Conflict Forecasting via Conformal Prediction for Markov Processes

arXiv:2604.25139 · stat.ME, stat.AP, stat.ML · Submitted 2026-04-28 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "Conflict Forecasting via Conformal Prediction for Markov Processes".

Jane: Here's the gist of the paper. Whether or not a country is at war, or experiencing escalating or deescalating levels of conflict, has massive ramifications on a country’s national and foreign policy.

Tom: In this paper, we present the use of conformal prediction on temporally-dependent data to obtain prediction sets of possible future conflict state-sequences.

Jane: The motivation for this work stems from the growing recognition that uncertainty quantification is equally critical in forecasting conflict dynamics, as incorrect predictions carry severe consequences and decision-makers require a calibrated sense of the range of plausible outcomes.

Tom: Early efforts focused on point predictions, but there is a growing recognition that uncertainty quantification is equally critical.

Paper summary: Tom: So we're talking about this paper here called "Conflict Forecasting via Conformal Prediction for Markov Processes," and the big idea is using conformal prediction on time-dependent data to get sets of possible future conflict sequences <ref:2604.25139#pg1>. Jane, can you lay out the core thesis for us?

Jane: Well, essentially, this paper addresses how we can predict conflict states—like whether a country is at war or escalating—when that data has a time component <ref:2604.25139#pg1>. The authors are comparing two main strategies: conformal prediction against a likelihood-based prediction strategy when the underlying data follows a discrete-state Markov process <ref:2604.25139#pg1>.

Tom: What does that comparison mean for us on the ground? Like, why not just pick one method?

Lu: Because point predictions just don't give enough information, and the penalty for being wrong is really high in conflict forecasting, so they need a way to show the range of plausible outcomes <ref:2604.25139#pg2>.

Meng: I'm wondering about that uncertainty quantification part. If we get a single forecast, how do we know if it's actually reliable when the underlying process is so messy?

Jane: Conformal prediction offers an alternative because it only needs an exchangeability assumption instead of having to perfectly specify the complex model of conflict dynamics <ref:2604.25139#pg2>. This approach produces prediction sets with guaranteed coverage for finite sample data <ref:2604.25139#pg3>.

Tom: But the paper immediately flags a problem, right? They say Markov processes aren't inherently exchangeable, which is the standard assumption for conformal methods <ref:2604.25139#pg3>. That sounds like a major hurdle.

Lu: Exactly, and they tackle this by looking at two different frameworks for Markov processes <ref:2604.25139#pg3>. The first idea is treating each state sequence as an individual observation, assuming those observations are independent and thus exchangeable <ref:2604.25139#pg3>.

Meng: So they are trying to force a form of exchangeability onto something that is supposed to be sequential and dependent over time, which is a tough spot to navigate <ref:2604.25139#pg3>.

Jane: And the second framework leverages the properties of Markov processes and permutation groups to create this notion of partial exchangeability for a single state sequence <ref:2604.25139#pg3>. That's where they build their prediction set methodology <ref:2604.25139#pg4>.

Tom: Okay, so they are trying to get a rigorous way to handle that temporal dependency while still using the principled approach of conformal prediction <ref:2604.25139#pg1>. How does this translate into actually predicting future conflict sequences?

Jane: The ultimate goal isn't just one future state sequence, but a prediction set C one-alpha for any level from zero to one, such that the true sequence is inside that set with at least a probability of one minus alpha <ref:2604.25139#pg5>. They compare this against the likelihood-based approach which constructs sets based on highest probability mass <ref:2604.25139#pg4>.

Paper summary: Lu: The likelihood-based method just fits a Markov model to the calibration data and picks the highest probability mass set, defined by finding all future sequences that have a probability greater than or equal to one minus alpha <ref:2604.25139#pg4>. It's straightforward math, but it relies heavily on having the correct model <ref:2604.25139#pg1>.

Meng: If that model is wrong, the likelihood-based prediction set can get really wide when things get uncertain <ref:2604.25139#pg4>. That sounds risky when you're trying to deploy a forecast in a real system <ref:2604.25139#pg7>.

Tom: And the simulation study showed something specific about this comparison, right? How did the two methods actually perform against each other?

Jane: Figure two shows that conformal prediction remains well calibrated for all confidence levels and test lengths, which aligns with what they saw in previous work <ref:2604.25139#pg7>. In contrast, the likelihood-based approach tends to over-cover as both the test length and confidence level decrease <ref:2604.25139#pg7>.

Lu: Look at Table one which compares the size of these prediction sets for different forecast lengths and target coverage levels <ref:2604.25139#pg8>. On average, the likelihood-based prediction sets contain fewer predicted state sequences than the conformal prediction sets, except in one specific case where forecast length is one and the target coverage level is zero point five five <ref:2604.25139#pg8>.

Meng: So for a typical scenario, CP gives us more options to consider than the likelihood method does in terms of the number of sequences it suggests <ref:2604.25139#pg8>. What about when the input data is very simple?

Tom: That brings up their limitations, I think. They point out that conformal prediction isn't as efficient as the likelihood-based prediction when all of the observed states in the calibration data are actually the same state <ref:2604.25139#pg7>.

Jane: That inefficiency falls dramatically when the last state in your test sequence is different from the dominant state seen during training <ref:2604.25139#pg7>. They also specifically noted that for a special case where all calibration data is just one state, CP is less efficient than the likelihood-based prediction <ref:2604.25139#pg7>.

