In 1907 Sir Francis Galton analyzed data from a weight-judging competition at the West of England Fat Stock and Poultry Exhibition. To compete for prizes, about 800 individuals paid to guess the weight of the ox once it had been slaughtered and dressed. The remarkable finding, repeated in many ways since, is that the average of all the guesses was within one percent of the correct wight, yet no individual guess was that close. This principle is known as the wisdom of crowds. This suggests it is too hard to build a single model that is perfect, but rather one is better off to build lots of imperfect models and average across them. EBMA improves prediction by pooling information from multiple forecast models to generate ensemble predictions.This approach has been extended to political forecasting models (here and here) and incorporated successfully into ongoing forecasting efforts undertaken by government and commercial enterprises.
Despite the overwhelming avalanche of data that characterizes our world, often times there is a lot of information that simply isn't available. This missing data problem was swept under the rug previously, leading to biased models and erroneous predictions. Now there are several well known techniques to impute missing data. Few of these are ideal for predictive models, and most of them are computationally cumbersome. We have adapted an approach that is both fast and easy, reducing the computational barriers while maintaining a principled flexibility. Ironically, this modern approach is based on a fundamental theorem from 1959 (Sklar's theorem). We use a type of copula method applied to rank order transformations to reduce bias in our models and predictions (Fast and Easy Imputation of Missing Social Science Data). Copulas are newly constructed distributions that link together interdependent random variables, which can be continuous, ordinal, or categorical.
Many models assume that the differences can be captured by looking at the right variables. Risk is probably different, and some places are immune from risk, while others are rife with risk. The risk of debilitating social protests in New Zealand is close to zero, while it is probably much higher in the Ukraine. Some models are able to split the cases being analyzed into two groups: those with substantial risk and the others which are essentially immune to the kind of events being studied. This approach is known as split population modeling, one of the ways that Predictive Heuristics tackles the analysis of risk in the entire world.
Many standard approaches to analyzing risk in countries only use variables that are measured at the country level. But in actuality, there are global variables-say oil prices-that impact all countries in some important fashion. In the same way, countries are interrelated such that bank failures in Greece affect the economies in many countries around the world, including its immediate neighbors in Europe. And, too sine groups of countries-say the OECD countries-have institutions that affect and constrain their behaviors as well. In reality variables from many different levels, global, regional, country-level, and even within country-affect the risk environment. As such, we employ a type of approach that embraces all of these kinds of aspects of the empirical world to aid in understanding risk in specific countries.
Click the icons to view recent reports and papers
In the social sciences, and in the subject matter community, there is a certain skepticism of prediction. This paper looks at the history of prediction going back over 1000 years and shows how prediction is not only important, but a necessary part of understanding the past.
Each month over the past several years, we've generated monthly predictions of risk in most countries in the world. These predictions are specific to 1) ethnic and religious violence, 2) insurgencies, 3) rebellions, 4) domestic political crises, and 4) international crises. We also predict the level of anti-governmental protests. An example of these monthly reports is available for download.
We also predict on a monthly basis the likelihood of irregular leadership changes in each country in the world. This includes coups d'etat, rebellions, and other forms of irregular political events. These are very rare events, but very consequential when they happen. It is important to try and predict events that have low probability but high impact. In fact most of politics is about such events. Feel free to download the example of our monthly ILC report.
Who we are
Michael D. Ward is one of the leading social scientists focused on prediction of social trends using modern data collection and statistical methods. He has published more than a dozen books as well as over 100 articles in political science, geography, sociology, economics, and statistics. He has been a consultant to governments and corporations over the past decade and was one of the principal scientists for the W-ICEWS project--now a program of record--to provide predictive models of political crises for the U.S. government.
Sandra L. Ward has a BS from Loyola University, and a MS and Ph. D. from Northwestern University. Her dissertation developed simulation methods for controlling automated chemical processes. She has worked in computer automation at Argonne National Laboratories, She also worked at the Science Center Berlin on methods for social science models of large scale political processes. Subsequently, she developed remote procedure technologies for Burroughs (Unisys) and Netwise. Subsequently, she worked for two decades at Microsoft helping with the SQL engine and applications in transaction processing.
Andreas Beger has a political science Ph.D. from Florida State University. He has worked on conflict forecasting since 2011 and previously also was a Military Intelligence officer in the Florida Army National Guard.
© 2016 Predictive Heuristics