Disruptive events are in the focus of this project. Of particular interest are geopolitical events that have a large impact due to their extents and sudden occurrence. Smoldering conflicts that suddenly turn into waves of protest, social unrest or even warlike events are examples of such disruptive developments. The advantages of being able to predict such disruptive changes are obvious. Important for the predictive capability is to understand the diverse mechanisms that can lead to disruptive behaviors. In dynamic and typically non-linear systems - including societal systems - disruptive events often arise due to so-called tipping point dynamics. After reaching a certain threshold, i.e. the tipping point, the system state shifts abruptly. From a mathematical point of view, these “outbreaks” correspond to bifurcations, e.g. saddle-node bifurcations, which explain the abrupt changes. In the context of forecasting, a theory called Early Warning Signals (EWS) has been developed in recent years. The basic idea of this prediction model is based on the observation of different statistical measures that change characteristically as the system approaches the tipping point. Variance, autocorrelation and kurtosis are examples of such measures, which roughly describe the resilience behavior of the system. Preliminary experiments with conflict data collected by the Armed Conflict Location & Event Data Project (ACLED) [1] has revealed that sudden outbreaks of protests or social riots are already indicated by EWS weeks before they occur. The major application fields for EWS forecasts are in ecology and climate models, with a clearly visible spread trend to many other disciplines such as economics and finance.

Goal

The goal of this project is to develop an application framework that allows to determine the system dynamics, specifically a bifurcation dynamics, in time series data. Given the appropriate dynamics in the time series, EWS models will be applied for predictions. Special attention should be paid to the statistical measures or indicators of the EWS model and how well they are suited as features in Machine Learning (ML) models, for example LSTM models [2], a deep learning approach that has proven to be very useful for time series predictions.

Requirements

  • Interest in predictive modelling
  • Mathematical skills and particular interest in bifurcation dynamics
  • Interest in time series analysis
  • Programming skills
  • Skills in machine learning approaches

Work packages

We propose three work packages for this project:

  1. At the beginning, a general familiarization with topics of time series analysis and subsequently with the specific methods of EWS is expected (e.g. [3], [4]). The main activity here is literature study.
  2. An introduction to conflict data (here ACLED data [1]) can be provided in a straightforward manner. However, the use of ACLED time series is not mandatory. In principle, any time series that manifest an appropriate dynamics and interesting predictive use cases can be used. An explorative data analysis at an early stage of the project is appropriate and helpful, to find at most interesting predictive scenarios.
  3. According to the goal of this project, the development of an application framework is required. The framework should enable appropriate preprocessing of the data and implement the EWS model. For R a simple framework based on available R-libraries already exists (cf. [5]). However, other technologies and programming languages (e.g. Python, Julia, …) are welcome.

Interesting predictive analysis and use cases should be considered for a publication.

References

[1] https://acleddata.com/

[2] https://en.wikipedia.org/wiki/Long_short-term_memory

[3] Scheffer, M., Bascompte, J., Brock, W. et al. (2009) Early-warning signals for critical transitions. Nature 461, 53–59. https://doi.org/10.1038/nature08227

[4] Kéfi S., Dakos V., Scheffer M., Van Nes E.H. & Rietkerk, M. (2013) Early warning signals also precede non-catastrophic transitions. Oikos, 122: 641-648. doi:10.1111/j.1600-0706.2012.20838.x

[5] https://duncanobrien.github.io/EWSmethods/articles/using_ewsnet.html