Forecasting skyrocketing unemployment with big data Mara Rosala - - PowerPoint PPT Presentation

New Techniques and Technologies for Statistics NTTS 2015 10-12 March 2015, Brussels

Forecasting skyrocketing unemployment with big data

María Rosalía Vicente (mrosalia@uniovi.es) Ana Jesús López (anaj@uniovi.es) Rigoberto Pérez (rigo@uniovi.es) University of Oviedo (Spain)

1.5e+006 2e+006 2.5e+006 3e+006 3.5e+006 4e+006 4.5e+006 5e+006 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013

Monthly evolution of registered unemployment in Spain Source: Spanish Ministry of Employment and Social Security (2014)

Unemployment rates. Year 2014

EU-28= 10.2% Spain= 24.5%

Source: Eurostat (2015)

BACKGROUND

• Two main references: Ettredge, Gerdes and Karuga (2005) and

Choi and Varian (2009).

• Evidence has been provided for different countries: France

(Fondeur and Karamé, 2013), Germany (Askitas and Zimmermann, 2009), Israel (Suhoy, 2009), Italy (D’Amuri, 2009), Norway (Anvik and Gjelstad, 2010), the UK (McLaren and Shanbhogue, 2011) and the US (D'Amuri and Marcucci, 2009). Literature on nowcasting and forecasting unemployment with

• nline search-related data:

DATA

Explanatory variables:

• On the demand side: The Employment Confidence Indicator (ECI) which shows

the balance between the positive and negative opinions of industrial firms on the current employment situation and their perspectives three-months ahead.

• Source: Spanish Ministry of Industry, Energy and Tourism.
• On the supply side: Google’s Trend Index which measures the volume of queries

made by internet users through this search engine.

• Note: This is a weekly index that takes value 100 in the week with the

highest number of searches for the words of interest.

• Keywords: “oferta de trabajo” and “oferta de empleo” (=job offer).
• Source: Google Trends service.

Variable of interest: Monthly registered unemployment in Spain.

• Source: Spanish Ministry of Employment and Social Security.
• Period of analysis: January 2004-December 2012.
• Forecasting horizon: January 2013-December 2013.

METHODOLOGY

Baseline B1: ARIMA(0,1,2)(0,1,1) (1-L)(1-L12)Yt=(1-q1L-q2L2)(1-Q1L12)ut Baseline B2: ARIMA(0,1,2)(0,1,1) with a level shift (LS) starting in March 2008 and a level shift with trend (t LS) (1-L)(1-L12)Yt= (1-q1L-q2L2)(1-Q1L12)ut + g1LSt

+ g2 t LSt

Model M1: (1-L)(1-L12)Yt= (1-q1L-q2L2)(1-Q1L12)ut + g1LSt

+ g2 t LSt + b1Xt ECI

Model M2: (1-L)(1-L12)Yt= (1-q1L-q2L2)(1-Q1L12)ut + g2 t LSt + b1Xt

ECI + b2 Xt Google-T

Model M3: (1-L)(1-L12)Yt= (1-q1L-q2L2)(1-Q1L12)ut + g2 t LSt + b1Xt

ECI + b3 Xt Google-E

Two baselines models: Three specifications including Google-related variables on job search:

Baseline B1 Baseline B2 Model M1 Model M2 Model M3

q1 0.7853 *** 0.7603 *** 0.7422 *** 0.6858 *** 0.6863 *** q2 0.4055 *** 0.4006 *** 0.3888 *** 0.3763 *** 0.3766 *** Q1

• 0.4618 ***
• 0.5526 ***
• 0.5339 ***
• 0.6607 ***
• 0.6555 ***

g1 (Level shift) 58439.3 ** g2 (Level shift with trend)

• 751.266 **
• 258.788 **
• 339.137 ***
• 304.633 ***

b1 (Employment Confidence Indicator)

• 1206.42 ***
• 704.939 *
• 785.996 *

b2 (Google index for “oferta de empleo”) 304.563 ** b3 (Google index for “oferta de trabajo”) 308.017 * S.D. of innovations 33237.26 33043.72 32428.39 31212.74 31598.58 Akaike Criterion 2259.380 2258.660 2255.088 2249.829 2252.163 Schwarz Criterion 2269.595 2273.983 2270.412 2267.706 2270.040 Normality test Chi- square Chi-2=2.57 p=0.27 Chi-2=1.79 p=0.40 Chi-2=1.34 p=0.51 Chi-2=2.41 p=0.30 Chi-2=2.49 p=0.29

Estimation results for ARIMA and ARIMAX models on Spanish unemployment

4.6e+006 4.65e+006 4.7e+006 4.75e+006 4.8e+006 4.85e+006 4.9e+006 4.95e+006 5e+006 5.05e+006 2013 Unemployment Unemployment_Baseline_forecast Unemployment_M1_forecast Unemployment_M2_forecast Unemployment_M3_forecast

Actual and forecasted unemployment in the horizon January-December 2013

Baseline B1 Baseline B2 Model M1 Model M2 Model M3 Root Mean Squared Error 219440 64065 67653 61639 59056 Mean Percentage Error

• 3.3073

1.2408 0.1319 0.8527 0.6042 Mean Absolute Percentage Error 3.5837 1.2408 1.1794 1.1678 1.075 Theil's U 3.2791 0.9023 0.9707 0.8678 0.8289

SUMMARY

• Emerging literature on the use of “Big Data” to improve the

nowcasting and forecasting of macroeconomic variables.

• This paper has focused on the data coming from individuals’

internet search behavior in order to analyze the evolution of unemployment in Spain.

• Searches on “job offers”.
• Results confirm the potential of the proposed approach: It

significantly improves the estimation and forecasting of unemployment’s figures in a context of important economic shocks.

More details in the paper: Vicente, M.R., López, A.J. and Pérez, R. (2015): Forecasting unemployment with internet search data: Does it help to improve predictions when job destruction is skyrocketing?, Technological Forecasting & Social Change, 92, 132-139.