Back to Product Wiki

 

Overview

When creating a machine learning regression model, users can choose the method that best suits their time-series data properties. Here is the list of supported models at the moment:

1. Optimized gradient boosting
2. Prophet
3. Auto Arima

You may select the method of your choice using the drop-down menu under Method advanced settings:

 

 

Choosing the best method, in addition to the properties of the data, depends on the size of the available dataset. Examples of dataset sizes are:

• Small dataset: Months of data at hourly sampling interval
• Medium dataset: Years of data at a 15-minute sampling interval
• Large dataset: Years of data at 1-minute sampling interval

 

1. Optimized gradient boosting (default)

This is a fast and accurate model that learns the behaviour of the data by building many lightweight decision trees. Each new tree tries to correct the mistakes of the previous ones. This “boosting” process helps the model capture complex relationships and subtle patterns in the data. It is best used for:

• Medium to large datasets
• Complex, noisy patterns that simpler models might miss

This method is NOT the best choice when the user needs extrapolation capabilities (forecasting) for strong upward or downward trends that have not been seen before in the data.

 

2. Prophet

This method decomposes the time series into two parts: trend and seasonality. Trend is modeled as a piecewise-linear curve that adapts when patterns change. Seasonality is modeled using smooth curves (e.g., daily, weekly, or yearly cycles). It is best used for:

• Medium-sized datasets where the runtime remains manageable
• Data with clear repeated cycles (daily, weekly, and/or yearly)
• Handling missing data, gaps, and temporary disturbances 

 

3. Auto Arima

A classical statistical method for simpler, stable time-series data using the auto-regressive moving average technique. Auto ARIMA examines your historical values and automatically finds the best configuration. The model focuses on the autocorrelation in the data; i.e., how past values influence future values. By disabling the seasonal-trend search, the model runs much faster but is better suited to non-seasonal or lightly seasonal series. It is best used for:

• Small to medium datasets
• Stable time-series data without abrupt trend shifts or complex seasonal patterns
• When you want a simple, explainable model with reliable short-term forecasts

Patterns Part 1: What does the Pattern tool do?

Patterns Part 2: How to Create a Pattern?

Patterns Part 3: How to Create a Regression Pattern?

Patterns Part 4: How to Create a Trend Pattern?

Patterns Part 5: How to Select a Machine Learning Regression Pattern Method?