Lu: The authors even suggest a workaround for that specific inefficiency by adding an artificial time point at the end of the sequence <ref:2604.25139#pg8>. That’s a bit of a hack, but it shows they are trying to find where CP breaks down in efficiency <ref:2604.25139#pg8>.

Meng: From an engineering standpoint, that sounds like something we'd have to build special handling for if we were deploying this at scale with very simple input data <ref:2604.25139#pg8>. It’s not a clean solution yet.

Tom: So to wrap up the summary of "Conflict Forecasting via Conformal Prediction for Markov Processes," the paper shows that conformal prediction sets maintain coverage across all forecasted steps, which is important because it preserves the temporal dynamics that matter in many applications <ref:2604.25139#pg9>.

Paper summary: Jane: That's true, and unlike likelihood-based prediction sets which might collapse to a stationary distribution when you project many steps ahead, conformal prediction is better positioned to capture the true trajectory uncertainty <ref:2604.25139#pg9>.

Lu: So the paper really advocates for CP because it’s better at reflecting genuine trajectory uncertainty rather than just fixed state probabilities <ref:2604.25139#pg9>. It’s about capturing the dynamics, not just a static probability distribution <ref:2604.25139#pg9>.

Meng: If we look at the implications for deployment, it suggests that for forecasting sequences, uncertainty quantification matters more than just getting one best guess <ref:2604.25139#pg1>. We need sets of possibilities <ref:2604.25139#pg7>.

Tom: Right, so the authors are basically arguing that for complex, time-dependent processes like conflict states, we need robust methods that don't rely on assuming perfect model accuracy <ref:2604.25139#pg2>. They give us a way to get calibrated sets of plausible outcomes <ref:2604.25139#pg1>.

Jane: And the comparison with likelihood methods shows that while they have some efficiency trade-offs, conformal prediction seems more reliable for capturing the actual evolution of the sequence over time <ref:2604.25139#pg9>. That’s a key difference for anyone trying to use this in policy decisions <ref:2604.25139#pg1>.

Lu: It’s about moving beyond just point predictions and giving decision-makers a calibrated range of outcomes that respect the temporal nature of the data <ref:2604.25139#pg5>. That's the core contribution here <ref:2604.25139#pg1>.

Meng: So, for production systems, it means we need to be careful about how we define our uncertainty because a simple model fitting approach might give us a deceptively narrow view of what could actually happen <ref:2604.25139#pg8>.

Tom: Indeed, this paper on "Conflict Forecasting via Conformal Prediction for Markov Processes" suggests that using conformal prediction to build prediction sets is a principled way to forecast sequences without needing a perfectly specified model <ref:2604.25139#pg1>. It’s about getting those plausible future conflict sequences right <ref:2604.25139#pg9>.

Jane: We should keep an eye on how these methods evolve, especially when dealing with more complex dependencies or when we move beyond the discrete-state Markov process that they tested <ref:2604.25139#pg3>. That’s where the next round of research will likely be needed <ref:2604.25139#pg3>.

Lu: Definitely, because handling those violations of exchangeability in a way that remains efficient is still an open problem in this area <ref:2604.25139#pg3>.

Meng: And for us, it means if we build forecasting tools, we have to account for the possibility that our simplest models might be misleading when they're projecting long sequences <ref:2604.25139#pg7>.

Tom: That’s the practical implication—uncertainty quantification isn't just academic; it directly affects how decisions are made about national security and foreign policy <ref:2604.25139#pg1>.

Jane: So, to wrap up this discussion on "Conflict Forecasting via Conformal Prediction for Markov Processes," the paper gives us a tool for generating prediction sets of possible future conflict state sequences using conformal prediction <ref:2604.25139#pg1>. It shows that CP is better at capturing temporal dynamics than likelihood-based approaches when projecting many steps ahead [ref:two thousand six hundred four point two five one three nine#pg9

Conclusion: Tom: So we're wrapping up this paper "Conflict Forecasting via Conformal Prediction for Markov Processes" <ref:2604.25139#pg9>. It’s about using conformal prediction to get sets of possible future conflict sequences <ref:2604.25139#pg1>.

Jane: Right, it's taking something that changes state over time—like war status—and giving us a range of what could happen <ref:2604.25139#pg5>. The authors are really comparing this conformal prediction approach against a likelihood-based method <ref:2604.25139#pg4>.

Lu: What they’re showing is that the conformal prediction sets keep their coverage across all the future steps <ref:2604.25139#pg7>. That's important because it preserves the actual temporal movement of the conflict state <ref:2604.25139#pg9>.

Meng: But they also admitted a limitation, which is that if all their training data is just one state, conformal prediction isn't as efficient as the likelihood-based method <ref:2604.25139#pg7>. They even have to add an extra time point to fix that <ref:2604.25139#pg8>.

Tom: So, the big picture here is that for forecasting conflict dynamics, we need methods that respect those temporal dependencies and provide a range of plausible outcomes <ref:2604.25139#pg1>.

Jane: Exactly. Likelihood-based methods can sometimes collapse to just fixed probabilities when you look way out into the future, which isn't helpful for understanding actual trajectory uncertainty <ref:2604.25139#pg9>.

