Conflict Intensity Metric

• Conflict Intensity Metric is a quantitative measure that operationalizes conflict by capturing escalation, persistence, and spatial distribution via customizable units of analysis.
• It employs methods such as Poisson point processes, dynamic elastic net, and latent variable models to forecast daily events, track changes in fatalities, and infer latent intensity regimes.
• The metric adapts to various temporal and spatial resolutions, enabling nuanced risk assessment and comparisons across different conflict scenarios.

Searching arXiv for the specified paper and closely related work on conflict intensity metrics. A conflict intensity metric is a quantitative operationalization of how much conflict is occurring, how rapidly it is unfolding, or how severe its consequences are, defined at a specified unit of analysis such as an event, day, week, month, grid cell, or grounded response. In recent arXiv literature, the term encompasses forward event rates in point processes, changes in log conflict fatalities, inverse interevent-time measures, heavy-tail exponents for fatalities per event, ordinal latent intensity classes, and spatial density/concentration indicators. These formulations do not measure a single invariant quantity; rather, each metric isolates a particular aspect of escalation, persistence, clustering, criticality, or contradiction at a chosen temporal and spatial resolution (Sun et al., 2022, Attinà et al., 2022, Bruggeman, 2024, Abdou et al., 2 Feb 2026, Stoehr et al., 2022, Walther et al., 2020).

1. Conceptual scope and units of measurement

Conflict intensity metrics differ first by the object they score. Some are rate measures, intended to estimate the expected number of events over a short horizon. Others are severity measures, intended to quantify the heaviness of the tail of event sizes or the expected burden per event, site, or unit area. Still others are latent-state measures, which infer an unobserved intensity class or regime from observed indicators such as casualties, action categories, or recent battle deaths. A further class comprises spatial measures, in which the magnitude and internal geographic distribution of violence are treated as distinct properties. This diversity reflects the fact that “intensity” is not a primitive observable; it is an operational choice tied to modeling purpose, data resolution, and the substantive mechanism under study.

Metric family	Unit of analysis	Interpretation
Forward Poisson rate $\lambda_t$	Country-day	Expected attacks per day
$\Delta_s Y_{c,t}$	Country-month	Change in log fatalities
$I^* = 1/\mathrm{median}(\tau)$	Event-time window	Inverse typical lull
$\hat{\alpha}$	Fatalities per event	Tail heaviness / large-event risk
$\bar{z}_n$	Event	Latent ordinal intensity
$CI_{jk}, CC_{jk}$	Cell-period	Density and clustering

At the daily country level, the Neural Forward-Intensity Poisson Process (NFIPP) defines conflict intensity as the forward Poisson rate $\lambda_t$ for the immediately following day, predicted from covariates and past event history. At the monthly country level, Dynamic Elastic Net defines the target as the step-specific change in log-transformed fatalities, $\Delta_s Y_{c,t} = Y_{c,t+s} - Y_{c,t}$, where $Y_{c,t} = \log(1 + F_{c,t})$. At the event scale, the ordinal latent variable model defines intensity through a latent class $z_n \in \{1,\dots,C\}$ and its posterior summaries, while the Spatial Conflict Dynamics indicator separates conflict intensity within a region, $\Delta_s Y_{c,t}$0, from conflict concentration, $\Delta_s Y_{c,t}$1 (Sun et al., 2022, Attinà et al., 2022, Stoehr et al., 2022, Walther et al., 2020).

This suggests that any encyclopedia-level definition of conflict intensity must be indexed by at least three design choices: the event process being modeled, the resolution of observation, and the inferential purpose. A rate-like metric is appropriate when forecasting or warning is the objective; a tail or latent-class metric is appropriate when escalation, severity, or regime structure is the central concern; and a spatial metric is appropriate when diffusion or clustering is substantively important.

