Exploring Enhancements to the Stacked LSTM Outlier Detection and Correction Method

Existence of outliers in a dataset render it problematic for Time Series Forecasting. Hence various mathematical and statistical methodologies such as Omission and imputation, Z-Score Normalization, Winsorization, and Robust Regression, have been utilized but often fall short especially when working with datasets containing extensive outlier zones. A novel solution proposed and evaluated by the authors in a previous paper has shown promise especially when utilizing Long Short-Term Memory (LSTM)s, but further enhancements in the form of increasing training data size and variation, ensemble strategies, autoencoders, and smoothing techniques were suggested. This paper reports on the exploration of the effects of these enhancements on said solution.