A0175
Title: IRF and nowcasting with functional approaches and mixed-frequency data
Authors: Laurent Ferrara - SKEMA Business School (France)
Anna Simoni - CREST - ENSAE and CNRS (France)
Catherine Doz - Paris School of Economics (France) [presenting]
Abstract: Prediction and causal models often involve economic time series with different sampling frequencies. The mix-frequency problem has received a lot of attention in the past nowcasting/forecasting literature, which has proposed solutions that work especially well for prediction when the frequency gap is small, like the gap between monthly and quarterly data. An alternative approach is proposed to deal with mix-frequency which is particularly designed for situations where the gap in sampling frequencies is large, like daily and quarterly. Such a large gap can be easily found when one mixes series from standard and non-standard sources, like the internet or newspapers. The approach focuses on both prediction and causality, and exhibits excellent performance. By treating the high-frequency variable as a realization of a stochastic process in continuous time, the estimation problem is cast in the class of ill-posed inverse problems, and different regularized estimators are proposed. Importantly, for each high-frequency covariate, a measure of its influence is recovered on the target variable as a function of the time gap between the forecasting horizon and the date of the information in the past. Because the analysis is conditional on several covariates, the fact that the influence of high-frequency series rapidly decreases as soon as information from standard data arises is well captured.