2. Forward-intensity and count-process formulations

In NFIPP, daily event counts are treated as realizations of a Poisson point process whose intensity is learned by a neural network. The daily increment on $\Delta_s Y_{c,t}$2 with $\Delta_s Y_{c,t}$3 day is

$\Delta_s Y_{c,t}$4

and the forward intensity is

$\Delta_s Y_{c,t}$5

In the daily discretization used there, the expected count over a window is $\Delta_s Y_{c,t}$6, and aggregated counts satisfy $\Delta_s Y_{c,t}$7 with $\Delta_s Y_{c,t}$8. The feature stack $\Delta_s Y_{c,t}$9 includes past violence measured as mean daily attacks in the past 15, 90, and 365 days; counts of floods and droughts in the past 6 and 36 months; and sums of absolute temperature anomalies in the past 6 and 36 months. The final layer uses a softplus nonlinearity, $I^* = 1/\mathrm{median}(\tau)$0, and training maximizes the Poisson log-likelihood

$I^* = 1/\mathrm{median}(\tau)$1

The model is explicitly excitation-driven: climate shocks and lagged attacks enter through a nonlinear mapping rather than through a prescribed Hawkes kernel (Sun et al., 2022).

Two further count-process formulations shift the target away from daily rates while preserving a probabilistic interpretation. Dynamic Elastic Net models, for each country and horizon $I^* = 1/\mathrm{median}(\tau)$2, the change in the log-transformed count of state-based conflict fatalities:

$I^* = 1/\mathrm{median}(\tau)$3

The forecasting equation is estimated with an elastic-net objective with $I^* = 1/\mathrm{median}(\tau)$4, allowing adaptive variable selection among approximately 753 predictors. In this setup, “intensity” is not the level of fatalities itself but the forecasted change in log fatalities over a fixed horizon, which is a distinct operationalization from NFIPP’s forward daily rate (Attinà et al., 2022).

The Bayesian AR-HMM formulation defines intensity through the emission mean of weekly battle-death counts conditional on a latent regime. Weekly counts satisfy

$I^* = 1/\mathrm{median}(\tau)$5

with

$I^* = 1/\mathrm{median}(\tau)$6

and

$I^* = 1/\mathrm{median}(\tau)$7

The latent states $I^* = 1/\mathrm{median}(\tau)$8 are interpreted ex post as non-violent, stable violence, and intensified violence. Accordingly, the metric may be reported either as the posterior mean count-scale intensity $I^* = 1/\mathrm{median}(\tau)$9, its log transform, or the regime-high probability $\hat{\alpha}$0 (Williams et al., 2021).

These three formulations share a common premise: intensity is most useful when embedded in a generative or predictive model rather than treated as a raw descriptive count. They differ, however, in whether the quantity of interest is the next-day hazard, the multi-month change in fatalities, or the expected count under latent regime-switching dynamics.

3. Temporal resolution, interevent structure, and burstiness

A separate tradition defines conflict intensity through the distribution of interevent times rather than through counts. The central empirical result is resolution dependence: fine-grained timestamps yield lognormal interevent times, while coarse-grained timestamps such as daily aggregation yield apparent power-law behavior. For fine-grained data, if $\hat{\alpha}$1 denotes the interevent time and $\hat{\alpha}$2 is approximately normal with parameters $\hat{\alpha}$3, then the lognormal-consistent intensity is

$\hat{\alpha}$4

where $\hat{\alpha}$5 is the median interevent time. For coarse-grained data with a power-law fit above $\hat{\alpha}$6, a robust alternative is the inverse median

$\hat{\alpha}$7

The paper’s unified recommendation is

$\hat{\alpha}$8

computed at the finest available resolution, with rolling-window updates of $\hat{\alpha}$9, $\bar{z}_n$0, and optionally a kernel intensity

$\bar{z}_n$1

The rationale is that the inverse median is robust to heavy tails and directly interpretable as the inverse of a “typical lull” (Bruggeman, 2024).

The theoretical basis is a multiplicative process,

$\bar{z}_n$2

which implies approximate normality of $\bar{z}_n$3 by the central limit theorem. Coarse-graining introduces a positive lower bound, censors small lulls, and creates mixtures of lognormals across contexts or periods, each of which can generate apparent power-law scaling. This matters because rate estimates built from daily or monthly aggregates can mis-specify the underlying burst structure when the true process operates on finer time scales.

A plausible implication is that count-based intensity and interevent-based intensity are complementary rather than interchangeable. NFIPP’s daily forward rate $\bar{z}_n$4 and the inverse-median metric $\bar{z}_n$5 may track similar escalation episodes, but they summarize different stochastic objects: one is an expected count conditional on covariates and history, and the other is a robust function of the lull distribution.

