POLI3148 Assignment 1 · ACLED event data

From Capital War to Fragmented Civilian Insecurity

Mapping and identifying (by predictive modelling) violence against civilians in Sudan's civil war using ACLED records from 15 April 2023 through 22 April 2025.

16,004

ACLED events

43,256

reported fatalities

4,469

civilian-targeting events

19

state-level areas

Research Question

This report asks: How has Sudan's civil war shifted geographically and politically since April 2023, and can ACLED event patterns help identify where violence against civilians is most likely to intensify? The analysis treats ACLED as an event-level record of reported political disorder rather than a complete census of harm. That distinction matters: the dataset is excellent for comparing reported patterns across time, place, event type, and actors, but fatalities and civilian harm are still shaped by source availability, access, and coding rules.

Background and Key Terms

Sudan's current war began on 15 April 2023 as a confrontation between the Sudanese Armed Forces (SAF) and the Rapid Support Forces (RSF). The conflict first produced intense fighting in Khartoum, where control over the capital carried both military and symbolic value, but it later expanded through Darfur, Al Jazirah, Kordofan, Sennar, and other state-level arenas. ACLED's Sudan analysis describes this evolution as a war shaped by shifting front lines, fragmented alliances, and external support, rather than a static two-party battlefield (Ali et al., 2025; Birru, 2024).

ACLED records reported political violence, demonstrations, and strategic developments by date, location, actors, event type, and reported fatalities (ACLED, 2024; Raleigh et al., 2010). The report uses these records to compare conflict patterns over time and across Sudan's states, while treating the dataset as a record of observed and coded events rather than a full measure of all harm.

Admin1 / state

ACLED's admin1 field is the largest subnational administrative unit. In this dashboard, it is shown as “state” or “state-level area” for readability.

State-month

A state-month is one Sudanese state in one calendar month. This is the unit used for trend summaries and the predictive risk model.

Civilian targeting

ACLED marks events where civilians are directly targeted. This report uses that coding to track civilian-facing insecurity, not total civilian harm.

Figure 1. Monthly ACLED events and reported fatalities in Sudan, April 2023-April 2025. Stacked bars show monthly event counts by ACLED event type; the black line shows reported fatalities.

Interactive dashboard layer

Figure 2. Conflict Story Explorer

Use the controls once and the time series, state ranking, event-type mix, actor ranking, and narrative insight update together. Click a state bar to drill into that state.

0

events

0

reported fatalities

0

civilian-targeting events

0%

civilian-targeting share

Finding 1: The War Begins as a Capital-Centered Contest

The first phase is unmistakably centered on Khartoum. Across the export, Khartoum records 6,838 events, more than any other state. The monthly series (Figure 1) shows an immediate surge after 15 April 2023, with battles and explosions/remote violence dominating the opening pattern. The peak month by event count is October 2024, when ACLED records 897 events. Fatalities follow a less stable trajectory, peaking in October 2024 with 3,688 reported deaths.

The linked Conflict Story Explorer (Figure 2) lets readers test this pattern interactively by changing region, state, event type, and measure. This distinction between events and fatalities is substantively important. Event counts capture the tempo of recorded disorder; fatalities capture reported lethality, which is more volatile and often concentrated in a small number of severe events. In Sudan, the capital fight generates enormous event volume, but later phases of violence in Darfur and the central Nile corridor carry a civilian-security logic that cannot be reduced to the Khartoum front line.

The event-type mix also matters. Battles account for 5,112 (31.9%) of events, explosions and remote violence for 3,830 (23.9%), violence against civilians for 3,374 (21.1%), and strategic developments for 3,419 (21.4%). Strategic developments are not equivalent to direct violent incidents; ACLED uses this category for politically important non-violent activity such as looting, recruitment, arrests, or territorial transfers, so it is best read as context for later violence rather than as a simple violence count.

Figure 3. Event-level geography of Sudan's civil war. Each point is an ACLED event. Use the legend to isolate event types and hover for event details.

Finding 2: Civilian Insecurity Fragments Across Regions

The event-level point map (Figure 3) and state choropleth (Figure 4) show a movement from the capital toward a wider arch of insecurity. Khartoum remains the largest event cluster, but the highest reported fatality burden is in North Darfur, with 11,229 reported fatalities. Civilian-targeting events are most numerous in Khartoum, which records 1,341 such events. That pattern supports the project's central claim: the war is not only a SAF-RSF battlefield; it is also a dispersed civilian-protection crisis.

The monthly state animation (Figure 5) makes the timing of this spread easier to inspect. Two spatial shifts stand out. First, Darfur remains a high-lethality theater, especially when North and West Darfur are viewed alongside South and Central Darfur. Second, Al Jazirah and surrounding central/Nile states become major sites of civilian targeting after the conflict expands beyond the capital. This matters politically because violence against civilians often signals territorial control, predation, reprisal, and local governance breakdown rather than conventional battlefield exchange.

