Simplicity, robustness and speed in time series forecasting: LEAF-learning by exponentially adaptive forecasting, a novel univariate predictive analytics benchmark
提出一种名为LEAF的新时间序列预测方法，从简单基础出发，融入序列误差学习，实证和模拟均显示其在不同频率和参数下优于现有基准模型，适合作为评估其他预测方法的基准。
In this article, we present a novel forecast benchmark for time series, which we call learning exponential adaptive forecasting or LEAF. We examine its theoretical properties and showcase that this new forecasting method can be explored within the context of a model that begins from simple foundations, expands to incorporate the essential features of the mean of the time series and the naïve forecast, and overlays these with sequential error learning. Our empirical exercise demonstrates the efficacy of the proposed method, which delivers superior forecasting performance across different frequencies, time periods, and parameter combinations. Both the original and enhanced forms of the method consistently outperform the standard benchmark models that mostly appear in the literature. We re-confirm the efficacy of the method also via a simulation analysis that indicates the same performance characteristics, now in a controlled environment. The foundational simplicity, and the simplicity of computations, as well as the substantially good performance we obtain, suggest the potential of this new method in real-world applications and as a powerful benchmark to evaluate other forecasting approaches.