4. Severity, tail risk, and latent event intensity

Event severity can itself serve as a conflict intensity metric. In the Yemen and Syria study, the power-law tail exponent $\bar{z}_n$6 of fatalities per event is used as a quantitative proxy for intensity and for the robustness of larger clusters of fighters. For $\bar{z}_n$7,

$\bar{z}_n$8

and the maximum-likelihood estimator is

$\bar{z}_n$9

Lower $CI_{jk}, CC_{jk}$0 means a heavier tail, a higher probability of very large-fatality events, and therefore heightened intensity. In the fusion–fission framework, $CI_{jk}, CC_{jk}$1 near $CI_{jk}, CC_{jk}$2 indicates balanced fusion and fission, while temporary reductions toward $CI_{jk}, CC_{jk}$3 suggest a temporary increase in the robustness and involvement of larger clusters, often ahead of the largest battles (Abdou et al., 2 Feb 2026).

A related scaling formulation uses conflict avalanches. There the basic intensity summaries are the temporal average $CI_{jk}, CC_{jk}$4, the areal average $CI_{jk}, CC_{jk}$5, and the event-normalized intensity $CI_{jk}, CC_{jk}$6, with instantaneous forms derived from $CI_{jk}, CC_{jk}$7. Virulence enters multiplicatively: $CI_{jk}, CC_{jk}$8, so early estimates of $CI_{jk}, CC_{jk}$9 are interpreted as informative about later totals and duration (Lee et al., 2020).

The ordinal latent-variable model approaches the same problem from the opposite direction: instead of fitting a tail to observed event sizes, it posits a latent ordinal class $\lambda_t$0 for each event. The event tuple comprises subject $\lambda_t$1, predicate $\lambda_t$2, quantifier $\lambda_t$3, and object $\lambda_t$4, with

$\lambda_t$5

$\lambda_t$6

The ordinal semantics are induced by ordered priors on the class-indexed parameters, and the practical score is the posterior mean class

$\lambda_t$7

which may then be rescaled to $\lambda_t$8 and aggregated across time and space. This formulation directly addresses a limitation of Goldstein-type action-only scales by incorporating “who,” “to whom,” and casualties (Stoehr et al., 2022).

Taken together, these approaches show that severity-based intensity can be estimated from at least three different statistical objects: the tail of event sizes, the aggregate growth law of conflict clusters, or a latent ordinal class inferred from event attributes. The choice among them depends on whether the primary interest is extreme-event risk, mesoscopic scaling, or event-level semantic/contextual intensity.

5. Spatial formulations and the geography of violence

The Spatial Conflict Dynamics indicator defines conflict intensity within a region as the density of events per unit area and conflict concentration as the internal patterning of event locations relative to complete spatial randomness. For cell $\lambda_t$9 in period $\Delta_s Y_{c,t} = Y_{c,t+s} - Y_{c,t}$0,

$\Delta_s Y_{c,t} = Y_{c,t+s} - Y_{c,t}$1

Conflict concentration is computed via the Average Nearest Neighbor ratio. If $\Delta_s Y_{c,t} = Y_{c,t+s} - Y_{c,t}$2, the observed mean nearest-neighbor distance is

$\Delta_s Y_{c,t} = Y_{c,t+s} - Y_{c,t}$3

the expected mean distance under complete spatial randomness is

$\Delta_s Y_{c,t} = Y_{c,t+s} - Y_{c,t}$4

and

$\Delta_s Y_{c,t} = Y_{c,t+s} - Y_{c,t}$5

The four resulting types are High intensity + Clustered, High intensity + Dispersed, Low intensity + Clustered, and Low intensity + Dispersed. In the North and West Africa application, the grid cells are $\Delta_s Y_{c,t} = Y_{c,t+s} - Y_{c,t}$6 km, so $\Delta_s Y_{c,t} = Y_{c,t+s} - Y_{c,t}$7 km$\Delta_s Y_{c,t} = Y_{c,t+s} - Y_{c,t}$8, and the high/low intensity threshold is the “generational mean” $\Delta_s Y_{c,t} = Y_{c,t+s} - Y_{c,t}$9 events/km$Y_{c,t} = \log(1 + F_{c,t})$0 computed over 1997–2016 (Walther et al., 2020).

