On trend estimation of time-series: a simple linear programming approach

Abstract

This paper presents a simple linear programme for the solution of the trend-estimation problem in time-series. We studied the trade-off between two important properties of the trend component: smoothness and fidelity (closeness to the data). The linear programming solution is a monotone sequence optimizing some (weighted) combination of both properties. A single coefficient in the objective function is user-dependent and represents the user’s preference with respect to the two properties. The method is illustrated on several empirical series.