Reference: Machine Learning Algorithms for Adaptive Planning

Adaptive Planning offers several algorithms for Predictive Forecaster and 1 for Anomaly Detection. All the algorithms provide sufficient results. Still, we recommend that you do a trial and error process to find the best algorithm for your data.

The table provides some general guidelines and capabilities to help you make your selections. Always check your forecasted data for accuracy.

Algorithm Comparison

In addition to the 3:2 ratio between reference and forecasted data points, you must also have the minimum data points listed in the table for each algorithm.

Algorithm	Details	Requirements and Capabilities
AutoFit	Classifier model	Seasonality : Yearly, Biannual, and Quarterly with a Weekly and a Multiplicative option. Minimum data points in reference : 15 Levers : Yes Notes : Ideal for data that’s: Complex and might require different algorithms per data series. Difficult for you to understand. AutoFit can apply algorithms that don’t support certain capabilities. Example: When AutoFit applies Croton, it doesn’t use seasonality or levers.
ARIMA	Auto-regressive and moving average model.	Seasonality: Automatically considers seasonal impacts without user-defined specifications. Minimum data points in reference : 10 Levers: Yes Notes: Ideal for short-term forecasting, not ideal for longer-term forecasting. Good at forecasting volatile metrics.
Croston	Exponential smoothing model.	Seasonality : No Minimum data points in reference : 2 Levers : No Notes : Ideal for sparse data with intermittent demand. The results are the average forecast for the period. Example: You’ll get the average (the same value) across the time period.
Holt-Winters	Additive or multiplicative seasonality model. Exponential smoothing.	Seasonality :Yearly, Biannual, and Quarterly with a Multiplicative option. Minimum data points in reference : 9 Levers : No Notes : Ideal for data where the most recent data requires more weight. Example: During the pandemic, purchases had unusual patterns. In recent years, purchase patterns normalized. Computationally efficient.
Kalman Filter	A state space model. Estimates future values based on the known historical values in a time series. Recursively estimates the average values of the forecasts considering the noise components at each step and the estimates obtained in the previous step.	Seasonality : Option where you can indicate whether or not you want to account for seasonality. Minimum data points in reference : 2 Levers : No Notes : Ideal for random fluctuations and no obvious patterns. Computationally efficient.
LightGBM	Supervised learning model. Uses lags and their impact on the forecast.	Seasonality : No Minimum data points in reference : 12 Levers : Yes Notes : Ideal for demand forecasting scenarios.
N-BEATS	Metadeep learning model bootstrapped with public datasets. Creates a neural network that has general knowledge about the historical time series. Trained on more than 60,000 time series that the public domain also uses. Fast because it uses a single model for every time series.	Seasonality : Automatically considers seasonal impacts without user-defined specifications. Minimum data points in reference : 2 Levers : No
Orbit DLT	Additive and damped trend model	Seasonality :Yearly, Biannual, and Quarterly. Minimum data points in reference : 2 Levers : No Notes : Ideal for data with: Clear trends. Consistent seasonality. Little to no outliers.
Prophet	An additive model with nonlinear trends Frames the forecasting process as a curve-fitting problem and accommodates seasonalities and external levers.	Seasonality : Yearly, Biannual, and Quarterly with a Weekly and a Multiplicative option. Minimum data points in reference : 2 Levers : Yes Notes : Ideal for data with: Shifts in trends. Outliers.

Setup Considerations: Predictive Forecaster

Create Machine Learning Forecasts

Build Lever Modeled Sheets for Predictive Forecaster

Concept: The AutoFit Algorithm

FAQ: Predictive Forecaster