The Impact of Outlier Detection Algorithms on Statistical, Machine Learning and Deep Learning Forecast Models

Abstract Understanding and eliminating the impact of outliers on statistical, machine learning, and deep learning forecast models is the primary objective of this paper. Steel, iron ore, and coal price series are used as inputs for both outlier detection and forecasting models. The underlying processes behind the outliers in the data set are mainly two disastrous events for humanity: The Covid-19 pandemic and the Russian-Ukrainian war. Outlier detection is done with one-class support vector machine (SVM), local outlier factor (LOF), isolation forest (iForest), and minimum covariance determinant (MCD) algorithms. The iForest proved superior in the determination of post-pandemic growth and Ukrainian war periods as outlier ensembles. After the elimination of outlier ensembles, statistical autoregressive integrated moving average (ARIMA) and ARIMA with the transfer function (ARIMA-TF), nonlinear autoregressive neural network (NAR), and long short-term memory (LSTM) models are used to forecast steel prices. The LSTM method proved superior with regard to all forecast accuracy measures; and also showed the highest improvement by a 25.5 % reduction in root mean square error (RMSE) score after outlier ensembles are deleted. Deleting outlier ensembles may be generalized and applied to other cyclical series where cycle length and magnitude are already not constant over time.