Big data and macroeconomic nowcasting from data access to modelling
2 likes557 views
El documento presenta una investigación sobre el uso de grandes datos para la previsión macroeconómica en el ámbito de la Eurostat, discutiendo su aplicación, beneficios y métodos de modelado. Se desarrollan recomendaciones para el uso efectivo de grandes datos en ejercicios de 'nowcasting', abarcando desde la evaluación de calidad hasta la preparación y modelado de datos. Un estudio de caso muestra que, aunque los resultados son mixtos, algunas combinaciones de tendencias de Google y técnicas econométricas sofisticadas pueden mejorar las previsiones, aunque los beneficios son generalmente pequeños.
Big data and macroeconomic nowcasting from data access to modelling
- 1. Session 6: Administrative and Alternative Data Sources Big Data and Macroeconomic Nowcasting: From Data Access to Modelling 15th Conference of International Association for Official Statistics Abu Dhabi, 6–8 December 2016 Dario Buono*, European Commission, Eurostat, dario.buono@ec.europa.eu Stephan Krische*, GOPA Consultants, stephan.krische@gopa.de Massimiliano Marcellino, Bocconi University, massimiliano.marcellino@unibocconi.it George Kapetanios, King’s College, george.kapetanios@kcl.ac.uk Gian Luigi Mazzi, European Commission, Eurostat, gianluigi.mazzi@ec.europa.eu Fotis Papailias, Queen's University Management School, f.papailias@qub.ac.uk *The views expressed are the author’s alone and do not necessarily correspond to those of the corresponding organisations of affiliation
- 2. Eurostat, the Statistical Office of EU • About 700 people with 28 different nationalities • Statistical Office of European Union, part of EC • Core business: • Euro-zone (19) & EU (28) aggregates • harmonization, best practices, guidelines, trainings & international cooperation • Methodology team: Time Series, Econometrics, SDC, Research & EA
- 3. Why interested in Big Data for nowcasting? • Big Data are 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
- 4. European research project: Apr 15 to Jul 16
- 5. Research questions and findings Can Big Data help for Macroeconomic Nowcasting? What are the potential Big Data sources? 1. Literature review 2. Models/methods to be used for Big data 3. Recommendations on how to handle Big Data 4. Case study: IPI, Inflation, unemployment of some EU countries
- 6. Big Data types & dimensionality • When the dimensionality increases, the volume of the space increases so fast that the available data become sparse. • For statistically significant result, the amount of data needed often grows exponentially with the dimensionality. • Use of a typology based on Doornik and Hendry (2015): • Tall data: many observation, few variables • Fat data: many variables, few observations • Huge data: many variables, many observations
- 7. Eurostat Models race • Dynamic Factor Analysis • Partial Least Squares • Bayesian Regression • LASSO regression • U-Midas models • Model averaging 255 models tested, macro-financial & google trend data
- 8. Eurostat Statistical Methods: findings • Sparse regression (LASSO) works for fat, huge data • Data reduction techniques (PLS) helpful for large variables • (U)-MIDAS or bridge modelling for mixed frequency • Dimensionality reduction improves nowcasting • Forecast combination: Data-driven automated strategy with model rotation based on forecasting performance in the past works well
- 9. From Data Access to Modelling Step-by-step approach, accompanied by specific recommendations for the use of big data for macroeconomic nowcasting, guiding to • the identification and the choice of Big Data • pre-treatment and econometric modelling • the comparative evaluation of results to obtain a very useful tool for decision about the use or not of Big Data
- 10. Step 1: Big Data usefulness within a nowcasting exercise Recommendations 1. Evaluate the quality of the existing nowcasts and identify issue (bias or inefficiency or large errors in specific periods), that can be fixed by adding information in Big Data based indicators 2. Use of Big Data only when expecting to improve the timeliness and/or the quality of nowcastings 3. Do not consider Big Data sources with spurious correlations with the target variable
- 11. Step 2: Big Data search Recommendations 1. Starting point for an assessment of the potential benefits/costs of the use of Big Data for macroeconomic nowcasting: identification of their source • Social Networks (human-sourced information) • Traditional Business Systems (process-mediated data) • Internet of Things (machine-generated data) 2. Choice is heavily dependent on the target indicator of the nowcasting exercise
- 12. Step 3: Assessment of big-data accessibility and quality Recommendations 1. Privilege data providers with guarantee of continuity and of the availability of a good metadata associated to the Big Data 2. Privilege Big Data sources ensuring sufficient time and cross- sectional coverage 3. If a bias is observed a bias correction can be included in the nowcasting strategy. 4. To deal with possible instabilities of the relationships between the Big Data and the target variables, nowcasting models should be re- specified on a regular basis (e.g. yearly) and occasionally in the presence of unexpected events.