Lu: The authors argue that conformal prediction is better at capturing that true trajectory uncertainty over many steps <ref:2604.25139#pg1>. It's about the sequence itself, not just a static probability distribution <ref:2604.25139#pg9>.

Meng: For me, that means if I’m building a system to predict conflict, I can't just pick the simplest model because it might be missing crucial long-term shifts in the state <ref:2604.25139#pg7>.

Tom: So, the paper suggests that using conformal prediction provides a principled way to get those calibrated sets of plausible outcomes for conflict sequences without needing a perfectly specified model <ref:2604.25139#pg1>. It's about getting those future conflict sequences right <ref:2604.25139#pg9>.

Jane: That’s the main message, and it sets up the next question: can this work in real-world policy decisions where time is measured in human terms?

Department of Statistics, North Carolina State University · Department of Mathematical Sciences, The University of Texas at Dallas · Department of War Studies, Swedish Defense University · Department of Mathematics, University of Oslo

stat.ME, stat.AP, stat.ML

Submitted: 2026-04-28

Updated: 2026-11-11

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 82/100

The gist: Whether or not a country is at war, or experiencing escalating or deescalating levels of conflict, has massive ramifications on a country’s national and foreign policy.

Key concepts

Conformal Prediction (CP)
CP is a method used to create prediction sets that guarantee a certain level of coverage for the true underlying data distribution without needing to perfectly specify the statistical model. It achieves this by requiring only an exchangeability assumption, making it robust against model misspecification in conflict forecasting.
Markov Process
A Markov process models conflict dynamics as a sequence of states transitioning over time, where the next state depends only on the current state and not on the entire history. This sequential nature makes it useful for tracking how a country's conflict status evolves over time.
Prediction Sets
Instead of predicting just one future outcome (point prediction), this method constructs a set of all plausible future sequences. These sets provide a range of possible outcomes, which is crucial for decision-makers to understand the uncertainty in conflict forecasting.

Terminology

Summary

Whether or not a country is at war, or experiencing escalating or deescalating levels of conflict, has massive ramifications on a country’s national and foreign policy. p. 0

Where the paper says it

Williams1 , and Gudmund H. Hermansen4 1Department of Statistics, North Carolina State University 2Department of Mathematical Sciences, The University of Texas at Dallas 3Department of War Studies, Swedish Defense University 4Department of Mathematics, University of Oslo Abstract Whether or not a country is at war, or experiencing escalating or deescalating levels of conflict, has massive ramifications on a country’s national and foreign policy.

In this paper, we present the use of conformal prediction on temporally-dependent data to obtain prediction sets of possible future conflict state-sequences. p. 0
Where the paper says it

Given a country’s history of conflict, or lack thereof, future predictions about the war-status of a country are valuable information. In this paper, we present the use of conformal prediction on temporally-dependent data to obtain prediction sets of possible future conflict state-sequences.

p. 0
Where the paper says it

Given a country’s history of conflict, or lack thereof, future predictions about the war-status of a country are valuable information. In this paper, we present the use of conformal prediction on temporally-dependent data to obtain prediction sets of possible future conflict state-sequences.

Motivation and Problem Statement

The motivation for this work stems from the growing recognition that uncertainty quantification is equally critical in forecasting conflict dynamics, as incorrect predictions carry severe consequences and decision-makers require a calibrated sense of the range of plausible outcomes. Early efforts focused on point predictions, but there is a growing recognition that uncertainty quantification is equally critical. p. 1

Where the paper says it

, 2025, Ward and Beger, 2017, Morgan et al. Early efforts focused on point predictions, but there is growing recognition that uncertainty quantification is equally critical: incorrect predictions carry severe consequences, and decision-makers require not just a single forecast but a calibrated sense of the range of plausible outcomes [Hegre et al.

Markov models are promising for conflict dynamics as they represent the evolution of a conflict system as a sequence of states transitioning over time. p. 1
Where the paper says it

This has motivated a shift toward probabilistic forecasting methods and prediction intervals/sets. Among the models that have shown promise for conflict dynamics are Markov models, which represent the evolution of a conflict system as a sequence of states transitioning over time [Randahl and Vegelius, 2022, Williams et al.

However, their uncertainty estimates are sensitive to model misspecification, which is a serious concern given the difficulty of correctly characterizing conflict processes. p. 1
Where the paper says it

Combined with estimated conditional distributions of fatality counts given the current state, these models can generate forecasts of future conflict trajectories [Randahl and Vegelius, 2024, Hegre et al. However, their uncertainty estimates are sensitive to model misspecification, a serious concern given the difficulty of correctly characterizing conflict processes. p. 1

Where the paper says it

Combined with estimated conditional distributions of fatality counts given the current state, these models can generate forecasts of future conflict trajectories [Randahl and Vegelius, 2024, Hegre et al. However, their uncertainty estimates are sensitive to model misspecification, a serious concern given the difficulty of correctly characterizing conflict processes.

Conformal prediction (CP) offers a principled alternative by requiring only an exchangeability assumption rather than a correctly specified model. p. 1
Where the paper says it

Conformal prediction (CP) offers a principled alternative. p. 1

Where the paper says it

However, their uncertainty estimates are sensitive to model misspecification, a serious concern given the difficulty of correctly characterizing conflict processes. Conformal prediction (CP) offers a principled alternative.

