Essays in real-time forecasting

This thesis contains three essays in the field of real-time econometrics, and more particularly forecasting. The issue of using data as available in real-time to forecasters, policymakers or financial markets is an important one which has only recently been taken on board in the empirical literature. Data available and used in real-time are preliminary and differ from ex-post revised data, and given that data revisions may be quite substantial, the use of latest available instead of real-time can substantially affect empirical findings (see, among others, Croushore’s (2011) survey). Furthermore, as variables are released on different dates and with varying degrees of publication lags, in order not to disregard timely information, datasets are characterized by the so-called “ragged-edge”structure problem. Hence, special econometric frameworks, such as developed by Giannone, Reichlin and Small (2008) must be used. The first Chapter, “The impact of macroeconomic news on bond yields: (in)stabilities over time and relative importance”, studies the reaction of U.S. Treasury bond yields to real-time market-based news in the daily flow of macroeconomic releases which provide most of the relevant information on their fundamentals, i.e. the state of the economy and inflation. We find that yields react systematically to a set of news consisting of the soft data, which have very short publication lags, and the most timely hard data, with the employment report being the most important release. However, sub-samples evidence reveals that parameter instability in terms of absolute and relative size of yields response to news, as well as significance, is present. Especially, the often cited dominance to markets of the employment report has been evolving over time, as the size of the yields reaction to it was steadily increasing. Moreover, over the recent crisis period there has been an overall switch in the relative importance of soft and hard data compared to the pre-crisis period, with the latter becoming more important even if less timely, and the scope of hard data to which markets react has increased and is more balanced as less concentrated on the employment report. Markets have become more reactive to news over the recent crisis period, particularly to hard data. This is a consequence of the fact that in periods of high uncertainty (bad state), markets starve for information and attach a higher value to the marginal information content of these news releases. The second and third Chapters focus on the real-time ability of models to now-and-forecast in a data-rich environment. It uses an econometric framework, that can deal with large panels that have a “ragged-edge”structure, and to evaluate the models in real-time, we constructed a database of vintages for US variables reproducing the exact information that was available to a real-time forecaster. The second Chapter, “Real-time nowcasting of GDP: a factor model versus professional forecasters”, performs a fully real-time nowcasting (forecasting) exercise of US real GDP growth using Giannone, Reichlin and Smalls (2008), henceforth (GRS), dynamic factor model (DFM) framework which enables to handle large unbalanced datasets as available in real-time. We track the daily evolution throughout the current and next quarter of the model nowcasting performance. Similarly to GRS’s pseudo real-time results, we find that the precision of the nowcasts increases with information releases. Moreover, the Survey of Professional Forecasters does not carry additional information with respect to the model, suggesting that the often cited superiority of the former, attributable to judgment, is weak over our sample. As one moves forward along the real-time data flow, the continuous updating of the model provides a more precise estimate of current quarter GDP growth and the Survey of Professional Forecasters becomes stale. These results are robust to the recent recession period. The last Chapter, “Real-time forecasting in a data-rich environment”, evaluates the ability of different models, to forecast key real and nominal U.S. monthly macroeconomic variables in a data-rich environment and from the perspective of a real-time forecaster. Among the approaches used to forecast in a data-rich environment, we use pooling of bi-variate forecasts which is an indirect way to exploit large cross-section and the directly pooling of information using a high-dimensional model (DFM and Bayesian VAR). Furthermore forecasts combination schemes are used, to overcome the choice of model specification faced by the practitioner (e.g. which criteria to use to select the parametrization of the model), as we seek for evidence regarding the performance of a model that is robust across specifications/ combination schemes. Our findings show that predictability of the real variables is confined over the recent recession/crisis period. This in line with the findings of D’Agostino and Giannone (2012) over an earlier period, that gains in relative performance of models using large datasets over univariate models are driven by downturn periods which are characterized by higher comovements. These results are robust to the combination schemes or models used. A point worth mentioning is that for nowcasting GDP exploiting crosssectional information along the real-time data flow also helps over the end of the great moderation period. Since this is a quarterly aggregate proxying the state of the economy, monthly variables carry information content for GDP. But similarly to the findings for the monthly variables, predictability, as measured by the gains relative to the naive random walk model, is higher during crisis/recession period than during tranquil times. Regarding inflation, results are stable across time, but predictability is mainly found at nowcasting and forecasting one-month ahead, with the BVAR standing out at nowcasting. The results show that the forecasting gains at these short horizons stem mainly from exploiting timely information. The results also show that direct pooling of information using a high dimensional model (DFM or BVAR) which takes into account the cross-correlation between the variables and efficiently deals with the “ragged-edge”structure of the dataset, yields more accurate forecasts than the indirect pooling of bi-variate forecasts/models.