- 13. Step 4: Big data preparation Recommendations 1. Big data often unstructured: proper mapping 2. Pre-treatment to remove deterministic patterns • Outliers, calendar effects, missing observations • Seasonal and non-seasonal short-term movements should be dealt accordingly to the characteristic of the target variable 3. Create a specific IT environment where the original data are collected and stored with associated routines 4. Ensure the availability of an exhaustive documentation of the Big Data conversion process
- 14. Step 5: Big Data modelling strategy Recommendations 1. Identification of appropriate econometric techniques 2. First dimension: choice between the use of methods suited for large but not huge datasets, therefore applied to summaries of the Big Data (Google Trends) • nowcasting with large datasets can be based on factor models, large BVARs, or shrinkage regressions 3. Huge datasets can be handled by sparse principal components, linear models combined with heuristic optimization, or a variety of machine learning methods such as LASSO & LARS regression 4. In case of mixed frequency data, methods such as UMIDAS and, as a second best, Bridge, should be privileged.
- 15. Step 6: Results evaluation of Big Data based nowcasting Recommendations 1. Run a critical and comprehensive assessment of the contribution of Big Data for nowcasting the indicator of interest based, e.g., on standard criteria such as MSE or MAE. 2. In order to reduce the extent of data and model snooping, a cross- validation approach should be followed: • various models and indicators, with and without Big Data, estimated over a first sample and selected and/or pooled according to their performance • then the performance of the preferred approaches re-evaluated over a second sample
- 16. Case study - Implementation of all these steps for nowcasting IP growth, inflation and unemployment in several EU countries in a pseudo out of sample context, using Google trends for specific and carefully selected keywords for each country and variable - Big Data specific features: transform unstructured into structured data, time series decompositions, handling mixed frequency data - Overall, the results are mixed but there are several cases where Google trends, when combined with rather sophisticated econometric techniques, yield forecasting gains, though generally small. - Gains in term of timeliness or revisions have not been considered
- 17. Eurostat Literature contribution Eurostat Statistical Working Paper "Big Data and Macroeconomic Nowcasting: From data access to modelling" Methodological finding will be included in 2 chapter of the Eurostat/UNECE Handbook on Rapid Estimates currently under 2nd peer review, (forthcoming in 2017)
- 18. What's next? Big Data Econometrics • 2017, a new project focusing on: • Econometrics, Filtering issues, advanced Bayesian estimation and forecasting methods • Real time empirical evaluations (including a direct comparison with Eurostat flash estimates), • New ways and new metrics to present nowcasts • Possible data timeliness/accuracy gains • Big data handling tool developed as R package • Scientific summary for Big Data Econometric strategy
- 19. Thank you for your attention!! Some References: - Eurostat, Big data and macroeconomic nowcasting, preliminary results presented at the ESS methodological working group (7 April 2016, Luxembourg) http://ec.europa.eu/eurostat/cros/content/item21bigdataandmacroeconomicnowcastingsl ides_en - Big data CROS portal, http://ec.europa.eu/eurostat/cros/content/big-data_en - Marcellino, M. (2016), “Nowcasting with Big Data”, Keynote Speech at the 33rd CIRET conference. - Harford, T. (2014, April). Big data: Are we making a big mistake? Financial Times. Available at http://www.ft.com/cms/s/2/21a6e7d8-b479-11e3-a09a-00144feabdc0.html #ixzz2xcdlP1zZ - Lazer, D., Kennedy, R., King, G., Vespignani, A. (2014). "The Parable of Google Flu: Traps in Big Data Analysis", Science, 143, 1203-1205. - Tibshirani, R. (1996). “Regression Shrinkage and Selection via the Lasso”, Journal of the Royal Statistical Society B, 58, 267-288.