Nowcasting payrolls employment with traditional mediacontent

Bortoli, C.∗, Combes, S.∗ & Renault, T.∗∗

Institut National de la Statistique et des Etudes Economiques∗

Universite Paris 1 Pantheon-Sorbonne∗∗

IESEG School of Management∗∗

thomas.renault@univ-paris1.fr

March 16, 2016

Bortoli, C., Combes, S., & Renault, T. Nowcasting payrolls employment March 16, 2016 1 / 20

Non-Farm Payrolls Employment

Quarterly data - Flash estimate (45 days) - First estimate (70 days) -Revised estimate (160 days)

Synthesis of administrative sources (URSSAF, Pole emploi, DADS,etc.) and company surveys (ACEMO) on employment

Estimation of employment levels and changes in different businesssectors

Bortoli, C., Combes, S., & Renault, T. Nowcasting payrolls employment March 16, 2016 2 / 20

Non-Farm Payrolls Employment

Figure: Change in Non-Farm Payrolls Employment (Flash Estimate)

Bortoli, C., Combes, S., & Renault, T. Nowcasting payrolls employment March 16, 2016 3 / 20

Methodology

Textual Analysis [Loughran & McDonald, JoF, 2011]

Quantification of Media Content [Baker & al., NBER (2013)]

Media Pessimism and Unemployment Expectations [Garz, JEP(2013)]

Alternative Data and Unemployment [Fondeur & Karame, EM (2013)]

Bortoli, C., Combes, S., & Renault, T. Nowcasting payrolls employment March 16, 2016 4 / 20

Methodology

1 Derive quantitative indicators from textual content published bytraditional media

2 Include textual indicators into benchmarks model used to nowcastFrench employment

3 Compare Root Mean Square Forecast Error (RMSFE)

Bortoli, C., Combes, S., & Renault, T. Nowcasting payrolls employment March 16, 2016 5 / 20

Baker & al. [2013] methodology

Search Le Monde (from 1987) using Lexis Nexis and Le Figaro (from2002) using Factiva

Count the number of articles containing the following triple :

(E) ”economie”, ”economique”, ”economiques”(P) ”taxe”, ”impot”, ”depense”, ”deficit”, ”politique”, ”banque centrale”,”budget”... (20 keywords)(U) ”incertitude”, ”incertain”, ”incertitudes”, ”incertains”

Scale by the total number of articles in the same newspaper andmonth. Standardize and normalize to create an aggregate indicator

Bortoli, C., Combes, S., & Renault, T. Nowcasting payrolls employment March 16, 2016 6 / 20

Baker & al. [2013] methodology

Figure: Economic Policy Uncertainty Index - France

Bortoli, C., Combes, S., & Renault, T. Nowcasting payrolls employment March 16, 2016 7 / 20

Methodology

Extract all articles published on Le Monde website between 1990 and2016

Remove articles related to foreign countries

Identify articles related to the topic ”Economy” using machine learningmethods

Construct two field-specific lexicon related (a) to employment and (b)to global sentiment

Derive article tonality using a dictionary-based approach

Bortoli, C., Combes, S., & Renault, T. Nowcasting payrolls employment March 16, 2016 8 / 20

Methodology

Bortoli, C., Combes, S., & Renault, T. Nowcasting payrolls employment March 16, 2016 9 / 20

Data

1,297,292 articles, of which 212,721 have been classified as ”economy”by journalists from Le Monde (195,051 ”international”, 158,969”society”, 122,718 ”politics, 95,426 ”europe”, 85,649 ”sports”, 63,579”planet”...). Classification starts in 2005.

Naive Bayes classifier on a training dataset of 20,000 economic newsand 20,000 non-economic news

Named-entity recognition to remove articles not related to France

Final dataset of 202,674 articles related to the French economy

Bortoli, C., Combes, S., & Renault, T. Nowcasting payrolls employment March 16, 2016 10 / 20

Data

We first create a field-specific lexicon with positive (negative) bigramsand trigrams related to employment / company’s outlook :

Positive words list (Total 53 ngrams) : ”job creation”, ”hiring plan”,”higher profit”, ”increase activity”, ”staff increase”...Negative words list (Total 121 ngrams) : ”job destruction”, ”redundancyplan”, ”lower profit”, ”industrial restructuring”, ”judicial liquidation”...

Then, we create a lexicon with global positive (negative) words :

Positive words list (Total 485 ngrams) : ”improve”, ”favorable”,”surplus”, ”cooperation”, ”expand”, ”success”...Negative words list (Total 1,507 ngrams) : ”instability”, ”trouble”,”uncertainty”, ”weaken”, ”depress”, ”erode”...