This decomposition is important because magnitude and spatial expression are not equivalent. A region can display high intensity while remaining tightly localized, or low intensity while being geographically diffuse. The empirical application reports that clustered cells dominated historically, but dispersed cells rose in later years, which the authors relate to changing spatial forms of violence. The metric therefore treats diffusion and concentration as analytically separate from event volume.

Spatial formulations also connect naturally to count and severity models. A daily forward rate $Y_{c,t} = \log(1 + F_{c,t})$1 may be forecast for a country, yet the same country can contain cells occupying different SCD categories. Likewise, a low tail exponent $Y_{c,t} = \log(1 + F_{c,t})$2 for fatalities per event can coexist with either clustered or dispersed spatial expression. This suggests that spatial intensity measures are best understood as orthogonal complements to temporal-rate and severity metrics rather than substitutes for them.

6. Evaluation, assumptions, and extensions beyond armed conflict

A recurring issue is that many conflict intensity metrics are predictive or operational rather than structurally causal. NFIPP explicitly states that “excitation causality” is assessed via predictability improvement, not structural causality, and operationalizes attribution through Likelihood Gain, $Y_{c,t} = \log(1 + F_{c,t})$3, by comparing a violence-only baseline to a violence-plus-climate model. Dynamic Elastic Net evaluates forecasts with MSE and TADDA, and the AR-HMM emphasizes posterior regime probabilities and credible intervals rather than point forecasts alone. Resolution-aware interevent metrics stress censoring, truncation, and matched-support model comparison, while the fatality-tail approach notes that the study intentionally did not perform model comparison against other heavy-tail families in that application (Sun et al., 2022, Attinà et al., 2022, Williams et al., 2021, Bruggeman, 2024, Abdou et al., 2 Feb 2026).

Common limitations recur across formulations. Sparse counts reduce power even when Poisson or Negative Binomial likelihoods are used. Reporting delays, under-reporting, and rounding affect both count and severity estimates. Spatial metrics depend on grid choice and are affected by edge effects. Latent-variable models inherit ontology and reporting biases from event coding schemes. Heavy-tail exponents can shift with windowing and cutoff choice. These are not peripheral implementation details; they determine what “intensity” actually means in practice.

Recent work extends the same general idea of conflict intensity to non-violent domains by redefining the object of conflict. In document-grounded language modeling, ConflictScore decomposes responses into atomic claims and aggregates the prevalence and severity of contradictory evidence through

$Y_{c,t} = \log(1 + F_{c,t})$4

In Dempster–Shafer sensing, the conflict weight is

$Y_{c,t} = \log(1 + F_{c,t})$5

with $Y_{c,t} = \log(1 + F_{c,t})$6 the mass assigned to mutually inconsistent evidence. In traffic conflict detection, the order-statistics intensity is

$Y_{c,t} = \log(1 + F_{c,t})$7

and in lateral autonomous-driving conflicts the Modified-Emergency Index is

$Y_{c,t} = \log(1 + F_{c,t})$8

For structured airspace, a composite cell-level index is

$Y_{c,t} = \log(1 + F_{c,t})$9

These formulations preserve the same logical architecture—define a conflict event, define a scale or horizon, and quantify either the expected rate, probability, or severity of that event—even though the underlying systems are textual, robotic, vehicular, or air-traffic rather than political (Liu et al., 24 Jun 2026, Carlson et al., 2012, Jiao et al., 2024, Cheng et al., 31 Oct 2025, Kendall et al., 12 Jun 2026).

Taken together, the literature indicates that a conflict intensity metric is best viewed as a model-based summary of conflict-generating structure at a chosen scale. In political violence, the dominant realizations are forward rates, changes in fatalities, interevent-time inverses, tail exponents, latent intensity classes, and spatial density/concentration. In adjacent literatures, the same term migrates to contradiction severity, collision criticality, or interpretation inconsistency. The unifying principle is not a shared formula but a shared task: compressing heterogeneous evidence about escalation, contention, or risk into a quantitatively interpretable signal.