Title page
Contents
Abstract 4
Résumé 5
1. Introduction 8
2. Data 10
2.1. National average SWB 10
2.2. Google Trends time series 12
2.3. Macroeconomic control variables 14
3. Methods 15
3.1. Exploratory analysis 15
3.2. Training design for base learners 15
3.3. Meta-learning with an ensemble model 17
4. Results 19
4.1. Exploratory analysis 19
4.2. Training design for base learners 20
4.3. Meta-learning with an ensemble model 24
5. Conclusion 26
References 27
Table 1. Results of yearly regression models for the three types of pre-processing 19
Table 2. Summary of experimental results related to the design of base learners 21
Table 3. In-sample and out-of-sample performance of base learners and meta learner 25
Figure 1. Number of monthly observations per country 11
Figure 2. Time series of the average SWB in Germany 12
Figure 3. Time series of two unrelated categories of searches and dates of breaks in Google Trends reporting methods (dashed lines) 13
Figure 4. Z-scores and standard errors of all variables of interest for the final specification of the yearly regression model 20
Figure 5. Time series of true values versus out-of-sample predictions by 4 base learners - Averages over pooled sample of country-month observations 22
Figure 6. Out-of-sample predictions of country-month observations against true values for the 4 base-learners 23
Figure 7. MSE of out-of-sample predictions by country and by base learner 24
Figure 8. Out-of-sample MSE by country for the Meta learner 25