A time series is a set y1, y2, ... ,yN of observations that are observed over a series of equally spaced time periods. Some examples of time series data include quarterly sales reports, monthly average temperatures, and counts of sunspots. The Time Series platform enables you to explore patterns and trends found in these types of data. You can then use these patterns and trends to forecast, or predict, into the future.
Characteristics that are common in time series data include seasonality, trend, and autocorrelation. Seasonality refers to patterns that occur over a known period of time. For example, data that are collected monthly might look similar in summer months across all years of data collection. Trend refers to long term movements of a series, such as gradual increases or decreases of values across time. Autocorrelation is the degree to which each point in a series is correlated with earlier values in the series.
There are many different models and forecasting methods available in the Time Series platform. However, not all methods can handle trend or seasonality. In order to choose an appropriate model, it is essential to determine which characteristics are present in the series. The Time Series platform provides graphs such as variograms, autocorrelation plots, partial autocorrelation plots, and spectral density plots that can be used to identify the type of model appropriate for describing and forecasting the evolution of the time series. There are also several differencing and decomposition methods in the platform that enable you to remove seasonal or general trends in the data to explore and simplify the analysis. You can also view and apply a Box-Cox transformation to your data.
Alternatively, the platform can fit more sophisticated models that can incorporate seasonality and long term trends. One such model in the platform that has this ability is Winter’s Additive Method, which is an advanced exponential smoothing model. In addition, the platform can fit AutoRegressive Integrated Moving Average (ARIMA) models and State Space Smoothing models. Both of these types of models are the most statistically complex, but also provide the most flexibility. Advanced exponential smoothing, ARIMA, and State Space Smoothing models are harder to interpret, but they are excellent tools for forecasting.
The Time Series platform can also fit transfer function models when supplied with an input series.
Note: ARIMA models are also known as Box-Jenkins models.