By requiring only an exchangeability assumption rather than a correctly specified model, CP produces prediction sets with finite-sample marginal coverage guarantees.

A key challenge addressed is that Markov processes, by definition, violate the standard exchangeability assumption underlying conformal methods. p. 1
Where the paper says it

In this paper, we explore applying CP to discrete-state Markov processes to generate valid prediction sets for future conflict state-sequences, and we compare its performance against likelihood-based approaches. A key challenge we address is that Markov processes, by definition, violate the standard exchangeability assumption underlying conformal methods, and we discuss the implications and limitations of existing approaches for handling their temporal dependence.

Methodology for Prediction Sets

The ultimate goal is to forecast state-sequences into the future rather than a point-prediction. Two approaches are presented: (1) CP in Section 3. p. 2

Where the paper says it

, one future state-sequence of length T1), our goal instead is to construct a prediction set, for any level 1 − α ∈ (0, 1), of state-sequences, say C1−α, that guarantees the true underlying state-sequence is an element of the set with probability at least 1 − α. We present two approaches to constructing prediction sets: (1) CP in Section 3.

1, and (2) likelihood-based prediction in Section 3. p. 2
Where the paper says it

We present two approaches to constructing prediction sets: (1) CP in Section 3. 1, and (2) likelihood-based prediction in Section 3.

2.

The CP methodology leverages a notion of partial exchangeability for a single state-sequence by treating each observed time point as an individual observation and then leveraging properties of Markov processes and permutation-groups to arrive at a notion of partial exchangeability for a single state-sequence. p. 2

Where the paper says it

The first idea is to treat each state-sequence as an individual observation, and then multiple observations of these state-sequences are assumed to be independent and thus exchangeable [Cherubin and Nouretdinov, 2016, Stankeviciute et al. However, instead of relying on multiple realizations of independent state-sequences, the second idea leverages properties of Markov processes and permutation-groups to arrive at a notion of partial exchangeability for a single state-sequence [Chernozhukov et al.

The algorithm involves:

  1. Estimating the transition probability matrix (P) using the calibration data. p. 4

    Where the paper says it

    The likelihood-based prediction sets for individual state-sequences are constructed as follows. Using the calibration data, {Xt} T t=1, we first fit a Markov model by estimating the transition probability matrix Pb (same as step 2 of Algorithm 1).

  2. Determining the set of permutable i-blocks, I =

I i, given XT +T1 = xT +T1 = i.

  1. Calculating a non-conformity score S(π) for permutations π in P. p. 3

    Where the paper says it

    , “identity permutation”) 5. For computational purposes Π := ( P, if |P| ≤ n a random sample without replacement of size n from P, if |P| > n for π ∈ Π do Calculate the non-conformity score S(π) S(π) = 1 − PT1 j=1 PPb X (π) T +j | X (π) T T1 end for 6.

  2. Calculating the p-value qb(x) =

q b(x) = P

π∈Π 1if S(π) > S(I) + u · 1if S(π) = S(I)Π, where I is the identity permutation (refer to step 4 of Algorithm 1.

The likelihood-based approach fits a Markov model to the observed statesequences in the calibration data and constructs prediction sets based on highest probability mass. p. 4

Where the paper says it

[2023] for proof of validity). 2 Likelihood-Based Prediction An alternative approach to CP is to simply fit a Markov model to the observed statesequences in the calibration data and obtain likelihood-based prediction sets for forecasted state-sequences of length T1.

The prediction set C like 1−α is taken to be the highest probability mass set, defined by K = min (k ∈

X T1 j=1 P(XT +T1 = x(j) XT = xT) ≥ 1 − α).

Evaluation and Comparison

The simulation study compares CP and likelihood-based approaches using a known true data generating mechanism. Figure 2 shows that the CP approach remains well calibrated for all confidence levels and test lengths, consistent with results in Nettasinghe et al. p. 7

Where the paper says it

Figure 2 presents the reliability curves for the conformal and likelihood-based predictions. We see that the CP approach remains well calibrated for all confidence levels and test lengths, consistent with the results in Nettasinghe et al.

[2023]. In contrast, the likelihood-based approach tends to over-cover as both the test length and confidence level decrease. p. 7
Where the paper says it

We see that the CP approach remains well calibrated for all confidence levels and test lengths, consistent with the results in Nettasinghe et al. In the case of the likelihood-based approach, however, we see that as both the test length and confidence level decrease, the prediction sets tend to over-cover.

Table 1 compares the cardinality of prediction sets using CP and likelihood-based prediction for different forecast lengths, T1, and target coverage levels. p. 8

Where the paper says it

87 Table 1: Comparison of the cardinality of prediction sets using CP and likelihood-based prediction (“Like. ”) for the simulation over different prediction lengths, T1, and target coverage levels.

On average, the likelihood-based prediction sets contain fewer predicted state-sequences than the CP sets (with exception to T1 = 1 and target coverage 0. p. 8
Where the paper says it

Lastly, Table 1 presents the differences in the cardinality of the prediction sets between the conformal and likelihood-based prediction algorithms. On average, the likelihood-based prediction sets contain fewer predicted state-sequences than the CP sets (with exception to T1 = 1 and target coverage 0.

55).

