Big Data Econometrics: Nowcasting and Early Estimates ESTAT Project (01/2017 – 01/2018) Implemented by GOPA Consultants; Technical Expertise Provided by Massimiliano Marcellino (Bocconi University), George Kapetanios (King’s College), Fabio Bacchini (ISTAT), Fotis Papailias (King’s College), Katerina Petrova (University of St. Andrews )
Why interested in Big Data for nowcasting? Big Data can provide relevant complementary information to standard data, being based on different information sets More granular perspective on the indicator of interest, both in the temporal and cross-sectional dimensions It is timely available, generally not subject to revisions
Presentation Outline Big Data Types with Macro Nowcasting Potential Feature Extraction: from Big Data to Time Series Filtering techniques Examples by Type Reuters-based (Big Data) Uncertainty Indexes Google-Based Uncertainty Indexes Econometric Methodologies for Big Data Next Steps
Big Data Types with Macro Nowcasting Potential Overview Financial markets data Scanner prices data Electronic payments data Online prices data Mobile phones data Online search data Sensor data Textual data Satellite images data Social media data
Nowcasting Specific Macroeconomic Variables Using Big Data Unemployment GDP and components Inflation Surveys Financial variables
Types of Big Data by Dominant Dimension Fat Tall Huge Financial Markets X Electronic Payments Mobile Phones Sensor Data / IoT Satellite Images Scanner Prices Online Prices Online Search Fat (big cross-sectional dimension, N, small temporal dimension, T) Tall (small N, big T) Huge (big N, big T)
Uncertainty Indexes based on Big Data Feature Extraction: from Big Data to Time Series for Econometric Modelling General Data Conversion Framework for Un- structured Numerical Big Data Empirical Examples Uncertainty Indexes based on Big Data In general, the majority of the examples indicate that linear aggregation seems a suitable choice in numerical exercises, although this largely depends on the conceptual setting of the nowcasting exercise too.
Feature Extraction of Big Data to Time Series for Econometric Modelling Outliers?
Filtering techniques for high-frequency data Outliers Seasonal Patterns and Signal Extraction From Signal Extraction to Uncertainty Indexes STL filtering approach based on local linear regression, loess method (“Seasonal and Trend decomposition using Loess“) seems particularly suited – computationally very fast and can also handle high frequencies.
Feature Extraction: from Big Data to Time Series for Econometric Modelling Huge dimension? Subsampling in order to reduce the dimensionality problem
Feature Extraction: from Big Data to Time Series for Econometric Modelling Financial Markets Data
Feature Extraction: from Big Data to Time Series for Econometric Modelling Electronic Payments Data
Feature Extraction: from Big Data to Time Series for Econometric Modelling Electronic Payments Data
Feature Extraction: from Big Data to Time Series for Econometric Modelling Mobile Phones Data
Feature Extraction: from Big Data to Time Series for Econometric Modelling Sensor Data and IoT
Feature Extraction: from Big Data to Time Series for Econometric Modelling Satellite Images Data
Feature Extraction: from Big Data to Time Series for Econometric Modelling Online Prices Data
Feature Extraction: from Big Data to Time Series for Econometric Modelling Online Data
Feature Extraction: from Big Data to Time Series for Econometric Modelling Textual: Reuters-based Uncertainty Indexes Use Reuters online database (big data) to build uncertainty indexes based on keywords
Feature Extraction: from Big Data to Time Series for Econometric Modelling Textual: Reuters-based Uncertainty Indexes
Feature Extraction: from Big Data to Time Series for Econometric Modelling Textual: Reuters-based Uncertainty Indexes
Feature Extraction: from Big Data to Time Series for Econometric Modelling Textual: Reuters-based Uncertainty Indexes
Feature Extraction: from Big Data to Time Series for Econometric Modelling Textual: Reuters-based Uncertainty Indexes
Feature Extraction: from Big Data to Time Series for Econometric Modelling Online Search: Google-Based Uncertainty Indexes
Feature Extraction: from Big Data to Time Series for Econometric Modelling In general, the majority of the examples indicate that linear aggregation seems a suitable choice in numerical exercises, although this largely depends on the conceptual setting of the nowcasting exercise too.
Econometric Methodologies for Big Data with Potential for Nowcasting Penalised Regressions Ridge Lasso Elastic Net Spike and Slab Regressions Compressed Regressions
Econometric Methodologies for Big Data with Potential for Nowcasting Bayesian VARs Reduced Rank VARs Bayesian Reduced Rank VAR Time Varying Parameter VAR Stochastic Volatility VAR Nonparametric time varying Quantile and Expectile Regressions
Econometric Methodologies for Big Data with Potential for Nowcasting Regression Trees Boosting Cluster Analysis Random Forests Neural Networks / Deep Learning
Econometric Methodologies for Big Data with Potential for Nowcasting Factors PCA Sparse PCA PLS Forecast Combination Bayesian Model Averaging Frequentist Model Averaging Mixed-Frequency Models
So Far… Encouraging evidence for the use of Big Data for nowcasting and flash estimation Official Statistical Agencies should consider Big Data Improved flash estimates and possibly refinements of other statistics
Next Steps… Empirical test on possible timeliness gains New metrics for official statistics using econometric techniques ESTAT Big data handling tool