Bortoli, C., Combes, S., & Renault, T. Nowcasting payrolls employment March 16, 2016 11 / 20

Data

For each article, we count the number of words in each lexicon, andwe define :

SENTIMENTi =# of Positive word − # of Negative word

Total word(1)

INDICATORi =

INDICATORi = 1 if SENTIMENTi > 0INDICATORi = 0 if SENTIMENTi = 0INDICATORi = −1 if SENTIMENTi < 0

(2)

Bortoli, C., Combes, S., & Renault, T. Nowcasting payrolls employment March 16, 2016 12 / 20

Data

Figure: Employment Media Indicator (EMI)

Bortoli, C., Combes, S., & Renault, T. Nowcasting payrolls employment March 16, 2016 13 / 20

Data

Figure: Global Media Indicator (GMI)

Bortoli, C., Combes, S., & Renault, T. Nowcasting payrolls employment March 16, 2016 14 / 20

Nowcasting French Non-Farm Payrolls Employment

The baseline model is a simple auto-regressive model. Different mediaindicators (X) are then added, and compared to the baseline model. Theemployment equation takes the form :

∆NFPt = α + β1∆NFPt−1 + β2∆NFPt−2 + ΦXt + εt (3)

We also consider an augmented baseline model by adding availableinformation from business climate surveys (monthly data, published at theend of the month for current month)

∆NFPt = α + β1∆NFPt−1 + β2∆NFPt−2 + β3∆Surveyt + ΦXt + εt (4)

Bortoli, C., Combes, S., & Renault, T. Nowcasting payrolls employment March 16, 2016 15 / 20

Nowcasting French Non-Farm Payrolls Employment

Table: Employment Regression Results [Month 2]

Baseline EMI GMI Survey Survey+GMI

[1] [2] [3] [4] [5]

α 0.0002 0.0012 0.0058 0.0002 0.0037

(0.000) (0.001) (0.001) (0.000) (0.001)

∆NFPt−1 0.5989*** 0.5336*** 0.4164*** 0.3832*** 0.3182***

(0.083) (0.102) (0.090 ) (0.071) (0.078 )

∆NFPt−2 0.2295*** 0.1988** 0.0509 0.4760*** 0.3038***

(0.087) (0.086) (0.100) (0.076) (0.085)

EMIt 1.7614*

(0.446)

GMIt 0.3927*** 0.2499***

(0.068) (0.073)

∆Surveyt 0.0383*** 0.0292***

(0.004) (0.005)

Adjusted R2 0.622 0.631 0.714 0.740 0.770

Bortoli, C., Combes, S., & Renault, T. Nowcasting payrolls employment March 16, 2016 16 / 20

Nowcasting French Non-Farm Payrolls Employment

Table: Root Mean Square Forecast Error [2000-2016]

Baseline EMI GMI Survey Survey+GMI

[1] [2] [3] [4] [5]

RMSFE [1 month] 0.2732 0.2737 0.2623 0.2162 0.2259

RMSFE [2 months] 0.2415 0.2471 0.2062 0.1876 0.1787

RMSFE [3 months] 0.2415 0.2392 0.2064 0.1751 0.1699

Bortoli, C., Combes, S., & Renault, T. Nowcasting payrolls employment March 16, 2016 17 / 20

Conclusion

Sentiment derived from articles published in traditional media seemsto help nowcasting French non-farm payrolls employment

When two months of data are available, both business climate andglobal media tonality are significant at the 1% level [Full Sample]

RMSFE is lower when a media indicator is added to anauto-regressive model augmented with business climate survey [Fittedon 1990-1999 and rolling forecast on 2000-2016]

Bortoli, C., Combes, S., & Renault, T. Nowcasting payrolls employment March 16, 2016 18 / 20

Directions for future research

Use more advanced Natural Language Processing techniques toimprove the detection and weighting of media sentiment

Extend time period (data from LeMonde are available since 1944).Could be especially useful for periods where other indicators are notavailable (for example, business climate is available since 1980)

Add more media to avoid bias related to the analysis of a uniquecontent provider

Analyze if media content helps nowcasting/forecasting othermacro-economics of financial variables (stock prices, market volatility,investment, GDP...)

Bortoli, C., Combes, S., & Renault, T. Nowcasting payrolls employment March 16, 2016 19 / 20

Questions

Thanks for your attention.

Bortoli, C., Combes, S., & Renault, T. Nowcasting payrolls employment March 16, 2016 20 / 20