Out-of-Sample time series forecasting is a common, important, and subtle task. The OOS package introduces a comprehensive and cohesive API for the out-of-sample forecasting workflow: data preparation, forecasting - including both traditional econometric time series models and modern machine learning techniques - forecast combination, model and error analysis, and forecast visualization.
The key difference between OOS and the other time series forecasting packages is that it operates out-of-sample by construction. That is, OOS functions are designed to re-clean data and re-train models each forecast date and are careful not to introduce look-ahead bias into its information set via data cleaning or forecasts via model training. This framework is in contrast to other packages which tend to focus on cleaning data or fitting a forecast model once, leaving the user to construct the out-of-sample exercise on their own.
Available tools and techniques are summarized below and may be reviewed under the Tools tab. While vignettes and other extended documentation of the OOS package’s capabilities may be found under the Workflow tab.
|Clean Outliers||Impute Missing Data (via imputeTS)||Dimension Reduction|
|Winsorize||Linear Interpolation||Principal Components|
|Univariate Forecasts (via forecast)||Multivariate Forecasts (via caret)||Forecast Combinations|
|Random Walk||Vector Autoregression||Mean|
|ETS||LASSO Regression||Trimmed (Winsorized) Mean|
|Theta Method||Elastic Net||Linear Regression|
|TBATS||Principal Component Regression||LASSO Regression|
|STL||Partial Least Squares Regression||Ridge Regression|
|AR Perceptron||Random Forest||Partial Egalitarian LASSO|
|Tree-Based Gradient Boosting Machine||Principal Component Regression|
|Single Layered Neural Network||Partial Least Squares Regression|
|Tree-Based Gradient Boosting Machine|
|Single Layered Neural Network|
If you should have questions, concerns, or wish to collaborate, please contact Tyler J. Pike