When to use TBATS

Share this article

Are you wondering whether to use a TBATS model for your next data science project? Or maybe you want to hear more about the differences between TBATS models and other time series forecasting models? Well either way, you are in the right place!

In this article, we tell you everything you need to know about using TBATs models for time series forecasting. We start out by discussing what kind of data TBATS models should be used on. After that, we discuss the advantages and disadvantages of TBATS models. Finally, we provide specific examples of situations where you should and should not use TBATS models.

What data to use for TBATS models

What type of data should TBATS models be used for? TBATS models are designed to be used on time series data. Time series data is data that contains repeated measurements on a single quantity that are taken over an extended period of time.

There are multiple differences between datasets that are used for time series models and datasets that are used for standard supervised learning models like linear regression models. One example is that when you are training a time series model, you do not need to include covariates that are associated with the outcome variable. Rather than using covariates to predict future values of the outcome variable, time series models use previous values of the outcome variable to predict future values of the outcome variable.

Advantages and disadvantages of TBATS models

Advantages of TBATS models

What are some of the main advantages of TBATS models? Here are some examples of advantages that TBATS models have over other time series forecasting models.

  • Multiple seasonality. One of the main advantages of TBATS models is that they can be used to model datasets that include multiple different types of seasonal trends. That means, for example, that they can be applied to datasets that have both daily and hourly trends.
  • Does not require stationary data. TBATS models can be considered an extension of a basic exponential smoothing model. Much like basic exponential smoothing models, TBATS models do not operate under the assumption that the underlying data is stationary. That means that they can be used even when your data is not stationary and cannot easily be made stationary by differencing.
  • Does not require linear relationships in data. Another advantage of TBATS models is that they do not strictly require the dependencies between different measurements to be linear. There are multiple types of transformations that can be applied during the modeling process to account for situations where the dependencies between one value of the variable and the next are not strictly linear.
  • Highly flexible. Another advantage of TBATS models is that they are highly flexible and can be extended to model many different types of time series data. There are many different parameters that can be adjusted to tailor the model framework to specific datasets.

Disadvantages of TBATS models

What are some examples of disadvantages of TBATS models? Here are some examples of disadvantages that TBATS models have compared to other time series forecasting models.

  • Can not incorporate covariates. One disadvantage of TBATS models is that it is not easy to incorporate additional covariate information into TBATS models. That means that TBATS models are not a great option for situations where there are additional covariates that should be used to predict the outcome variable.
  • Not as well studied. Another disadvantage of the TBATS models is that they are relatively niche and not well understood by practitioners who do not have experience working with time series data. That means that it is not as easy for colleagues to give feedback on or contribute to projects that use TBATS models. Before they can contribute, colleagues will have to set aside some time to learn about TBATS models.
  • Difficult to explain. Another disadvantage of TBATS models is that they can be relatively difficult to explain. There are many different types of transformations and modifications that can be made to the model, so they are not straightforward to explain.
  • Many hyperparameters. Since there are many different types of transformation and modifications that can be made to TBATS models, there are many different hyperparameters and configuration options that need to be considered. That means that you may have to spend more time optimizing the model and determining what will work well for your dataset.
  • Computationally complex. Since there are many different hyperparameters and configurations that can be set, that means that it can be computationally intensive to arrive at a final model. Many implementations of TBATS models automatically select the best model configuration for you by trying out multiple different models and evaluating their performance. This process can take a lot of time and computational resources.
  • Only appropriate for univariate time series. TBATS models are only designed to be used on univariate time series, which means that they are not appropriate for situations where you have multiple time series that you want to model jointly.
  • Can be sensitive to outliers. Like basic exponential smoothing models, TBATS models can be sensitive to outliers. This is particularly true if the outliers occurred relatively recently before the time when the forecast is being generated.
  • Can be sensitive to mean shifts. Similarly, TBATS models can be somewhat sensitive to large mean shifts or disruption in the data. This is particularly true if the mean shift or disruption in the data happened relatively recently before the point where the forecast is being generated.
  • Common implementations cannot handle missing values natively. Another disadvantage of TBATs models is that it can be tricky to handle missing data. Most common implementations cannot handle missing data in a robust way.

When to use TBATs models

When should you use a TBATS model over another type of time series model? Here are some examples of situations where you should use a TBATS model.

  • When your data has multiple seasonality. The main reason that you should reach for TBATS over another type of time series model is if you suspect there are multiple different seasonal periods in your data. TBATS models can incorporate multiple different seasonal patterns easily, which makes they a good option for real world situations that require this capability.

When not to use TBATs models

When should you avoid using a TBATS model? Here are some examples of situations where you should avoid using TBATS models.

  • When you need a simple baseline model. TBATS models are relatively complex models that require a lot of choices to be made to specify the correct model configuration. They are not a good option if you are looking for a simple baseline model.
  • When you need computational efficiency. Since TBATS models have many different hyperparameters and configuration options that need to be specified, it can be time and resource consuming to find the right configuration. TBATS models are not a great option if you have many time series you need to model and you need a process that is computationally efficient.

Related articles

Time series models

Are you trying to figure out which machine learning model is best for your next data science project? Check out our comprehensive guide on how to choose the right machine learning model.

Other articles


Share this article

Leave a Comment

Your email address will not be published. Required fields are marked *