CSV Forecast

Forecast any time series from a CSV

Drop a file with a date and a number. Four models are fitted, tested on the data you already have, and the winner's forecast comes with 80% and 95% ranges. Everything runs in this tab; nothing is uploaded.

1 Your data

Drop a CSV here or choose a file
.csv, .tsv or .txt · read locally, never uploaded
Or try an example (free, instant):

How it works

Reading your file
The date and value columns are found for you. The spacing between dates sets the frequency (hourly, daily, business days, weekly, monthly, quarterly, yearly). Repeated dates are added up or averaged, and missing periods are filled in, with a note saying how many.
Finding the cycle
The series is detrended and checked for a repeat at the natural lengths for its frequency: 7 days, 52 weeks, 12 months, 4 quarters, 24 or 168 hours. You can override it.
The models
A seasonal-naive baseline repeats the last cycle. Holt-Winters tracks a level, a trend and a seasonal pattern, either a fixed-size swing (additive) or one that grows with the level (multiplicative). STL + ETS peels the seasonal pattern off with loess smoothing, forecasts what is left, then adds the pattern back.
The test
The last stretch of your data is held back, each model forecasts it from what came before, and the errors are scored. Rolling backtests repeat this at several cut-off points. The winner is then refitted on all of the data to forecast the future.
The ranges
80% and 95% prediction intervals widen with the horizon. For the multiplicative model they are an approximation; the coverage column tells you how well they held up on your own data.

Questions

What format does my CSV need?
One column with dates and one with numbers. Dates can be 2024-03-01, 03/01/2024, 1/3/2024, Mar 2024, 2024Q1, 2024-03 or a timestamp with a time. Commas, semicolons and tabs all work, and numbers like 1,234.50 or $99 are read correctly. Extra columns are fine; a text column such as region can be used as a filter.
Which forecasting models does it use?
Four: a seasonal-naive baseline that repeats the last cycle, Holt-Winters with an additive season, Holt-Winters with a multiplicative season, and STL decomposition followed by exponential smoothing on the seasonally adjusted series. The trend type (none, linear or damped) is chosen by AICc. Each is fitted in your browser.
How does it pick the best model?
It hides the last stretch of your data, fits every model on what comes before, and measures how close each forecast lands. With rolling backtests this repeats for several cut-off points and the errors are averaged. Models are ranked by MAPE, or by MASE when the series has zeros or values near zero.
Is my data uploaded anywhere?
No. The file is read and every model is fitted inside your browser tab. The page makes no network request with your data, and there is no sign-in.