Figure 4. State-level distribution of ACLED events, fatalities, civilian targeting, and civilian-targeting share. Dropdown controls switch the state-level map between measures.

Figure 5. Monthly spread of conflict and civilian targeting by state. The animation aggregates events to state-month bubbles. Bubble size reflects total events; color reflects civilian-targeting events.

Finding 3: Actor Structure Is Bipolar, But Localized

The actor network (Figure 6) makes the SAF-RSF axis visible without making it the whole story. Nodes represent actors and edges represent co-involvement in the same ACLED event. SAF and RSF are central because they appear across the largest number of recorded interactions, but civilian nodes, unidentified armed groups, Darfur communal militias, joint forces, police, and local armed movements fill the surrounding structure. This is what fragmented insecurity looks like in event data: a national confrontation generates local actor constellations that vary by state and month.

ActorTypeWeighted degree
RSFRSF11,007
SAFSAF7,759
Civilians Civilians5,489
Unidentified Armed Group Other405
Darfur Communal Militia Armed group351
SPLM-N-Abdelaziz: SPLM (North) (Abdelaziz...Armed group152
Darfur Joint Forces/JSAMF: Joint Force of...Armed group90
Darfur Arab Militia Armed group81

The network should not be read as an alliance map. ACLED actor1-actor2 ties indicate co-presence in recorded events, not durable cooperation. Still, the network is useful because it shows where a simple two-actor war narrative is too thin. Civilian targeting is partly associated with the front line, but also with militia activity, looting/property destruction, abductions, and local armed governance.

Figure 6. Actor co-involvement network for major conflict actors. The graph displays the strongest actor-pair ties by event count. Larger nodes have higher weighted degree.

Finding 4: Recent Conflict Intensity Predicts Civilian-Targeting Risk

The machine-learning task predicts whether a state-month will experience a high civilian-targeting count in the following month. Here, a state-month means one Sudanese admin1 unit in one calendar month, and “high” means above the 75th percentile among labelled state-months. The feature-effects plot (Figure 7) and held-out confusion matrix (Figure 8) summarize how the transparent logistic classifier behaves. The model is trained on lagged conflict features: previous-month events, fatalities, battles, remote violence, SAF and RSF involvement, distinct actor count, three-month rolling conflict totals, and broad macro-region indicators. It is intentionally simple so that its results can be interpreted rather than treated as a black box.

Target:

1 if next month's civilian-targeting count is above the 75th percentile of labelled state-months; 0 otherwise.

Validation:

Held-out state-months from 2024-11-01 to 2025-03-01; train rows = 361, test rows = 95.

Performance:

Accuracy 0.86, precision 0.72, recall 0.81, F1 0.76, AUC 0.93.

Interpretation:

Probabilities are early-warning indicators, not deterministic forecasts; they summarize patterns in reported ACLED events.

The model's strongest positive predictors are lagged civilian targeting, recent event concentration, actor diversity, and location spread. In plain terms, areas with a recent mix of fighting, civilian targeting, and multiple actors are more likely to remain high-risk in the following month. This is useful for early warning, but partly reflects conflict autocorrelation: violence is often most likely where violence has recently occurred. The latest model-estimated risk map (Figure 9) is based on April 2025 features and estimates risk for the next month after the export window.

State (admin1)Predicted high-risk probabilityCurrent eventsCurrent civilian targeting
Khartoum100.0%7327
North Darfur99.9%8129
North Kordofan96.2%4110
South Darfur90.9%204
Al Jazirah70.7%31
West Darfur52.2%52

Figure 7. Feature effects in the civilian-targeting risk model. Positive coefficients increase predicted high-risk probability; negative coefficients reduce it.

Figure 8. Held-out test confusion matrix for the simple logistic model.

Figure 9. Model-estimated next-month high civilian-targeting risk by state. The map is generated from the latest month in the available export and should be refreshed with newer ACLED data before policy use.

Conclusion

The evidence supports a three-part answer to the research question. Geographically, the war begins as a capital-centered contest, then widens into Darfur, Kordofan, Al Jazirah, Sennar, White Nile, and other state-level arenas. Politically, the SAF-RSF confrontation remains central but does not exhaust the conflict structure; local militias, civilians, police, armed movements, and unidentified groups shape the lived geography of insecurity. Predictively, recent civilian targeting and broader conflict intensity provide useful warning signals for where civilian-targeting risk may intensify next.

A main limitation is measurement: ACLED records reported events and estimated fatalities, so access constraints and source coverage can affect comparisons across remote and urban areas. The analysis also does not adjust for population size, displacement, humanitarian access, or local media density, all of which could change how state-level risk should be interpreted. Secondly, the model predicts state-month risk, not individual incidents or causal mechanisms. Its value is as an interpretable early-warning layer to guide closer qualitative and geographic investigation.

References