Limitations of Approaches

The limitations highlight that CP is less efficient than the likelihood-based prediction when X T,..., XT are all the same state. p. 0

Where the paper says it

Williams1 , and Gudmund H. Hermansen4 1Department of Statistics, North Carolina State University 2Department of Mathematical Sciences, The University of Texas at Dallas 3Department of War Studies, Swedish Defense University 4Department of Mathematics, University of Oslo Abstract Whether or not a country is at war, or experiencing escalating or deescalating levels of conflict, has massive ramifications on a country’s national and foreign policy.

p. 0
Where the paper says it

Williams1 , and Gudmund H. Hermansen4 1Department of Statistics, North Carolina State University 2Department of Mathematical Sciences, The University of Texas at Dallas 3Department of War Studies, Swedish Defense University 4Department of Mathematics, University of Oslo Abstract Whether or not a country is at war, or experiencing escalating or deescalating levels of conflict, has massive ramifications on a country’s national and foreign policy.

When the last state differs from the dominant state in the calibration data, CP's efficiency falls dramatically.

Furthermore, for a special case where all of the calibration data are state 1, CP is less efficient than likelihood-based prediction. p. 8

Where the paper says it

2, the likelihood-based prediction sets are constructed as efficiently as possible given that the model is correctly specified in the simulation study, in the sense of minimizing the number of forecasted statesequences while guaranteeing a cumulative probability mass of at least 1−α. Although the greater cardinality suggests that CP is less efficient than the likelihood-based approach, Figure 2 indicates that the likelihood-based approach tends to over-cover due to fewer sequences accounting for most of the probability mass.

The paper presents a naive solution by introducing an artificial T + T1 + 1 time point where XT +T1+1 =

XT +T1+1 =

The CP approach is less efficient than the likelihood-based prediction when X T,..., XT are all the same state. This drawback is certainly an area for future research because ideally, we can construct CP sets that are still valid, but are as efficient as the likelihoodbased prediction sets when the calibration data is dominated by a single state. p. 9

Where the paper says it

The second part focuses on constructing prediction sets for a real forecasted state-sequence T1 time points into the future, beyond the time-frame of the observations in the data. In this case, the calibration data for each country is comprised of all observations.

Conclusion

CP sets retain coverage of plausible state-sequences across all forecasted steps, preserving the temporal dynamics that are of substantive interest in many applications. Likelihood-based prediction sets may converge to a stationary distribution (collapsing to fixed state probabilities), thereby losing the ability to reflect genuine trajectory uncertainty. p. 8

Where the paper says it

and likelihood-based approaches. We see that as the test length, T1, increases, the composition of the prediction sets produced from the likelihood-based approach converges to its stationary distribution; however, the CP approach produces a more diverse set of forecasted state-sequences, with a difference more clearly pronounced for the later time points (e.

Therefore, CP sets are better positioned to capture the true dynamics of the forecasted sequence when projecting many steps ahead. p. 6
Where the paper says it

For each combination of T1 and 1 − α, we simulate 500 state-sequences of length T + T1 according to the Markov process defined by (4). The CP and likelihood-based prediction sets are constructed using the first T time points (the calibration data), and then for each of the 500 simulations, we can determine whether or not the prediction sets [for the last T1 time points] contain the true forecasted state-sequence.

The paper is 2604. p. 0

Where the paper says it

Lastly, we comment on the possible limitations of existing approaches for applying conformal prediction to Markovian data, where the exchangeability assumption is violated. Keywords: conflict dynamics, multistate model, non-exchangeable data, state-space model, time series 1 arXiv:2604.

25139. Every sentence must end with a reference token. p. 3
Where the paper says it

, “identity permutation”) 5. For computational purposes Π := ( P, if |P| ≤ n a random sample without replacement of size n from P, if |P| > n for π ∈ Π do Calculate the non-conformity score S(π) S(π) = 1 − PT1 j=1 PPb X (π) T +j | X (π) T T1 end for 6.

The token names the page and quotes the words it rests on, copied word for word from that page. N is the page number counting from 1 — the same N as that page's --- Page N --- line.

--- Page 1 ---

Conflict Forecasting via Conformal Prediction for Markov Processes Aditya Basarkar, Emmett B. Kendall, David Randahl, Jonathan P Williams, and Gudmund H Hermansen 1Department of Statistics, North Carolina State University 2Department of Mathematical Sciences, The University of Texas at Dallas 3Department of War Studies, Swedish Defense University 4Department of Mathematics, University of Oslo Abstract Whether or not a country is at war has massive ramifications on a country’s national and foreign policy. p. 0

Where the paper says it

Williams1 , and Gudmund H. Hermansen4 1Department of Statistics, North Carolina State University 2Department of Mathematical Sciences, The University of Texas at Dallas 3Department of War Studies, Swedish Defense University 4Department of Mathematics, University of Oslo Abstract Whether or not a country is at war, or experiencing escalating or deescalating levels of conflict, has massive ramifications on a country’s national and foreign policy.

Given a country’s history of conflict, or lack thereof, future predictions about the war-status of a country are valuable information. p. 0
Where the paper says it

Hermansen4 1Department of Statistics, North Carolina State University 2Department of Mathematical Sciences, The University of Texas at Dallas 3Department of War Studies, Swedish Defense University 4Department of Mathematics, University of Oslo Abstract Whether or not a country is at war, or experiencing escalating or deescalating levels of conflict, has massive ramifications on a country’s national and foreign policy. Given a country’s history of conflict, or lack thereof, future predictions about the war-status of a country are valuable information.

