Lux Actuaries has developed the Lux Conflict, War and Strife (CWS) Predictive Model which has been designed to assess the risk of future armed conflicts globally, considering a multitude of indicators that cover all areas of development, governance and overall socio-economic strength of countries across time. This blog post discusses feature importance analysis - a key strength of the Lux CWS model - which helps to elucidate some of the black-box workings typical of most machine learning models.

The Lux CWS Model

The Lux CWS model aims to predict, at a country level, whether that country will be in a state of armed conflict during each of the next few years. A country is defined to be in a state of armed conflict if there is a contested incompatibility that concerns government and/or territory where the use of armed force between two parties, of which at least one is the government of a state, results in at least 25 battle-related deaths in one calendar year. 

The model considers a broad range of indicators and conflict history for all countries and determines a probability that a country will be in a state of armed conflict for each of the next three years. By monitoring the evolution of these probabilities over time, we can assess whether the risk of conflict in a particular country is increasing or decreasing, based on changes in the underlying features identified.

Features affecting conflict risk

The CWS model considers 108 indicators to assess armed conflict risk. These have been broadly classified into 5 categories – Demographic, Economic, Employment, Political and Resource Wealth. Together, these indicators provide a holistic picture of the socio-economic and political strength of each country over a 60 year timeframe (1960 to 2020). The biggest advantage of using a data-driven, Machine Learning-based approach is that ML algorithms can automatically determine which features have more predictive power – a task which is quite difficult to accomplish using more conventional stochastic methods given the size of the data and the large numbers of features under consideration. 

The chart below shows the most predictive features on a global level – we observe that Demographic and Economic factors are the most important.

The two most important features as shown above are:

• Number of births (in thousands), and
• Education expenditure as a percentage of gross national income.

These two factors have substantially greater predictive power than the remaining features. The next most important factors include:

• Percentage of men in labor force with advanced education who are unemployed,
• Ratio of Male to Female Population for the 41-50 age group, and
• Maternal Mortality ratio (maternal deaths per 100,000 live births).

Monitoring these indicators closely over time can potentially help gauge the risk of armed conflict well in advance and take necessary measures to mitigate it. By identifying the most important features, we can narrow our focus from 108 features to just 8-10 and still capture most of the risk. 

Feature importance varies by geography

Feature importance can be performed globally, as above, but also performed separately for different categories of countries, revealing interesting insights into the differences between the risk profiles of these categories. For example, comparing two regions: the Caucasus/Central Asia and Sub-Saharan Africa shows substantial differences in which features have greater predictive power. Below is a chart showing the top predictive features for both regions.

We observe that the Caucasus/Central Asia and Sub-Saharan Africa have risk profiles that are quite different from each other, and that restricting the analysis only to an overall, global level would not help identify an important nuance that the data-driven nature of the CWS model is able to extract.

Similar to a region-based analysis, we can consider segmentations of countries by income group and grade (i.e., first-world, second-world and third-world countries).

It is important to note that this analysis does not reveal the exact nature of the relationship between different indicators and the likelihood of armed conflict – it only shows us which features have a high degree of predictive influence on the likelihood of armed conflict. The nature of the relationship itself is shown through other analyses, which will be explored in later articles.