642,"children":643},"ul",{},[644,654,677,687,697],{"type":40,"tag":156,"props":645,"children":646},{},[647,652],{"type":40,"tag":53,"props":648,"children":649},{},[650],{"type":46,"value":651},"Stationarity within model type",{"type":46,"value":653}," — SES assumes no trend or seasonality; using SES on a strongly trending series produces forecasts that lag far behind; use Holt or Holt-Winters instead",{"type":40,"tag":156,"props":655,"children":656},{},[657,661,663,668,670,675],{"type":40,"tag":53,"props":658,"children":659},{},[660],{"type":46,"value":499},{"type":46,"value":662}," — if the seasonal amplitude grows proportionally with the level (percentage swings are constant), use ",{"type":40,"tag":53,"props":664,"children":665},{},[666],{"type":46,"value":667},"multiplicative",{"type":46,"value":669},"; if amplitude is constant in absolute terms, use ",{"type":40,"tag":53,"props":671,"children":672},{},[673],{"type":46,"value":674},"additive",{"type":46,"value":676},"; the model choice affects all three components",{"type":40,"tag":156,"props":678,"children":679},{},[680,685],{"type":40,"tag":53,"props":681,"children":682},{},[683],{"type":46,"value":684},"Sufficient seasonal cycles",{"type":46,"value":686}," — Holt-Winters requires at least 2 full seasonal cycles for seasonal parameter estimation; with fewer cycles the γ estimate is unreliable",{"type":40,"tag":156,"props":688,"children":689},{},[690,695],{"type":40,"tag":53,"props":691,"children":692},{},[693],{"type":46,"value":694},"Outliers",{"type":46,"value":696}," — exponential smoothing is not robust to outliers; a single extreme value permanently shifts the level component due to the recursive smoothing; detect and handle outliers before fitting",{"type":40,"tag":156,"props":698,"children":699},{},[700,705],{"type":40,"tag":53,"props":701,"children":702},{},[703],{"type":46,"value":704},"Forecast horizon",{"type":46,"value":706}," — exponential smoothing forecasts degrade with longer horizons because the uncertainty compounds; prediction intervals widen rapidly; treat long-horizon forecasts (> one seasonal cycle) with caution",{"type":40,"tag":41,"props":708,"children":710},{"id":709},"related-tools",[711],{"type":46,"value":712},"Related Tools",{"type":40,"tag":49,"props":714,"children":715},{},[716,718,724,726,732,734,740,742,748],{"type":46,"value":717},"Use the ",{"type":40,"tag":207,"props":719,"children":721},{"href":720},"/tools/moving-average",[722],{"type":46,"value":723},"Moving Average Calculator",{"type":46,"value":725}," for simpler non-forecasting smoothing (SMA, EMA) without the full forecasting framework. Use the ",{"type":40,"tag":207,"props":727,"children":729},{"href":728},"/tools/time-series-decomposition",[730],{"type":46,"value":731},"Time Series Decomposition",{"type":46,"value":733}," tool to formally extract trend, seasonal, and residual components for inspection before or instead of exponential smoothing. Use the ",{"type":40,"tag":207,"props":735,"children":737},{"href":736},"/tools/seasonality-analysis",[738],{"type":46,"value":739},"Seasonality Analysis",{"type":46,"value":741}," tool to characterize the seasonal pattern that Holt-Winters will capture. Use the ",{"type":40,"tag":207,"props":743,"children":745},{"href":744},"/tools/autocorrelation-plot",[746],{"type":46,"value":747},"Autocorrelation Plot (ACF)",{"type":46,"value":749}," to check the residuals from an exponential smoothing fit — well-fitted residuals should show no ACF spikes.",{"type":40,"tag":41,"props":751,"children":753},{"id":752},"frequently-asked-questions",[754],{"type":46,"value":755},"Frequently Asked Questions",{"type":40,"tag":49,"props":757,"children":758},{},[759,764,769,771,776,778,783],{"type":40,"tag":53,"props":760,"children":761},{},[762],{"type":46,"value":763},"What is the difference between SES, Holt, and Holt-Winters?",{"type":40,"tag":53,"props":765,"children":766},{},[767],{"type":46,"value":768},"SES",{"type":46,"value":770}," (one parameter α) is for stationary series — no trend, no seasonality. The forecast is a constant level. ",{"type":40,"tag":53,"props":772,"children":773},{},[774],{"type":46,"value":775},"Holt",{"type":46,"value":777}," (two