Forecasting and preventing conflict using Artificial Intelligence

Forecasting and Preventing Conflict using Artificial Intelligence

Armed conflicts pose significant challenges for the international community.  They prevent countries from developing, lead to a persistent outflow of refugees, and have a significant death-toll.  For this reason, the ability of policymakers to prevent outbreaks and escalations is crucial.  However, it is not an easy task. Decision-making requires good data, reliable risk evaluations, and a conceptual framework that helps evaluate the options at hand.

The BSE Working Paper 1355, “Dynamic Early Warning and Action Model,” by Hannes Mueller, Cristopher Rauh, and Alessandro Ruggieri, fills this gap by presenting two modules developed for the UK Foreign, Commonwealth Development Office. First, a forecast model which uses machine learning and text downloads to predict conflict outbreaks and intensity. Second, a decision-making module that uses this information provides the optimal intervention to prevent conflict damages.  Together these two modules create a quantitative benchmark for policymakers to choose the best intervention.

The Forecasting Module

By building a cutting-edge machine learning model, the authors can track the entire conflict cycle, from forecasting new outbreaks, escalation, and de-escalation out of conflict to the re-emergence of conflict in post-conflict phases where countries are particularly fragile.  To do so, the authors integrate a text-based forecast of conflict outbreaks with a geo-spatial and temporal forecast of conflict dynamics during conflict.

The forecast is derived from supervised and unsupervised machine learning and the natural processing of millions of news articles. Using data from the Uppsala Conflict Data Program (UCDP), the authors can train the model to forecast future conflict using news articles that provide monthly warning flags.  

The estimated risk of conflict is accessible through the webpage: To support prevention efforts, the authors provide monthly updates of the forecasts at the national and subnational level shown in Figures 1a and 1b. The performance of this system is relatively good. If the national model suggests that a conflict will break out within the next 12 months, it does with a likelihood of 80 percent. Furthermore, at this level of precision the model is able to spot half of all outbreaks. The subnational model performs even better. A visual impression of the good performance can be grasped from Figure 2 which shows the forecast of the subnational level together with its realization.

Figure 1a: National Forecast of Armed Conflict Risk
Figure 1b: Subnational Forecast of Armed Conflict Risk

However, generally good, performance of this system is driven by the conflict trap. It is easy to predict that conflict will break out in places that just suffered from conflict. For outbreaks that follow long periods of peace performance falls. At the national level precision falls from 80 to just above 10. A warning sent to a policy-maker will only materialize in 1 out of 10 instances. Could it still be useful to act on such imprecise forecasts?  More generally, how much should policy-makers focus on ongoing conflicts vs. places that are not suffer from violence at the moment? This can only be evaluated if we take into account the future costs caused by conflict. This is where the second module is important.

Figure 2a: Word map of intensity prediction
Figure 2b: Word map of intensity realization

The Decision-Making Module

The problem with evaluating preventive interventions, like mediating in peace negotiations, providing security aid or sending peacekeepers, is that they are actions taking with a cost in the present but targeted at changing the future. The decision-making module must therefore take into account how interventions change the future expected conflict costs of the country. The second module brings together the following factors to make these decisions: i) the (forecast) likelihood of conflict; (ii) the damage caused by future conflict; (iii) the effect of interventions on conflict dynamics; and (iv) the likelihood of success of an intervention.

To do that, the module simulates possible futures using a conflict risk state model. The idea is to use the forecast information to classify countries into risk states and then simulate the expected future costs using the past dynamics between these states. In the model, conflict states are grouped into three main categories: pre-conflict, post-conflict, or in-conflict. Given the persistence of conflict and its incredible costs in terms of human suffering and economic damages the main question is then whether policy-makers would ever want to intervene in states that are non in-conflict.

Here is where the forecasts and the dynamic view are so important. Present conflict damages do not necessarily indicate high future costs. Dynamics are important. To capture the dynamics of countries moving from one state to another, the authors use a Markov chain that provides the transition probability from each possible state today to each possible state tomorrow.  This allows interventions to be modeled as a change in the future path of a country. Different interventions, such as early or late prevention, will reduce the likelihood of escalating into a conflict. Stabilization increases the probability of de-escalating into a lower risk state from conflict. In this way the future repercussions of interventions can be captured.

What should be done?

The main finding of the authors is that interventions in the post-conflict states can be extremely efficient as conflict risk is objectively quite high in these states which makes future expected conflict costs extremely high. At the same time these states are not violent which means that a focus on ongoing violence can be a bad policy advisor. Overall, the authors consider that the total expected economic benefit of reinforced preventive efforts would bring monthly savings in expected costs of 26 billion USD with a monthly gain to the UK of 630 million USD.

However, there are also some limits to the current research. Most importantly, intervention costs are not modeled explicitly. The reason is that conversations with the FCDO did not reveal clear policy cost estimates. This is important because potential gains from interventions needs to be balanced with their costs. If it is, for example, much cheaper to de-escalate a non-violent conflict than a violent conflict then this would speak in favor of a more preventive approach. It would also allow the authors to model optimal future interventions.