Time series algorithms

The Time Series mining function provides algorithms that are based on different underlying model assumptions with several parameters. The learning algorithms try to find the best model and the best parameter values for the given data.

If you do not specify a seasonal cycle, it is automatically determined. Also, missing values and non-equidistant time series are automatically interpolated.

The Time Series mining function provides the following algorithms to predict future trends:

Which of the algorithms creates the best forecast of your data depends on different model assumptions. You can calculate all forecasts at the same time. The algorithms calculate a detailed forecast including seasonal behavior of the original time series. With the Time Series Visualizer, you can evaluate and compare the resulting curves.

Autoregressive Integrated Moving Average (ARIMA)

The ARIMA algorithm also incorporates seasonal components. Therefore this algorithm is also referred to as Seasonal ARIMA (SARIMA).

The autoregressive part of the algorithm uses weighted previous values while the moving average part weighs the previously assumed errors of the time series.

The ARIMA algorithm assumes the error to be independent and identically distributed from a normal distribution with zero mean. The basic ARMA model works for stationary time series only. Stationary time series contain equal mean and equal variance for the whole time series. Therefore the integrated part creates stationary series by differentiation.

Exponential Smoothing

Exponential Smoothing can consist of the following components:
  • Basic level at a certain point in time.
  • Trend.

    The trend can have additive or multiplicative characteristics. Also, it can be damped or non-damped.

  • The seasonal component.

Dependant on the data, trend and the seasonal component are optional. There are ARIMA models that correspond to Exponential Smoothing models and vice versa.

Seasonal Trend Decomposition

Seasonal Trend Decomposition fits different seasonal trend functions to the given data and selects the best seasonal trend function according to an error measure. The following trends are used during the training run:
  • Linear trend
  • Quadratic trend
  • Cubic trend
  • Logarithmic trend
  • Exponential trend
  • Hyperbolic trend

The seasonality is incorporated in an additive or multiplicative way.



Feedback | Information roadmap