Real Time Trend-Cycle Prediction

Abstract

For real time trend-cycle analysis, official statistical agencies generally produce estimates derived from asymmetric moving average techniques. However, the use of the asymmetric filters introduces revisions as new observations are incorporated to the series and also delays in detecting true turning points. This chapter deals with a reproducing kernel approach to obtain time-varying asymmetric trend-cycle filters that converge fast and monotonically to the corresponding symmetric ones. It shows that the bandwidth parameter that minimizes the distance between the gain functions of the asymmetric and symmetric filters is to be preferred. Respect to the asymmetric filters incorporated in the X12ARIMA software currently applied by the majority of official statistical agencies, the new set of weights produces both smaller revisions and time delay to detect true turning points. The theoretical results are empirically corroborated with a set of leading, coincident, and lagging indicators of the US economy.