Optimización matemática del rendimiento de un modelo de forecasting con suavizamiento exponencial simple

Abstract

In time series forecasting, the simple exponential smoothing technique allows one to forecast the next state based on past data. This model automatically assigns a weight to each instance with an exponential function, using a smoothing coefficient α, in the domain of ℝ∈[0,1]. Assigning an appropriate value of α is not a trivial task as a transfinite defines it, and generically, a coefficient is empirically selected to find a balance between computational complexity and model error. This scientific research develops a mathematical optimization process to improve the performance of a forecasting model in time series with simple exponential smoothing using traditional optimization techniques and models derived from Artificial Intelligence. Five optimization techniques and two trained machine learning algorithms were used to find an optimized α. The results were analyzed with twenty datasets of different natures and sizes, where the model error was analyzed using four metrics that were minimized. For each optimization technique, the convergence time of the algorithm was analyzed, to analyze the feasibility of implementation in a production environment. Regarding running time, the computational cost is 0.01 seconds for machine learning algorithms on datasets of any size. In contrast, a traditional optimization algorithm converged in less than a second, even for datasets from more than 100,000 instances.