Meet the individuals who submitted the winning proposals for the model challenge, which sought new and innovative models to predict where and when atrocities will occur in the near future, based on sociopolitical indicators from around the world and data on past atrocities. These winning submissions were chosen based on the predictive performance of their algorithm
and their use of key data.
For each five day period, this algorithm builds a decision tree using data from recent atrocity and social-political records. The algorithm then makes the prediction as the weighted linear combination of the recent decision tree predictions.
This model combines simple statistic methods with correlations observed in the given sociological and regional data. For example, there are often significant increases of distinct sociological events right before an atrocity occurs. Additionally, atrocity events have the tendency to spread locally, crossing regions or country borders. The core of the application rests on a few simple formulas reflecting the observations made and fine-tuned using the given test data.
Different events will result in different atrocity frequencies. Therefore, one can make predictions by setting conditions for different situations and combining multiple models. Conditions could include what happened in a region, what happened in a country, what happened in each region's neighbor, etc.
This model uses several, independent algorithms that offer predictions. These algorithms are based on either statistics of past atrocities that happened in the same region or in the same country, or on the correlation between the number of sociopolitical events and atrocities. The yields of these predictors are mixed into the final output, and the weighting used is periodically recalculated to reflect changes in the global situation.
Using data from past atrocities, this algorithm forecasts atrocities using weighted sums of predicted variables in each region and country.
This algorithm views the world as a mosaic of harmful and neutral regions tenuously interacting with each other. It creates a network of calculated breaking points between regions, then tests the tension induced by current sociopolitical events against the stress-bearing capacity of the network. If the tension between two regions exceeds the dynamically determined tension capacity, it calculates the likelihood that these regions will harm each other.