必博娱乐

主题：Multi‐step forecasting with high‐dimensional time series data – a new approach

时间：6月8日 13:30-15:30

地点：浦东校区6306

主要内容：Prediction is of central importance in the analysis of time series data. Univariate ARMA models andvector autoregressions provide a cornerstone for methods that can be used to predict the variable ofinterest any number of periods ahead. A recent thrust of the forecasting literature exploits large crosssectional information to formulate more precise forecasts for a single series of interest. In this case,vector autoregressions require estimation of an infeasibly large number of parameters so the forecastermust either identify separate forecasting equations for each time horizon of interest, or make additionalassumptions on the structure of the high dimensional data such as a factor model. The former makesthe already formidable problem of high dimensional model selection more difficult as many modelsneed to be chosen ‐ one for each horizon. This approach also fails to take advantage of the fact thatforecasts at different time horizons are related – the forecasting problems are solved independently foreach horizon. The latter requires assumptions that are hard to verify and may not be a goodapproximation for some time series data. This paper bridges this gap by proposing a new class of timevarying parameter models with an ARMA type structure. The model is highly flexible on the crosssectionaldependence of the data accommodating highly dependent data such as a factor model orpredictors that are independent of each other. At the same time, it is more parsimonious in the timeseries dimension resulting in the simple structure and convenience of ARMA type models for multisteppredictions. Simulation evidence shows that model works very well. Long run forecasts leverage highdimensional data to better identify long‐run and transitory components to macro data.