anuragpamuru / la_tanuki

Data Hack 2021 Report Economics Track II

Team: Tanuki Members: Takashi Yabuta, Anurag Pamuru, Mizuki Kadowaki, Laura Diao

INTRODUCTION

(a. Introduction to data) Our dataset contains score values and rankings for 9 different societal pillars from 149 different countries from the years 2007-2014. Each pillar has multiple subcategories that contributed to the category ranking and score. The societal pillars are as follows:

• Business
• Economy
• Education
• Natural Environment
• Governance
• Health
• Personal Freedom
• Safety and Security
• Social Capital

DATA

Due to contract terms, we were not granted access to the results from the Gallup World Poll. Because of this, we removed variables that we did not have access to and only kept datapoints which had values for all variables.

In order to accomplish the task of finding which countries saw the highest prosperity growth over the period from 2007 to 2014, we calculated the Prosperity Score as the arithmetic mean of the scores for all pillars for each individual Country. We then found the average growth/change in this metric over the 8 years for which we had data. These table manipulations gave us access to the names of the top 5 countries in terms of "Average Prosperity Growth Per Year", which are seen below.

OBSERVATIONS

Country	Average Prosperity Growth Per Year	Most Important Pillar
Togo	0.955408	Safety and Security
Zimbabwe	0.904582	Governance
Chad	0.820001	Education
Georgia	0.800023	Health
Macedonia	0.745948	Social Capital

We calculated the most important pillar for each nation by seeing which pillar score was furthest in absolute distance from the Prosperity Score each year. Then, we found which pillar was most frequently furthest from the prosperity across all 8 years, which gives us the "Most Important Pillar" in determining the Prosperity Growth.

After doing so, we then trained a linear regression model on all the categories of each pillar and then peek into the models to find the largest linear regression coefficients of each pillar, which should ideally be correlated with the most important categories in that pillar. Here they are, listed with their coefficients.

Most Impactful Categories for Business	Coefficient
ippr	2.4785
logis	2.040
affs	2.024

Most Impactful Categories for Economy	Coefficient
gdp5	18.610
eqi	5.717
amp	2.674

GDP is the greatest determiner of Economy.

Most Impactful Categories for Education	Coefficient
qslnl	1.714
Ginih15	-13.976
gbrata	-17.244

Most Impactful Categories for Environment	Coefficient
pestreg	0.375
improved_drinkwater	0.224
airpollution	-18.581

Air pollution is the greatest determiner of Environmental Health.

Most Impactful Categories for Governance	Coefficient
womenparl	7.635
rlaw	2.632
geff	2.537

More women in Parliament correlates to a stronger government!

Most Impactful Categories for Health	Coefficient
leb	0.253
tblnl	-0.258
meandiabetes	-0.444

Diabetes prevalence is the greatest determiner of national health

Most Impactful Categories for Personal Freedom	Coefficient
propright	6.759
lgbt_rights	4.285
deathpen	-7.695

The Death penalty/Capital Punishment is the greatest determiner of a nation's beliefs on personal freedom

Most Impactful Categories for Safety & Security	Coefficient
warcasual	-1.599
pts	-3.069
homilnl	-3.507

Homicide/murder is the best determiner of a nation's safety standards.

Most Impactful Categories for Social Capital	Coefficient
vtf	0.141

DATA VISUALIZATION

• Here, on the right, we have a regional analysis of changes in global prosperity over the 8 year period.
  • It is important to note that low growth can correlate to both economically stagnated countries or highly developed countries
    • For example, Australia and Afghanistan could both have similar levels of prosperity growth because nothing changed dramatically in both countries. However, both nations obviously have drastically different standards of living.
• Meanwhile, on the left we have a ranking of which continent was the global leader in prosperity growth each year.
  • There is no clear winner here; however, Africa seems to be the most consistent leader in growth.

MACHINE LEARNING METHODS

Methods

For the input to our machine learning models, we used the 50 features that were identified as important in the previous section.

Data Preparation

• Principal Component Analysis (PCA)

Since there are 50 features while there are only about 780 data points, we reduced the number of features so that the models can learn parameters.

Machine Learning Models

• Linear Regression
• Random Forest Regression
• Vector Autoregression

We used Linear Regression for our baseline model, and then we also trained Random Forest Regression to improve the performance.

