Skip to content

Time series

Time series applies machine learning techniques to sequential data, where observations are ordered over time. This approach is crucial for predicting future values or events and finds applications in finance, healthcare, weather forecasting, and more. ATOM support two time series tasks: univariate forecast and multivariate forecast.

Example

See our time series examples for univariate and multivariate forecast.


Exogenous variables

Exogenous variables are external factors that can influence the target variable and, unlike endogenous variables, are not part of the time series being analyzed.

Incorporating exogenous variables into time series models helps capture additional information that may impact the observed patterns. This inclusion allows for a more comprehensive understanding of the underlying dynamics and can lead to more accurate predictions.

Exogenous variables are added to atom with the X variable. When no exogenous variables are provided, atom.X returns an empty dataframe. Note that not all models make use of exogenous variables. Read more here about how to check specific model characteristics.


Seasonality

Seasonality refers to the recurring patterns that repeat at regular intervals over time, often corresponding to specific time periods, such as days, weeks, or months, and can significantly influence the observed data.

Add seasonality to atom using the sp parameter or attribute. You can add a single value for single seasonality or a sequence of values for multiple seasonalities. If you don't know the seasonality a priori, you can use the get_seasonal_period method to automatically detect the seasonality, e.g. atom.sp = atom.get_seasonal_period() or directly from the constructor atom = ATOMForecaster(y, sp="infer").

The majority of models only support one seasonal period. If more than one period is defined, such models only use the first one. Read here how to check which models support multiple seasonality.

Info

In a multivariate setting, the same period is used for all target columns.

In addition to the period, it's possible to further tune the seasonality by specifying the trend and seasonal models. In an additive model, the components are added together. It implies that the effect of one component does not depend on the level of the other components. In a multiplicative model, the components are multiplied together. This suggests that the effect of one component is proportional to the level of the other components.

Specify the trend and/or seasonal models providing the sp parameter (or attribute) with a dictionary, e.g., atom.sp = {"sp": 12, "seasonal_model": "multiplicative"}. Both the seasonal_model and trend_model values default to None.


Forecasting with regressors

All of ATOM's regressors can also be used in forecasting tasks. Simply select the regressor like any other model, e.g., atom.run(models="RF") to use a RandomForest model.

The regressor is automatically converted to a forecaster, based on reduction to tabular or time-series regression. During fitting, a sliding-window approach is used to first transform the time series into tabular or panel data, which is then used to fit a tabular or time-series regression estimator. During prediction, the last available data is used as input to the fitted regression estimator to generate forecasts.

See below a graphical representation of the reduction logic using the following symbols:

  • y: forecast target.
  • x: past values of y that are used as features (X) to forecast y.
  • *:observations, past or future, neither part of the window nor forecast.

Assume we have the following training data (15 observations):

|------------------------------| | * * * * * * * * * * * * * * *| |------------------------------|

The reducer targets the first data point after the window, irrespective of the forecasting horizons requested. In the example, the following five windows are created:

|------------------------------| | x x x x x x x x x x y * * * | | * x x x x x x x x x x y * * | | * * x x x x x x x x x x y * | | * * * x x x x x x x x x x y | | * * * * x x x x x x x x x x y| |------------------------------|

Warning

Regressor forecasters do not support in-sample predictions. Scores on the training set always return NaN.