World Leader Tips
with Sam Lavigne
Otherwise, here’s my Original post on Medium.com:
December 2, 2014
It must be really hard to be a world leader. Constantly thinking about international trade negotiations, dealing with civil unrest and outbreaks of violence, struggling with the opposition party, hearing the complaints of environmental groups; it must be exhausting. I thought to myself, if I was in their shoes, what would I want to be doing as one of the most powerful people in the world? I would probably want to just relax and catch up on gossip or commiserate about the struggles of governing with my fellow world leaders. And to enable me to do this, wouldn’t it be nice to sit back, put my feet up, and rely on an unbiased and data-driven system for suggesting political actions to alleviate the burdens of running a country?
So for the past couple of weeks, I’ve been working on a project titled the “International Political Tele-Conferencing Equilibrium System” (IPTCES). IPTCES uses an automated process to suggest different government-level actions to political leaders around the world, with the ambitious, and possibly utopian goal of creating a scenario where all world leaders are in constant telephone communication with each other.
IPTCES uses data from GDELT, a project that modestly describes itself as a “the largest, most comprehensive, and highest resolution open database of human society ever created”. IPTCES queries the GDELT’s database, selecting all events where a given world leader is the primary actor, runs that data through machine-learning algorithms (similar to the algorithms used by hedge funds to analyze and predict the stock market fluctuations), predicts actions the world leader might take, and then suggests an alternative action. When the program detects that world leaders seem to be acting in a way that would support a tele-conferencing homeostasis singularity scenario, ITPCES celebrates their achievement and encourage their continued status by automatically suggesting they call one of their fellow world leaders.
As of this post, the project generates a weekly forecast for many world leaders in the form of a line chart depicting the trend of their action history for the past 200 days, and their projected actions for the coming week. At the bottom of the graphic is a suggested action for the week.
The next phase of the project will include a website where the suggestions will be posted.
How it Works:
IPTCES is based on a metric called the Goldstein Scale, which is an aspect of the data captured by the GDELT. It’s used to tag and score all political, military, and societal actions on a scale of -10.0 to +10.0. Where -10.0 is the worst of all geopolitcal action, for example: ‘engage in ethnic cleansing’, ‘use weapons of mass destruction’, or ‘use chemical or biological weapons’. Zero is neutral, with actions like: ‘explain policy’, ‘vote’, or ‘engage in symbolic act’. And on the other end of the spectrum, a +10.0 signifies ‘retreat or military surrender’. To give a sense of the scale, a 8.6 is ‘provide military protection or peacekeeping’, a 7.0 is ‘apologize’, a -3.7 is ‘halt negotiation’, and a -5.7 is to ‘threaten to impose restrictions on freedoms of speech and expression’. Finally, 1.0 on the Goldstein scale is ‘talk on the phone’. The number on the Goldstein Scale is used to determine the politician’s action history.
For many world leaders, the GDELT is queried to find all instances of their name being associated with an event, for Barack Obama (the president of the United States) the number of instances is 386,362, but for Joyce Banda (the president of Malawi) it’s 1,815. Despite this discrepancy, in most cases there was plenty of data to work with to generate a profile of their shifting Goldstein Score. However, some prominent figures are strangely absent from the GDELT data, such as Enrique Peña Nieto, the president of Mexico, or Park Geun-hye, the president of South Korea.
In order to find political leader’s daily average Goldstein Score, for every event associated with the day, I used the ‘number of mentions’ column in the data to ‘weight’ the scores with the number of mentions associated with them. This helped to lessen the impact of outliers.
Days when the GDELT has no records for the leader were forward filled based on their previous entry. The trend of the Goldstein Score was then determined by finding their rolling mean or moving average with a 30-day threshold. Once the trend was calculated, I used the Autoregressive Moving Average model (ARMA) on the previous 400 days with twelve lags for the model in order to produce confident predictions. And finally, to produce the prediction, a sample of data from the previous two weeks was used to forecast the coming week.
I then average the forecasted Goldstein numbers to find the forecast score for the coming week. The suggested action is based on the inverse of forecast score’s distance from the target score of 1.0: ‘talking on the phone’.