Are you wondering whether you should use an exponential smoothing model for your data science project? Or maybe you are more interested in learning about the differences between exponential smoothing models and other time series models? Well either way, you are in the right place!
In this article, we tell you everything you need to know to understand when to use exponential smoothing models. First, we talk about what types of datasets exponential smoothing models should be used for. After that, we talk about a few different models that fall under the category of exponential smoothing models. Next, we discuss the main advantages and disadvantages of exponential smoothing models. Finally, we provide specific examples of cases where you should and should not use exponential smoothing models.
What kind of data to use exponential smoothing models for
What kind of data should you use exponential smoothing models for? Exponential smoothing models should be used when you are working with time series data. Time series data is data such that there are many repeated measurements of a specific quantity that are taken over time.
One thing that differentiates time series models like exponential smoothing models from traditional supervised learning models like linear regression models is that you do not have to have features that are associated with your outcome variable to build a linear regression model. All you need is an outcome variable that is measured repeatedly over time. Rather than using features to forecast future values of the outcome variable, time series models use previous values of the outcome variable to forecast future values of the outcome variable.
Types of exponential smoothing models
What are some specific examples of models that fall into the exponential smoothing family? Here are some common examples of models that fall into the exponential smoothing family.
- Simple exponential smoothing. A simple exponential smoothing model is the most simple implementation of an exponential smoothing model. Simple exponential smoothing models are used for data that does not have any trend or seasonality. Rather than including projections related to trend or seasonality into the forecast, simple exponential smoothing models produce flat forecasts that have the same value for all future observations.
- Holt’s method. Holt’s method is an extension of a simple exponential smoothing model that allows you to model data that has a constant trend. This can be a trend that is increasing or decreasing at a constant rate. This method is only appropriate for data where the trend is expected to increase or decrease at the same rate indefinitely. While Holt’s method can count for a simple trend, it cannot account for seasonality and should not be used on seasonal data.
- Holt’s dampened method. Holt’s dampened method is an extension to Holt’s method that introduces a dampening parameter that enables you to model a reduction in the magnitude of the trend over time. This means that you do not have to make the assumption that the trend will continue to increase (or decrease) at exactly the same rate indefinitely. This method still cannot be used to model data with seasonal patterns.
- Holt Winters. Holt Winters is an extension to other exponential smoothing models that incorporates the ability to account for seasonality. This model can be used to model data that has both seasonal patterns and a trend.
Advantages and disadvantages of exponential smoothing
Advantages of exponential smoothing
What are some of the main advantages of exponential smoothing models compared to other time series models? Here are some of the main advantages of exponential smoothing models.
- Well understood. One of the main advantages of exponential smoothing models is that they are relatively common and well understood. This means that collaborators and teammates will be more likely to be familiar with and understand this type of model. As a result, teammates will be able to ramp onto and contribute to projects that use exponential smoothing models more quickly than projects that use more obscure models.
- Explainable. In addition to being common and well understood, it is also relatively straightforward to explain how exponential smoothing modes work. That means that they are a good option if you are working with skeptical stakeholders who do not trust models that they do not understand.
- Does not require stationary data. Another advantage of exponential smoothing models is that they do not require the data that is used to train them to be stationary. This means that they can be used to handle datasets that cannot be handled by other models like ARIMA models.
- Suitable for small datasets. Another advantage of exponential smoothing models is that they are suitable for relatively small datasets. That is because they do not have many parameters that need to be tuned.
- Relatively small set of hyperparameters. Another advantage of exponential smoothing models is that they have a relatively small number of hyperparameters that need to be specified before training the model. This means that it is faster and easier to arrive at an appropriate choice of hyperparameters..
- Lower time complexity. Another advantage of exponential smoothing models is that they generally have lower time complexity than more complex time series models such as neural network models. Even simple models like ARIMA models can have high time complexity under the right conditions, so this is a relatively large advantage if you have limited computational resources.
- Handles missing data. Similarly to ARIMA models, exponential smoothing models can be formulated to handle data that has missing values. That being said, many common statistical packages do not include functionality for handling missing values natively in their implementation. Practically, it may be easier to use other models like ARIMA models in these cases.
- Easier to handle nonlinear dependencies in data. It is easier to extend exponential smoothing models to handle nonlinear dependencies than it is to do so for other models like ARIMA modes.
Disadvantages of exponential smoothing
What are some of the main disadvantages of exponential smoothing models? Here are some of the main disadvantages of exponential smoothing models.
- Cannot handle covariates. One of the main advantages of exponential smoothing models is that they generally cannot incorporate information related to covariates. There are some extensions of common exponential smoothing models that can incorporate covariate information, but they are not always available in forecasting packages. That means that exponential smoothing models are not the best option for use cases where there are important covariates that need to be incorporated into the model.
- Cannot handle multiple seasonality. Another disadvantage of exponential smoothing models is that they do not perform well in situations where there is multiple seasonality. That means they might not be suitable for time series that have multiple types of seasonal patterns, such as time series that have both weekly trends and daily trends.
- Struggle with recent mean shifts. Another disadvantage of exponential smoothing models is that they can face difficulties when presented with data with mean shifts or similar disruptions. This is especially true if the mean shifts were relatively recent.
- Only intended for univariate time series. Another disadvantage of exponential smoothing is that it is only intended to be used for univariate time series. That means there may be better methods to use if you have multiple correlated time series that you want to model jointly.
- Sensitive to recent outliers. Another disadvantage of exponential smoothing models is that they are somewhat sensitive to outliers in historical data. This is particularly true for outliers that occurred relatively recently.
When to use exponential smoothing models
When should you use exponential smoothing models over other common time series models? Here are some examples of situations where it makes sense to each for exponential smoothing models.
- Baseline that is simple and understandable. Exponential smoothing models are a good option to reach for when you want to use a model that is simple, easy to understand, and well studied. This means that it is a good option in situations where you want to use a common model so that other teammates can easily onboard onto a project. It is also a good option for situations where you have skeptical stakeholders who do not trust more complex models that cannot be easily explained.
- Data is not stationary. Exponential smoothing is a particularly good option to reach for when you want to use a simple model and your data is non stationary and cannot easily be made stationary. Unlike other models like ARIMA models, exponential smoothing models do not require the data to be made stationary.
When not to use exponential smoothing models
When should you avoid using an exponential smoothing model? Here are some examples of situations where you should avoid using exponential smoothing models.
- Peak performance is required. While exponential smoothing models perform well as a simple baseline, they rarely offer the best performance when they are compared against other time series models. That means that they are generally not the best choice if you are operating in an environment where small increases in predictive performance will result in large increases in delivered business value. If your data can be made stationary, even other simple models like ARIMA models tend to offer stronger performance.
- You need to incorporate covariates. It is not straightforward to incorporate covariates into exponential smoothing models. If you have covariates that need to be included, you are better off turning to an ARIMA model or a more complex model.
- When your data has multiple seasonality. If your data has multiple different seasonal trends, then you may be better off using a model that can handle multiple seasonality. A TBATS model is a great example of a model that can handle this type of data natively.
Time series models
- When to use ARIMA models
- When to use TBATS models
- When to use Facebook Prophet
- When to use Fourier ARIMA 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.