In fact, the regression models performed decently well, but they failed to recognize trends. So, we also trained Vector Autoregression, which is capable of recognizing trends.

Hyperparameter tuning and PCA

After our innitial models created through Linear Regression and Random Forest Regressor, we used PCA to change the dimensionality of our data to see whether this would improve the accuracy. We can see that increasing the number of principle components improved the performance of our Linear Regression drastically, but did not have much of an effect on the Random Forest model. The highest accuracy can be observed on the Random Forest model run on 17 components.

Predictions

After comparing the performances of machine learning models, it turned out that Vector Autoregression model performed the best based on our observations from the plots as shown below.

Here are the results of our predictions given by the Vector Autoregression model.

Pillar Value Predictions for 2015

country	busi	rank_busi	econ	rank_econ	educ	rank_educ	envi	rank_envi	gove	rank_gove	heal	rank_heal	pers	rank_pers	safe	rank_safe	soci	rank_soci
Georgia	52.9735	69	59.9654	87	59.1949	63	53.4916	120	54.0779	50	65.9174	102	56.0263	77	66.28	72	45.3822	100
Macedonia	55.536	60	59.253	92	66.4805	28	64.9589	54	51.2761	60	72.848	64	55.64	78	76.0674	34	45.0173	104
Chad	39.7406	134	45.7739	138	21.5546	146	55.9769	104	25.3702	143	47.9827	144	36.2053	132	56.4767	122	39.8931	136
Togo	45.6765	117	50.8053	120	32.023	135	58.3408	92	34.2592	129	54.1159	137	54.1778	82	61.3347	99	39.3812	139
Zimbabwe	48.1731	103	60.0522	85	47.0835	107	53.439	121	36.7532	118	68.2035	96	41.5871	118	61.4612	98	46.5467	90

Pillar Value Predictions for 2016

country	busi	rank_busi	econ	rank_econ	educ	rank_educ	envi	rank_envi	gove	rank_gove	heal	rank_heal	pers	rank_pers	safe	rank_safe	soci	rank_soci
Georgia	50.5652	90	61.4895	84	57.6592	72	51.3529	120	52.6507	61	63.3463	106	57.1895	75	66.8425	74	46.5718	95
Macedonia	53.5217	71	59.4928	98	65.5915	34	57.7827	93	49.1805	73	70.7027	79	56.6112	77	74.8046	36	43.1385	124
Chad	42.8513	119	45.6832	139	22.6597	145	57.1193	98	25.9784	141	49.6053	143	36.2493	128	61.3478	98	40.4285	133
Togo	45.6366	112	50.5939	122	29.5387	139	56.0203	102	30.512	131	54.501	135	45.9548	108	67.8589	67	41.2209	131
Zimbabwe	42.1331	122	66.2768	56	53.7926	82	55.3031	109	42.7921	93	73.1755	65	48.9734	98	67.7488	68	57.6737	32

Prosperity Predictions for 2015

country	proseprity	rank_proseprity
Georgia	57.0344	80
Macedonia	60.7864	52
Chad	40.9971	142
Togo	47.7905	126
Zimbabwe	51.4777	113

Prosperity Predictions for 2016

country	proseprity	rank_proseprity
Georgia	56.4075	80
Macedonia	58.9807	66
Chad	42.4359	140
Togo	46.8708	125
Zimbabwe	56.4299	79

CONCLUSION

Our prediction task was to create a model that determines for the years 2015 and 2016, each country’s pillar scores and ranks, as well as each country’s overall prosperity scores and ranks. To accomplish this we tested out various models such as Linear regression, Random Forest regression, and Vector Autoregression models. Given 50 features and 780 data points, we implemented Principal Component Analysis to prepare the data and reduce the features going into the models we tested. Despite the vast improvements in the linear regression model after editing the number of principal components, we saw the highest accuracy in the random forest model with fewer features.

Training and graphing the results of the Vector Autoregression model and random forest model, it turned out that in comparison, the Vector Autoregression model had the best performance. This performance may be attributed to how the Vector Autoregression models can recognize trends, handle time series data, and implicitly assume that the future predictions will follow after the past.