Time Series Forecaster
Paste a series of numbers and project it forward. Choose simple exponential smoothing, Holt's linear trend, a damped trend, or Holt-Winters seasonality — or let the tool pick by AICc. You get the fitted smoothing weights, accuracy metrics, and a forecast table with prediction intervals, computed entirely in your browser.
About this tool
Forecast a single numeric series without leaving the browser. Paste one value per row, a labelled
CSV-style row such as Jan,120, or a single comma-separated series, then choose an exponential
smoothing model: simple level smoothing, Holt's linear trend, a damped trend, or Holt-Winters
seasonality. If you leave Model on auto, the tool fits every applicable candidate and picks the
lowest-AICc result.
The output reports the selected model, fitted smoothing weights, in-sample accuracy metrics, and a future forecast table with lower and upper prediction bands. Optional fitted values make it easier to see where the model tracked the history well and where residuals were large. Everything is computed locally with deterministic WebAssembly, so the page, CLI, and chat tool use the same core logic.
Worked example
Paste this monthly sales series:
month,sales
Jan,120
Feb,132
Mar,141
Apr,158
May,166
Jun,181
Jul,190
Aug,203
Set Model to holt, Periods to forecast to 3, Prediction interval to 95%, and
Decimal places to 2. The forecast continues the fitted trend and returns a table like:
period forecast lower upper
9 ... ... ...
10 ... ... ...
11 ... ... ...
Use Holt-Winters additive with a season length such as 12 for monthly data with a yearly cycle,
or 4 for quarterly data. Multiplicative seasonality is useful when seasonal swings grow with the
level, but it requires strictly positive observations.
Limits and edge cases
- Input needs at least 3 observations and accepts at most 10,000 observations.
- Forecast horizon accepts 1–240 periods; the page slider focuses on the common 1–60 range.
- Seasonal models need
season_length >= 2and at least two full cycles of history. - Smoothing weights (
alpha,beta,gamma,phi) left at0are fitted automatically; set a positive value to pin that weight. - Multiplicative Holt-Winters rejects zero or negative observations because seasonal ratios would be undefined.
- Prediction intervals are approximate ETS residual bands, not a guarantee. Damped and multiplicative seasonal intervals are especially approximate at long horizons.
MAPEis omitted when actual values are zero;MASEneeds enough naive seasonal differences to establish a scale.- This is a univariate forecaster. It does not model promotions, regressors, holidays, multiple series, missing timestamps, or machine-learning features.
FAQ
Which model should I choose?
Use auto when you want a quick baseline: it tries the applicable exponential-smoothing models
and chooses by AICc. Use simple for a mostly flat series, Holt for a clear trend,
damped when the trend should flatten over time, and Holt-Winters when the same seasonal
pattern repeats every season_length observations.
What does season length mean?
Season length is the number of observations in one repeat cycle. Monthly data with yearly seasonality
uses 12, quarterly data uses 4, weekly data with daily observations uses 7, and non-seasonal
data uses 0. Seasonal models need at least two complete cycles so the starting seasonal indices can
be estimated.
Do I need to set alpha, beta, gamma, or phi?
Usually no. Leave them at 0 and the tool fits deterministic values by minimising one-step-ahead
squared errors. Set a positive value only when you need to reproduce a known model or compare a
specific smoothing weight against the automatic fit.
Can I paste labelled rows or currency values?
Yes. Rows like Jan, $120, percent-marked values, underscores, and accounting negatives such as
(12.5) are normalised before fitting. The last field in each row is treated as the numeric value;
earlier fields become labels for the optional fitted table.
Are the prediction intervals statistical confidence guarantees?
No. They are practical residual-based forecast bands for the selected ETS model. They are useful for rough planning and anomaly checks, but they do not replace a full statistical modelling workflow, especially for intermittent demand, structural breaks, or long seasonal horizons.
Developer & Automation Access
Run it from the terminal
Same engine as this page, headless — via the gizza CLI:
gizza tool time-series-forecaster "month,sales
Jan,120
Feb,132
Mar,141
Apr,158"New to the CLI? Get gizza →
Open it by URL
Pre-fill and auto-run this tool with query parameters — the names match the API/CLI:
https://gizza.ai/tools/time-series-forecaster/?data=month%2Csales%0AJan%2C120%0AFeb%2C132%0AMar%2C141%0AApr%2C158&model=auto&horizon=6&season_length=0&alpha=0&beta=0&gamma=0&phi=0&confidence=95&show_fitted=true&header=auto&decimals=3&format=textMachine-readable descriptor: tool.json — title + parameters JSON Schema for agents.