# CSV Forecast

> A free, in-browser time-series forecaster at https://csv-forecast.skillsafe.ai/. Upload or paste a
> CSV with a date column and a numeric column; the page detects the frequency and seasonal cycle,
> fits four forecasting models, backtests them on the most recent data, and charts the best forecast
> with 80% and 95% prediction intervals. There is no AI model, no sign-in, no API and nothing
> metered: every computation runs in the visitor's browser and the data is never uploaded.

## What it does

- Reads CSV, TSV or semicolon-separated text, with or without a header row. Finds the date column
  and the value column automatically (both can be changed). Understands ISO dates, US and
  day-first slash dates (decided per column, with a manual override), month names (Mar 2024,
  15 Jan 2024), year-month (2024-03), quarters (2024Q1, Q1 2024), plain years and timestamps.
  Reads numbers like 1,234.50, 1.234,50, $99, (12) and 45%.
- Optional filter on a text column (for example region or product) or all groups combined.
- Detects the frequency from the spacing between dates: minute, hourly, business-day, daily,
  weekly, monthly, quarterly, yearly. Can resample to a coarser frequency (sum or average).
  Repeated dates are added up or averaged; missing periods are filled by linear interpolation or 0,
  and the page says how many.
- Detects the seasonal cycle by detrending and checking the autocorrelation at the natural lags
  (7 days, 5 business days, 52 weeks, 12 months, 4 quarters, 24 or 168 hours). Prefers the longest
  cycle that is clearly present with at least three repeats. Overridable.
- Models: seasonal naive (baseline); Holt-Winters with additive season; Holt-Winters with
  multiplicative season (only when all values are above zero); STL decomposition (loess,
  Cleveland et al. 1990) followed by exponential smoothing on the seasonally adjusted series.
  The trend type (none, linear, damped) is chosen by AICc unless set by hand. Smoothing
  parameters are fitted by Nelder-Mead on the one-step-ahead squared error.
- Backtest: one holdout or 3 or 5 rolling cut-offs. Reports MAPE, sMAPE, MAE, RMSE, MASE and how
  often the actual values fell inside the 80% and 95% ranges. Ranks by MAPE, or by MASE when the
  series has zeros or values near zero. The winner is refitted on all data for the forecast.
- Prediction intervals: analytic ETS variance for the additive models; an approximation for the
  multiplicative model. Floored at 0 when the history is never negative.
- STL decomposition panels: observed, trend, seasonal, remainder, with seasonal and trend
  strength.
- Exports: forecast CSV for the chosen model (date, model, forecast, lower_80, upper_80, lower_95,
  upper_95), the same for all models, the backtest predictions as CSV, the chart as SVG, and a
  plain-text summary to the clipboard.

## Verification

The ETS recursions (additive and multiplicative season, linear and damped trend) and the STL
decomposition (plain and robust) reproduce statsmodels' ETSModel and STL to within 1e-10 on the
airline-passenger and Mauna Loa CO2 series for identical parameters.

## Example data bundled

- Airline passengers, monthly 1949 to 1960 (Box and Jenkins).
- Mauna Loa monthly mean CO2 since 1995 (NOAA Global Monitoring Laboratory, public domain).
- Synthetic: daily online-shop orders, hourly electricity load, quarterly revenue by region.

## Limits

- Up to 20,000 periods after resampling; files up to 60 MB.
- One seasonal cycle per model. Daily data uses the weekly cycle; resample to monthly for a
  yearly pattern.
- Excel files must be saved as CSV first.