The term structure of judgement: interpreting survey disagreement (2023), with Žymantas Budrys. Presentations: MacroFor Seminar Series (April 2023), IAAE 2023
Abstract: Consensus forecasts by professionals are highly accurate yet hide large heterogeneity. We build a framework to extract the judgement component from survey forecasts and analyse to what extent it contributes to respondents' disagreement. We find for the average respondent a sizable contribution of judgement about the current quarter, which often steers unconditional forecasts towards the realisation, improving accuracy. We identify the structural components of judgement exploiting stochastic volatility and give an economic interpretation to expected future shocks. For individual respondents, about one-third of disagreement is due to differences in coefficients or models used, and the remaining is coming from different assessments about future shocks; the latter mostly concerns the size of shocks, while there is a general agreement on their source.
Abstract: Forecasts produced by experts can influence the expectations of the general public, and ultimately the real economy. In this paper I ask what type of structural drivers do professional forecasters think will affect their projections? To what extent do they disagree about these drivers, and how uncertain are they about their magnitude? I model forecasts in a novel empirical macroeconomic setting, which allows me to decompose them into a model implied part and a judgement part, reflecting individual expectations of future shocks. The model takes into account multi-step ahead conditional forecasts, includes subjective uncertainty measured via different methods, and identifies shocks exploiting the time-varying volatility present in the forecasts. I find that throughout the sample, forecasters mostly disagree on the size of the shocks, while in periods of high volatility they give a larger weight to judgement and also disagree on the nature of shocks. My findings can inform policy makers by providing a deeper insight into the expectations formation process of forecasters from a structural perspective.
Macro-financial feedbacks through time (2022), with Ferre De Graeve and Raf Wouters. Presentations: International Monetary Fund, CEF 2022, IAAE 2022
Abstract: Changing (co-)variances of macroeconomic and financial series provide strong identification power in disentangling real-financial interactions. “Identification through heteroskedasticity" assumes changing (co-)variances stem only from changing structural shock-volatility. This paper generalizes the approach to encompass time-varying parameters. Imposing as constant either coefficients or shock volatilities does not reproduce real-financial (co-)variances for the US. The set of structural models that match the data contains both models with negative feedbacks and boom-bust theories. Alternative identification approaches unduly exclude plausible theories. The elasticity of financial to real variables increased around the 2000’s, while that of real to financial variables fell.
Combining Bayesian VAR and survey density forecasts: does it pay off? (2022), with Marta Bańbura, Joan Paredes and Francesco Ravazzolo (WP version). Presentations: ECB (2021), CEF 2021, IAAE 2021, RCEA Time Series Workshop, ISF 2021, ESOBE 2021, IIF 2021, ICMAIF 26, ESOBE 2022, Cleveland FED (2022)
Abstract: This paper studies how to combine real-time forecasts from a broad range of Bayesian vector autoregression (BVAR) specifications and survey (judgemental) forecasts by optimally exploiting their properties. To do that, we compare the forecasting performance of optimal pooling and tilting techniques, incorporating the survey information in various forms. We focus on predicting euro area inflation and GDP growth at medium-term forecast horizons and exploit the information from the ECB's Survey of Professional Forecasters (SPF). Results show that the SPF exhibits good point forecast performance but scores poorly in terms of densities for all variables and horizons. Accordingly, when individual models are tilted to the SPF's first moments and then optimally combined, point accuracy and calibration improve, whereas this is not always the case when the SPF's second moments are included in the tilting. Therefore, judgement incorporated in survey forecasts can considerably increase model forecast accuracy, however, the way and the extent to which it is incorporated matters. We demonstrate the usefulness of our analysis on a case study covering the COVID-19 pandemic period.