In this paper, we present the use of conformal prediction on temporally-dependent data to obtain prediction sets of possible future conflict state-sequences. More specifically, we compare the results of conformal prediction to a likelihood-based prediction strategy when the data are assumed to come from a discrete-state Markov process. p. 0
Where the paper says it

In this paper, we present the use of conformal prediction on temporally-dependent data to obtain prediction sets of possible future conflict state-sequences. More specifically, we compare the results of conformal prediction to a likelihood-based prediction strategy when the data are assumed to come from a discrete-state Markov process.

A point-prediction may not supply sufficient information because the penalty for a wrong prediction is extreme, and so we consider a machine learning alternative that gives valid uncertainty quantification and is robust to model misspecification. p. 0
Where the paper says it

More specifically, we compare the results of conformal prediction to a likelihood-based prediction strategy when the data are assumed to come from a discrete-state Markov process. A point-prediction may not supply sufficient information because the penalty for a wrong prediction is extreme, and so we consider a machine learning alternative that gives valid uncertainty quantification and is robust to model misspecification.

In the data analysis, we present real forecasts of conflict dynamics across multiple countries. p. 0
Where the paper says it

A point-prediction may not supply sufficient information because the penalty for a wrong prediction is extreme, and so we consider a machine learning alternative that gives valid uncertainty quantification and is robust to model misspecification. In the data analysis, we present real forecasts of conflict dynamics across multiple countries.

Lastly, we comment on the possible limitations of existing approaches for applying conformal prediction to Markovian data, where the exchangeability assumption is violated. p. 0
Where the paper says it

In the data analysis, we present real forecasts of conflict dynamics across multiple countries. Lastly, we comment on the possible limitations of existing approaches for applying conformal prediction to Markovian data, where the exchangeability assumption is violated.

--- Page 2 ---

Motivation The last decade has seen a surge of interest in forecasting conflict dynamics within peace and conflict research. Driven by increased access to fine-grained data and advances in machine learning and computing, several large-scale forecasting projects have emerged targeting armed conflict and related phenomena such as coups, electoral violence, and regime change [see for instance Hegre et al. p. 1

Where the paper says it

1 Motivation The last decade has seen a surge of interest in forecasting conflict dynamics within peace and conflict research. Driven by increased access to fine-grained data and advances in machine learning and computing, several large-scale forecasting projects have emerged targeting armed conflict and related phenomena such as coups, electoral violence, and regime change [see for instance Hegre et al.

, 2019, 2021, Randahl et al. p. 1
Where the paper says it

Driven by increased access to fine-grained data and advances in machine learning and computing, several large-scale forecasting projects have emerged targeting armed conflict and related phenomena such as coups, electoral violence, and regime change [see for instance Hegre et al. , 2019, 2021, Randahl et al.

, 2025, Ward and Beger, 2017, Morgan et al. p. 1
Where the paper says it

, 2019, 2021, Randahl et al. , 2025, Ward and Beger, 2017, Morgan et al.

, 2019]. Early efforts focused on point predictions, but there is growing recognition that uncertainty quantification is equally critical. p. 1
Where the paper says it

, 2025, Ward and Beger, 2017, Morgan et al. Early efforts focused on point predictions, but there is growing recognition that uncertainty quantification is equally critical: incorrect predictions carry severe consequences, and decision-makers require not just a single forecast but a calibrated sense of the range of plausible outcomes [Hegre et al.

This has motivated a shift toward probabilistic forecasting methods and prediction intervals/sets. p. 1
Where the paper says it

, 2025, Randahl et al. This has motivated a shift toward probabilistic forecasting methods and prediction intervals/sets.

Among the models that have shown promise for conflict dynamics are Markov models, which represent the evolution of a conflict system as a sequence of states transitioning over time. p. 1
Where the paper says it

This has motivated a shift toward probabilistic forecasting methods and prediction intervals/sets. Among the models that have shown promise for conflict dynamics are Markov models, which represent the evolution of a conflict system as a sequence of states transitioning over time [Randahl and Vegelius, 2022, Williams et al.

Combined with estimated conditional distributions of fatality counts given the current state, these models can generate forecasts of future conflict trajectories. p. 1
Where the paper says it

Among the models that have shown promise for conflict dynamics are Markov models, which represent the evolution of a conflict system as a sequence of states transitioning over time [Randahl and Vegelius, 2022, Williams et al. Combined with estimated conditional distributions of fatality counts given the current state, these models can generate forecasts of future conflict trajectories [Randahl and Vegelius, 2024, Hegre et al.

However, their uncertainty estimates are sensitive to model misspecification, a serious concern given the difficulty of correctly characterizing conflict processes. Conformal prediction (CP) offers a principled alternative. By requiring only an exchangeability assumption rather than a correctly specified model, CP produces prediction sets with finite-sample marginal coverage guarantees. p. 1
Where the paper says it

Conformal prediction (CP) offers a principled alternative. By requiring only an exchangeability assumption rather than a correctly specified model, CP produces prediction sets with finite-sample marginal coverage guarantees.

