Neighbor Weighting and Distance Metrics in Nearest Neighbor Nowcasting of Swedish GDP

Forecasting and nowcasting of economic activity can be of great importance in many settings. Business tendency survey data, when employed in a nearest neighbor (NN) algorithm, can produce nowcasts of Swedish GDP that compare well, in terms of predictive performance, to the often-used linear indicator models. The current article probes deeper into the choices available when implementing the nearest neighbor algorithm for nowcasting Swedish GDP and traces out the possible effects on nowcasting accuracy. The dimensions explored include the number of neighbors used for producing the nowcasts, the distance metric employed and distance-weighting of neighbors. The main results indicate the so-called Manhattan distance, or \(L_1\) $L_1$ norm, together with equal weighting of 4 or 5 neighbors, could improve nowcasting accuracy for Swedish GDP compared to a setting where a different number of neighbors is used, the \(L_2\) $L_2$ or \(L_\infty\) $L_{\infty }$ norms are employed and/or distance-based weighting of neighbors is applied.