A time series is a sequence of data points logged at regular intervals of time. Time series analysis is used to analyze trends, seasonality, and other patterns to understand past behavior and predict future values; it is valuable across various fields, such as marketing, finance, healthcare, and so on, enabling informed decision-making and forecasting based on historical data sets.
In general, time series forecasting can be classified into two main types, namely univariate and multivariate. The former involves predicting future values based on a single variable’s historical data, for example, if you are recording hourly wind speed values for a given geographical area and intend to forecast the future wind speed using historical wind speed records.
Time Series Forecasting Models and Techniques
On the other hand, Multivariate Time Series forecasting uses multiple related variables to predict future values. An example scenario would be the example above for wind speed forecasting, but multivariate forecasting also accounts for other weather-related features, such as atmospheric pressure, solar storms intensity, and temperature, among others. In summary, the multivariate scenario takes several variables into account to predict a target variable, in this case, the wind speed.
Overall, time series forecasting algorithms can be split into the following classes:
- Classical / Statistical Models: Additive models, moving averages, Exponential Smoothing, ARIMA, SARIMA, TBATS
- Machine Learning approach: Linear Regression, XGBoost, Random Forest
- Deep Learning: recurrent neural networks (RNNs), long short-term memory (LSTM), convolutional neural networks (CNNs), and transformers which can perform very well in tasks needing to account for both short-term and long-term trends, namely share prices forecasting and weather prediction
Time-series additive models consider that a time series can be split into its components (trend, seasonality, cyclical, and irregular) by adding them together. That is, the observed value at any given time point is the sum of these individual components.
In SAP HANA, an additive model [1] is a type of time series forecasting model where the observed value is the sum of trend, seasonality, and residual components. It is a key part of the SAP HANA Predictive Analysis Library (PAL) and is particularly useful for data exhibiting strong seasonal patterns.