Regressor cascading for time series forecasting

Abstract

Time series forecasting is the process of predicting future values of a time series based on its historical data patterns. It is a critical task in many domains, including finance, supply chain management, the environment, and more as accurate forecasts can help businesses and organizations make better decisions and improve their metrics. Although there have been significant advances in time series forecasting systems, thanks to the development of new machine learning algorithms, hardware improvements, and the increasing availability of data, it remains a challenging task. Common pitfalls, especially of single-model approaches include susceptibility to noise and outliers and inability to handle non-stationary data, which can lead to inaccurate and non-robust forecasts. Model-combining approaches, such as averaging the results of multiple predictors to produce a final forecast, are commonly used to mitigate such issues. This work introduces a novel application of Cascade Generalization or Cascading for time series forecasting, where multiple predictors are used sequentially, with each predictor’s output serving as additional input for the next. This methodology aims to overcome the limitations of single-model forecasts and traditional ensembles by incorporating a progressive learning mechanism. We adapt Cascade Generalization specifically for time series data, detailing its implementation and potential for handling complex, dynamic datasets. Our approach was systematically evaluated against traditional two-model averaging ensembles across ten diverse datasets, employing the Root Mean Square Error (RMSE) metric for performance assessment. The results revealed that cascading tends to outperform voting ensembles in most cases. This consistent trend suggests that cascading can be considered a reliable alternative to voting ensembles, showcasing its potential as an effective strategy for improving time series forecasting across a wide range of scenarios.