parameters α, β) adds a trend component — the forecast is a straight line projected from the current level at the current slope. ",{"type":40,"tag":53,"props":779,"children":780},{},[781],{"type":46,"value":782},"Holt-Winters",{"type":46,"value":784}," (three parameters α, β, γ) adds a seasonal component — the forecast is a trended line modulated by the seasonal pattern. Choose based on what the data shows: if the series is flat, use SES; if it trends without seasonality, use Holt; if it has both trend and repeating cycles, use Holt-Winters.",{"type":40,"tag":49,"props":786,"children":787},{},[788,793,795,799,801,805,807,812],{"type":40,"tag":53,"props":789,"children":790},{},[791],{"type":46,"value":792},"How do I choose between additive and multiplicative Holt-Winters?",{"type":46,"value":794},"\nLook at the seasonal amplitude over time. If December sales are always $50k above the annual average regardless of the overall sales level — the amplitude is constant — use ",{"type":40,"tag":53,"props":796,"children":797},{},[798],{"type":46,"value":674},{"type":46,"value":800},". If December sales are always 30% above the annual average (the amplitude grows as total sales grow) — use ",{"type":40,"tag":53,"props":802,"children":803},{},[804],{"type":46,"value":667},{"type":46,"value":806},". A quick visual check: plot the series; if the peaks and troughs spread further apart as the series trends up, multiplicative is likely better. Quantitatively, fit both and compare AIC — the model with lower AIC is preferred. Ask the AI to ",{"type":40,"tag":191,"props":808,"children":809},{},[810],{"type":46,"value":811},"\"fit both additive and multiplicative Holt-Winters and compare AIC\"",{"type":46,"value":813},".",{"type":40,"tag":49,"props":815,"children":816},{},[817,822],{"type":40,"tag":53,"props":818,"children":819},{},[820],{"type":46,"value":821},"What does alpha = 0.3 mean in practical terms?",{"type":46,"value":823},"\nWith α = 0.3, the current observation contributes 30% of the new smoothed level; the previous smoothed level contributes 70%. Equivalently, the effective memory of the filter extends back roughly 1/α = 3.3 periods — the most recent 3–4 observations carry the majority of the weight. A smaller α (e.g. 0.05) spreads weight across 20 periods and changes slowly; a larger α (e.g. 0.8) reacts to the last 1–2 periods. The optimal α reflects the noise level: noisy series need smaller α for stability; rapidly changing series need larger α to stay current.",{"type":40,"tag":49,"props":825,"children":826},{},[827,832,834,840,842,847],{"type":40,"tag":53,"props":828,"children":829},{},[830],{"type":46,"value":831},"How are prediction intervals computed for exponential smoothing?",{"type":46,"value":833},"\nPrediction intervals for exponential smoothing are based on the forecast error variance, which grows with the forecast horizon. For SES, the h-step-ahead forecast variance is approximately σ² × (1 + (h−1)α²) where σ² is the one-step residual variance. The 95% PI is forecast ± 1.96 × RMSE × √(variance multiplier). Statsmodels computes these analytically with the ",{"type":40,"tag":269,"props":835,"children":837},{"className":836},[],[838],{"type":46,"value":839},"simulate_smoother",{"type":46,"value":841}," method. The intervals widen more steeply for larger α (more uncertainty per step) and less steeply for smaller α. Ask the AI to ",{"type":40,"tag":191,"props":843,"children":844},{},[845],{"type":46,"value":846},"\"plot 80% and 95% prediction intervals for the Holt-Winters forecast\"",{"type":46,"value":813},{"title":7,"searchDepth":849,"depth":849,"links":850},2,[851,852,853,854,855,856,857,858],{"id":43,"depth":849,"text":47},{"id":147,"depth":849,"text":150},{"id":225,"depth":849,"text":228},{"id":351,"depth":849,"text":354},{"id":507,"depth":849,"text":510},{"id":636,"depth":849,"text":639},{"id":709,"depth":849,"text":712},{"id":752,"depth":849,"text":755},"markdown","content:tools:058.exponential-smoothing.md","content","tools/058.exponential-smoothing.md","tools/058.exponential-smoothing","md",{"loc":4},1775502471873]