Neural network prediction of crude oil futures using B-splines

Highlights: • We propose two approaches to improve forecasting accuracy of FTDNN on the term structure of crude oil futures. • The convergence based FTDNN makes consistently more accurate predictions. • Using B-spline to fit the term structure curves leads to consistently better one and three months ahead predictions. • Under some conditions the B-spline based approach is better for longer term predictions. We propose two ways to improve the forecasting accuracy of a focused time-delay neural network (FTDNN) that forecasts the term structure of crude oil futures. Our results show that a convergence based FTDNN makes consistently more accurate predictions than the fixed-epoch FTDNN in Barunik and Malinska (2016). Further, we suggest using basis splines (B-splines), instead of Nelson-Siegel functions, to fit the term structure curves. The empirical results show that the B-spline expansions lead to consistently better 1 and 3 months ahead predictions compared to the convergence based FTDNN. We also explore conditions under which the B-spline based approach may be better for longer-term predictions.