This course teaches students to build, refine, extrapolate, and, in some cases, interpret models designed for a single, sequential series. There are three modeling approaches presented. The traditional, Box-Jenkins approach for modeling time series is covered in the first part of the course. This presentation moves students from models for stationary data (or ARMA) to models for trend and seasonality (ARIMA) and concludes with information about specifying transfer function components in an ARIMAX, or time series regression, model. A Bayesian approach to modeling time series is considered next. The basic Bayesian framework is extended to accommodate autoregressive variation in the data as well as dynamic input variable effects. Machine learning algorithms for time series is the third approach. Gradient boosting and recurrent neural network algorithms are particularly well suited for accommodating nonlinear relationships in the data. Examples are provided to build intuition on the effective use of these algorithms. The course concludes by considering how forecasting precision can be improved by combining the strengths of the different approaches. The final lesson includes demonstrations of creating combined (or ensemble) and hybrid model forecasts.
Learn How To
Generate diagnostic plots to identify and understand components of systematic variation in time series data. Understand irregular or ARMA type variation in stationary data, and identify an appropriate model. Perform tests for trend and seasonality in time series, and accommodate these components in an ARIMA model. Specify, estimate, and generate forecasts from an ARIMAX or dynamic times series regression model. Extend the basic Bayesian modeling framework to accommodate autoregressive and dynamic regressor effects. Create predictions of the future from the Bayesian time series model. Understand how machine learning algorithms like gradient boosting and recurrent neural networks work on sequential data. Build, refine, and generate forecasts from machine learning algorithms. Combine predictions using ensemble and hybrid modeling approaches.Who Should Attend
Analysts with a background in linear models
Prerequisites
This course uses a variety of different software tools. Familiarity with Base SAS, SAS/ETS, SAS/STAT, and SAS Visual Forecasting, as well as open-source tools for sequential data handling and modeling, is helpful but not required. The lessons on Bayesian analysis and machine learning models assume some prior knowledge of these topics. One way that students can acquire this background is by completing these SAS Education courses: Bayesian Analyses Using SAS(R) and Supervised Machine Learning Pipelines Using SAS(R) Viya(R).
SAS Products Covered
SAS/STAT;SAS Visual Forecasting;SAS/ETS
Course Outline
Introduction to Time Series Models
A review of time series components and concepts.Simple models for time series. Exponential smoothing models.Traditional Time Series ModelsARMA models, stationary. ARMA models, trend. ARIMA models, seasonality. ARIMAX models, inputs.Bayesian Time Series AnalysisReview of Bayesian. Bayesian time series structure. Exogenous variables. Forecasting in Bayesian. Machine Learning Approaches to Time Series ModelingUsing machine learning models for time series forecasting. Deep learning with recurrent neural networks for time series forecasting. Hybrid Modeling Approaches and External ForecastsA hybrid or ensemble approach to forecasting. Combining traditional time series methods with machine learning methods. Using externally generated forecasts.
