Financial transaction prices, Autoregressive conditional multinomial model, GLARMA, Count data, Market microstructure effects, C22, C25, G10,

No abstract

We investigate price duration variance estimators that have long been neglected in the literature. In particular, we consider simple-to-construct non-parametric duration estimators, and parametric price duration estimators using autoregressive conditional duration specifications. This paper shows (i) how price duration estimators can be used for the estimation and forecasting of the integrated variance of an underlying semi-martingale price process and (ii) how they are affected by discrete and irregular spacing of observations, market microstructure noise, and finite price jumps. Specifically, we contribute to the literature by constructing the asymptotic theory for the non-parametric estimator with and without the presence of bid/ask spread and time discreteness. Further, we provide guidance about how our estimators can best be implemented in practice by appropriately selecting a threshold parameter that defines a price duration event, or by averaging over a range of non-parametric duration estimators. We also provide simulation and forecasting evidence that price duration estimators can extract relevant information from high-frequency data better and produce more accurate forecasts than competing realized volatility and option-implied variance estimators, when considered in isolation or as part of a forecasting combination setting.

a b s t r a c tWe use a cross-section of economic survey forecasts to predict the distribution of US macro variables in real time. This generalizes the existing literature, which uses disagreement (i.e., the cross-sectional variance of survey forecasts) to predict uncertainty (i.e., the conditional variance of future macroeconomic quantities). Our results show that cross-sectional information can be helpful for distribution forecasting, but this information needs to be modeled in a statistically efficient way in order to avoid overfitting. A simple one-parameter model which exploits time variation in the cross-section of survey point forecasts is found to perform well in practice.

We present an improved method for inference in linear regressions with overlapping observations. By aggregating the matrix of explanatory variables in a simple way, our method transforms the original regression into an equivalent representation in which the dependent variables are non-overlapping. This transformation removes that part of the autocorrelation in the error terms which is induced by the overlapping scheme.Our method can easily be applied within standard software packages since conventional inference procedures (OLS-, White-, Newey-West-standard errors) are asymptotically valid when applied to the transformed regression. Through Monte Carlo analysis we show that they perform better in finite samples than the methods applied to the original regression that are in common usage. We illustrate the significance of our method with two empirical applications.JEL classification: C20, G12