--- Page 3 ---

Background CP offers a robust approach to construct prediction sets that obtain valid coverage for exchangeable data [see Shafer and Vovk, 2008, Vovk et al., 2022, for more on CP]. p. 1

Where the paper says it

This has motivated a shift toward probabilistic forecasting methods and prediction intervals/sets. Among the models that have shown promise for conflict dynamics are Markov models, which represent the evolution of a conflict system as a sequence of states transitioning over time [Randahl and Vegelius, 2022, Williams et al.

However, CP relies on the assumption that the data are exchangeable, which in many applications, need not be the case. p. 1
Where the paper says it

, 2023], and more. However, CP relies on the assumption that the data are exchangeable, which in many applications, need not be the case.

The focus for this work is on time series data, particularly Markov processes, where sequential observations are temporally dependent. p. 2
Where the paper says it

not inherently exchangeable. The focus for this work is on time series data, particularly Markov processes, where sequential observations are temporally dependent.

--- Page 4 ---

Methodological developments in CP for time series data has a well-studied literature [e.g., Xu and Xie, 2021, 2023, Zheng and Proutiere, 2024]. p. 2

Where the paper says it

Methodological developments in CP for time series data has a well-studied literature [e. , Xu and Xie, 2021, 2023, Zheng and Proutiere, 2024, Dua et al.

Many of these approaches focus on developing provably valid, or approximately valid, methods for data that exhibit dependencies which violate the exchangeability assumption. p. 2
Where the paper says it

, 2025, Neglia et al. Many of these approaches focus on developing provably valid, or approximately valid, methods for data that exhibit dependencies which violate the exchangeability assumption.

One example is the development of adaptive CP, which calibrates the guaranteed confidence level in an online setting to account for on-going distributional changes in the data. p. 2
Where the paper says it

Many of these approaches focus on developing provably valid, or approximately valid, methods for data that exhibit dependencies which violate the exchangeability assumption. One example is the development of adaptive CP, which calibrates the guaranteed confidence level in an online setting to account for on-going distributional changes in the data [Gibbs and Candes, 2021, Zaffran et al.

Other approaches seek to re-frame where the notion of exchangeability is applied. p. 2
Where the paper says it

, 2022, Su et al. Other approaches seek to re-frame where the notion of exchangeability is applied.

This latter approach can be split into two different frameworks, specifically for handling discrete-state Markov processes. p. 2
Where the paper says it

Other approaches seek to re-frame where the notion of exchangeability is applied. This latter approach can be split into two different frameworks, specifically for handling discrete-state Markov processes.

The first idea is to treat each state-sequence as an individual observation, and then multiple observations of these state-sequences are assumed to be independent and thus exchangeable. p. 2
Where the paper says it

This latter approach can be split into two different frameworks, specifically for handling discrete-state Markov processes. The first idea is to treat each state-sequence as an individual observation, and then multiple observations of these state-sequences are assumed to be independent and thus exchangeable [Cherubin and Nouretdinov, 2016, Stankeviciute et al.

The second idea leverages properties of Markov processes and permutation-groups to arrive at a notion of partial exchangeability for a single state-sequence (more detail in Section 3. p. 2
Where the paper says it

The first idea is to treat each state-sequence as an individual observation, and then multiple observations of these state-sequences are assumed to be independent and thus exchangeable [Cherubin and Nouretdinov, 2016, Stankeviciute et al. However, instead of relying on multiple realizations of independent state-sequences, the second idea leverages properties of Markov processes and permutation-groups to arrive at a notion of partial exchangeability for a single state-sequence [Chernozhukov et al.

1).

--- Page 5 ---

Methodology Let X t T t=1 denote a discrete-state Markov process observed over T > 0 time points, where Xt ∈ X = the set of states = the set of states = m in Z+. We refer to X t T t=1 as the calibration sequence/data. p. 2

Where the paper says it

, m}, m ∈ Z +. We refer to {Xt} T t=1 as the calibration sequence/data.

The ultimate goal for this work is to forecast state-sequences T1 > 0 time points into the future. p. 2
Where the paper says it

We refer to {Xt} T t=1 as the calibration sequence/data. The ultimate goal for this work is to forecast state-sequences T1 > 0 time points into the future.

Rather than constructing a point-prediction (i. p. 2
Where the paper says it

The ultimate goal for this work is to forecast state-sequences T1 > 0 time points into the future. Rather than constructing a point-prediction (i.

p. 2
Where the paper says it

The ultimate goal for this work is to forecast state-sequences T1 > 0 time points into the future. Rather than constructing a point-prediction (i.

e., one future state-sequence of length T1), our goal instead is to construct a prediction set, for any level 1 − α ∈ (0, 1), of state-sequences, say C 1−α, that guarantees the true underlying state-sequence is an element of the set with probability at least 1 − α. p. 2
Where the paper says it

Rather than constructing a point-prediction (i. , one future state-sequence of length T1), our goal instead is to construct a prediction set, for any level 1 − α ∈ (0, 1), of state-sequences, say C1−α, that guarantees the true underlying state-sequence is an element of the set with probability at least 1 − α.

p. 2
Where the paper says it

Rather than constructing a point-prediction (i. , one future state-sequence of length T1), our goal instead is to construct a prediction set, for any level 1 − α ∈ (0, 1), of state-sequences, say C1−α, that guarantees the true underlying state-sequence is an element of the set with probability at least 1 − α.

