Transcript Négociations intégratives

ECIG 2007
Modeling www time series

The research opportunity

A word on time series models



Data
Models
Results


What have we learned?
Next steps
 2007
Stéphane Gauvin
FSA - ULaval
Research opportunity

CSR: Organizations manage a widening set of stakeholders



The digital sphere has become the Übermedia



Voices are innumerable
Which voice will become dominant? (eg: anti-smoking, fat lawsuits, vegetarianism)
General question is:


Power of exit
Power of voice
Can we measure and forecast real-world opinions merely by listening to the digital sphere?
Today’s question is:

How strong is the signal in the digital sphere?
 2007
Stéphane Gauvin
FSA - ULaval
A word on timeseries models

Marketing is concerned with theory building


Data mining is atheoretical
Trends are as a nuisance


First step is to take first and second differences
VAR and/or co-integration


Dekimpe & Hanssens IJRM 2000, WP 2006
Franses JMR 2005
 2007
Stéphane Gauvin
FSA - ULaval
Into the looking glass

The digital sphere is invisible. It is queried (googled)

We all google all the time to retrieve specific instances

Swammer searches to count instances
 2007
Stéphane Gauvin
FSA - ULaval
Swammer

Build an intelligent set of queries to compute index

Shown to be close to survey data
 2007
Stéphane Gauvin
FSA - ULaval
Illustrative data
 2007
Stéphane Gauvin
FSA - ULaval
Robust or else
 2007
Stéphane Gauvin
FSA - ULaval
Storms obscure trends
 2007
Stéphane Gauvin
FSA - ULaval
French presidental
 2007
Stéphane Gauvin
FSA - ULaval
Royal / Sarkozy
 2007
Stéphane Gauvin
FSA - ULaval
Industry data
 2007
Stéphane Gauvin
FSA - ULaval
Models

Parametric trend models

Robust estimator (M-reg)
 2007
Stéphane Gauvin
FSA - ULaval
SSA

Singular Spectrum Analysis (SSA) (Golyandina et al. 2000)

Non parametric applications to the digital sphere
 Bagchi & Mukhopadhyay (2006) (overall growth of the Internet)
 Papagiannaki et al. (2005) (overall backbone traffic)

SSA applications
 Ghil et al. (2002) (climatology)
 Balazs & Chaloupka (2004) (biology)
 Koelle & Pascual (2004) (epidemiology)
 Antoniou et al. (2003) (wavelet model / Internet traffic)
 Edwards (2006) (dissertation / US Navy related series)
 2007
Stéphane Gauvin
FSA - ULaval
Caterpillar-SSA
It is based on the idea of time series embedding into finitedimensional space and following application of singular value
decomposition (SVD) to the trajectory matrix (that is the result of
time series embedding). The components of SVD are uniquely
juxtaposed to the additive components of the original time series.
Thereby we obtain the decomposition of the time series into
additive components together with the information about them.
This information is represented by the collection of singular vectors
and signular values of the SVD.
 2007
Stéphane Gauvin
FSA - ULaval
Caterpillar-SSA

Opérationnellement
1. Construire une matrice de vecteurs décalés (dim L/2)
2. Extraire les valeurs propres
3. Regrouper les eigen-vecteurs en trois groupes
1. Tendance (auto-corrélations varient lentement)
2. Cycles (auto-corrélations varient rapidement)
3. Bruit (cycles de fréquence arbitraire)
 2007
Stéphane Gauvin
FSA - ULaval
Caterpillar-SSA
 2007
Stéphane Gauvin
FSA - ULaval
Results - presidential
 2007
Stéphane Gauvin
FSA - ULaval
Results - presidental
 2007
Stéphane Gauvin
FSA - ULaval
Results - Industry
 2007
Stéphane Gauvin
FSA - ULaval
Results - Industry
 2007
Stéphane Gauvin
FSA - ULaval
Results - Industry
 2007
Stéphane Gauvin
FSA - ULaval
Conclusions

Good signal-to-noise ratio

Estimation must be robust

SSA
 Trend is easily extracted and follows closely the original series
 Not robust to extreme values

M-NL
 Dominant technique for large scale scenario
 Sometimes, sensitive to seed values
 2007
Stéphane Gauvin
FSA - ULaval
Next

Build a tracking system
 M-NL to signal shifts
 autoSSA to produce rich trend summaries

Explore forecasting models
 Fitting and forecasting are not the same
 Longer series to test rolling holdout samples

Validity issues
 Anecdotal evidence of close tracking
 Presidential series raises questions as to what the signal means
 2007
Stéphane Gauvin
FSA - ULaval