Auto PROPHET

Prophet algorithm for forecasting with a limited number of change points

  • Fourier seasonality (periods: 4, 7, 12, and 52)
  • Holiday effects with date windows
  • Robust data handling (outlier detection and missing value interpolation)
  • Competitive performance compared to Python Prophet (3% to 5% difference)
  • Fast execution (less than 1 second compared to Python 3 to 5 seconds)
  • Linear trend with optional change point detection (limited number of change points)

Piecewise Linear Trend

y(t) = (k + Σ δᵢ·I(t ≥ sᵢ)) * t + m

Where:

k = base growth rate

δᵢ = rate adjustment at changepoint i

m = base offset

sᵢ = changepoint location

Fourier Seasonality

Reuse of BATS implementation logic:

  • Support for periods: 4, 7, 12, and 52
  • Automatic Fourier term selection
  • Efficient least squares fit

Robust data handling

  • Missing value interpolation
  • Outlier detection using the modified Z score with MAD
  • Automatic fallback strategies

Holiday with presets

  • Preset selection (none or one of seven industries)
  • Optional regional selection (Global, USA, EU, and APAC). This option applies to three industries.

Automatic changepoint detection

Detects trend breaks that use a variance-based method:

  • Analysis of rolling windows for variance changes
  • Identification of the top N most significant changepoints
  • Fitting of piecewise linear trend segments

We recommend using 0 to 1 changepoints for production because Bayesian priors are rarely available and were not implemented.

Comparison with BATS and ARIMA

This table shows the feature comparison of Prophet with BATS and ARIMA:

Feature Prophet BATS ARIMA
Trend Breaks ✔ Optional (0 -1) x Smooth only ! Through differencing
Seasonality ✔ Fourier ✔ Multiple Fourier ! Decomposition
Holiday Effects ✔ Production-ready x x
Long-term Forecast ✔ Excellent ✔ Good ! Uncertain
Short-term Forecast ✔ Good ✔ Excellent ✔ Excellent
Speed ✔ <1 second ✔ <1 second ✔ <1 second

You can use these guidelines to help select the appropriate forecasting method based on your business needs and data characteristics.

Use Prophet when:

  • Holiday or event effects are critical (for example, retail, e‑commerce, hospitality).
  • Long-term forecasting is required (more than 12 periods).
  • You need to separate holiday-driven spikes from the underlying trend.
  • Business users require interpretable models.
  • We recommend using 0 changepoints to improve production stability.

Use BATS when:

  • The data contains complex, multiple seasonalities.
  • Smooth trends are required without structural breaks.
  • Short- to medium-term forecasting is sufficient.
  • Holiday effects are not needed.

Use ARIMA when:

  • Short-term forecasting is the primary requirement.
  • The data is stationary or near-stationary.
  • Confidence intervals are required for the forecast.