Pruning redundant reasoning tokens

The ultimate goal for this work is to forecast state-sequences T1 > 0 time points into the future rather than constructing a point-prediction (i.e., one future state-sequence of length T1). p. 2

Where the paper says it

Rather than constructing a point-prediction (i. , one future state-sequence of length T1), our goal instead is to construct a prediction set, for any level 1 − α ∈ (0, 1), of state-sequences, say C1−α, that guarantees the true underlying state-sequence is an element of the set with probability at least 1 − α.

Rather than constructing a point-prediction (i.e., one future state-sequence of length T1), our goal instead is to construct a prediction set, for any level 1 − α ∈ (0, 1), of state-sequences, say C 1−α, that guarantees the true underlying state-sequence is an element of the set with probability at least 1 − α.

How the training signal is built

The calibration sequence/data X t T t=1 denote a discrete-state Markov process observed over T > 0 time points, where Xt ∈ X = the set of states = the set of states = m in Z+. The state-label Xi,t is determined by the following rules. p. 5

Where the paper says it

, 191}, at month t. The assigned state-label, Xi,t, is determined by the following rules: Xi,t =    1, if Yi,t−1 = Yi,t = 0 for t > 1, or Yi,t = 0 for t = 1 2, if Yi,t−1 = 0 and Yi,t > 0 for t > 1 3, if Yi,t−1 > 0 and Yi,t > 0 for t > 1, or Yi,t > 0 for t = 1 4, if Yi,t−1 > 0 and Yi,t = 0 for t > 1.

Where the gains come from

The CP algorithm involves:

  1. Set XT +T+1 = xT +T+1,. p. 2

    Where the paper says it

    , 2023] (more detail in Section 3. In either case, the output of the CP algorithm remains a prediction set of forecasted state-sequences.

    .., XT +T+T1 = xT +T+T1.

  2. Using

P i i,j = P T + T-1

t=1

(Xt = i∧ Xt+1 = j) / (PT + T-1 t=1

t=1 (Xt = i)).

  1. Determine the set of permutable i-blocks, I = p. 3
    Where the paper says it

    Using {Xt} T +T1 t=1 , estimate the transition probability matrix (P ) by h Pb i i,j = PT +T1−1 t=1 1(Xt = i ∧ Xt+1 = j) PT +T1−1 t=1 1(Xt = i) 3. Determine the set of permutable i-blocks, I = {I1, .

I i, given XT +T+T1 = xT +T+T1 = i.

  1. Let P be the permutation group of I where each ρ ∈ P results in

X(ρ) t T + T1 t=1 =

(I i(ρ) 1,..., I i(ρ) D, ID+1), X T +T+T1 t=1 = i.

  1. For computational purposes Π:=

Π:= (P, if P ≤ n a random sample without replacement of size n from P, if P > n for π ∈ Π do Calculate the non-conformity score S(π) =

S(π) = 1 − PT1 j=1 PPb X(π) T +j X(π) T T1 / T.

  1. Calculate the p-value qb(x) =

The CP algorithm involves:

  1. Set XT +T+T1 = xT +T+T1,. p. 2

    Where the paper says it

    , 2023] (more detail in Section 3. In either case, the output of the CP algorithm remains a prediction set of forecasted state-sequences.

    .., XT +T+T1,..., XT + T i T i.

  2. Using

P i i,j = P T + T-1 t=1

(Xt = i)).

  1. Determine the set of permutable i-blocks, I =

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improve uncertainty quantification for conflict forecasts by using Conformal Prediction to obtain prediction sets of possible future conflict state-sequences that are robust to model misspecification and provide finite-sample marginal coverage guarantees.

  2. Enhance trajectory forecasting accuracy by comparing Conformal Prediction against likelihood-based prediction strategies, noting that the CP approach remains well calibrated for all confidence levels and test lengths while the likelihood-based approach tends to over-cover.

  3. Improve long-term conflict forecasting by leveraging the distinction between the two methods, as CP sets retain coverage of plausible state-sequences across all forecasted steps, preserving the temporal dynamics that are of substantive interest in many applications, unlike likelihood-based sets which converge quickly to a stationary distribution.

  4. Address model misspecification risks by utilizing Conformal Prediction's reliance only on an exchangeability assumption rather than a correctly specified model, making it more robust when applied to Markov processes.

  5. Improve performance in data-sparse regimes by investigating the limitation where CP is less efficient than the likelihood-based prediction when X1, . . . , XT are all the same state, and exploring solutions like introducing an artificial time point to improve efficiency in such cases.

Abstract

Whether or not a country is at war, or experiencing escalating or deescalating levels of conflict, has massive ramifications on a country's national and foreign policy. Given a country's history of conflict, or lack thereof, future predictions about the war-status of a country are valuable information. In this paper, we present the use of conformal prediction on temporally-dependent data to obtain prediction sets of possible future conflict state-sequences. More specifically, we compare the results of conformal prediction to a likelihood-based prediction strategy when the data are assumed to come from a discrete-state Markov process. A point-prediction may not supply sufficient information because the penalty for a wrong prediction is extreme, and so we consider a machine learning alternative that gives valid uncertainty quantification and is robust to model misspecification. In the data analysis, we present real forecasts of conflict dynamics across multiple countries. Lastly, we assess the differences between the conformal and likelihood-based prediction sets applied to training data which fall under a pervasive edge-case, not discussed in existing literature.

Sources

Related papers