Comment on page
Explore the different forecasting models available to use in the Claret App to help you best leverage your past data and make informed predictions about the future.
The statistical models available to use in Claret are mathematical models that use historical data, or other sale types you wish, to predict what will happen in future.
Autoregression uses observations from previous time steps (lags) as input to a linear regression equation to predict the value at the next time step. It will first establish the lags that correlate most and then apply them as input variables to a linear regression model to predict the future value.
Autoregressive models assume the future to tightly resemble the past and as such will most likely be chosen for clean and seasonal time series.
ARIMA models are an evolution of Autoregressive models. They too predict future values based on past values with the help of a linear regression, which takes past values as an input (that's the AR part of ARIMA). But they transform the past values a little more by (1) removing trend and seasonality to focus the prediction on the underlying data and (2) smoothing these values with a moving average.
ARIMA models are widely used and popular as they can adapt to many types of time series.
Forecasts produced by Exponential Smoothing methods are weighted averages of past observations, with the weights decaying exponentially as the observations get older. In other words, the more recent the observation the higher the associated weight.
ETS models come in different degrees of complexity from simple methods to models including trend and seasonal components.
The Naïve method simply uses the last value of the observations as future predictions. Seasonal Naïve methods (SNAÏVE) identify seasonal periods in the data (months, quarters, or else) to set each future value accordingly.
While simple, naïve methods are surprisingly accurate particularly for noisy data.
A Time Series Linear Model (TSLM) is a linear regression model that predicts the course of one time series based on another time series. An example would be predicting sales based on marketing expenditure. An example for a single time series would be predicting future values based on previous values.
The VAR model is a more complex model allowing all variables of the model to affect each other. With simple linear regression forecasts, inputs were affecting outputs unidirectionally. A VAR makes sense when relating multiple time series to each other that can affect each other, for example if we wanted to predict sales based on income, we would assume that higher incomes drive sales, which might positively affect incomes.