A novel range prediction model using gradient descent optimization and regression techniques

Abstract

Predictive models learn relationships between dependent and independent features of a dataset to forecast future outcomes. The point forecasting models aim to predict a single numeric value for future input. Most of the time, point prediction models lead to inaccurate predictions due to the scattering variance of data from the fitted curve. However, the decision-makers often choose a range of plausible estimates leading to range forecasting to overcome human cognitive bias and catastrophic forecasting errors. In this paper, we present a novel range prediction model that predicts a range of most probable outcomes for future input. We first fit a robust nonlinear regression model to the data. Then, we compute the slope of the tangent line at each data point on the nonlinear fitting curve. We introduce an angular confidence interval and use it to generate conical (angular) sectors at each data point on fitted curve. The conical sector produces an arbitrarily shaped range residual interval. Finally, we apply gradient descent optimization to the range residual sum of squares to get the optimum range prediction results. Experiments are performed on four publicly available data sets, and the results show the viability of our approach.