CrisisLandMark: Quantitative Crisis Indicators

• CrisisLandMark is a framework that defines crisis markers using quantitative observables across networks, time series, and remote sensing data.
• It employs diverse methodologies such as k‐shell decomposition, topological data analysis, recurrence quantification, and copula-based regime detection to highlight crisis structure.
• It provides actionable early-warning signals for financial, economic, and Earth system crises, while also revealing methodological limitations in static and dynamic models.

Searching arXiv for "CrisisLandMark" and closely related papers to ground the article. CrisisLandMark is a term used in several distinct research settings to denote quantitative markers of crisis structure, crisis onset, crisis propagation, or crisis-oriented retrieval. In the global-economic contagion literature, it denotes a country in the network nucleus identified by $k$-shell decomposition and therefore a maximal “super-spreader” of crisis (Garas et al., 2010). In financial time-series and network analysis, the term is reused for early-warning signatures extracted from persistence landscapes, recurrence quantification, copula geometry, interbank motifs, and cross-border portfolio networks (Gidea et al., 2017, Piskun et al., 2011, Chalkis et al., 2021, Squartini et al., 2013, Joseph et al., 2013). In Earth-system modeling, it denotes mathematically defined threshold crossings and syndrome indices in a coupled sociopolitical-event–planetary-boundary framework (Bertolami et al., 22 Jul 2025). In remote sensing, CrisisLandMark is the name of a large-scale multimodal corpus of Sentinel-1 and Sentinel-2 imagery with structured crisis and land-cover annotations, together with associated contrastive retrieval models (Cambrin et al., 14 Jul 2025). The shared theme across these usages is the operationalization of crisis-relevant structure by explicit observables rather than by informal qualitative description.

1. Global-economic contagion and the original “crisis landmark”

The 2010 study “Worldwide spreading of economic crisis” defines the concept most directly. It constructs two weighted global-economic networks for 2007: a Corporate Ownership Network (CON) with 206 countries, where the directed link weight $w_{ij}$ is the number of subsidiary corporations in country $j$ owned by parent companies in country $i$; and an International Trade Network (ITN) with 82 countries, where the directed link weight $E_{ij}$ is the dollar value of exports from country $i$ to country $j$. For each country pair, the paper forms undirected total weights by summing the two directions, and defines node strength as total economic weight in CON and as $GDP_i$ in ITN. The reported empirical result is that node strength correlates nearly linearly with official GDP. Crisis propagation is then modeled with an SIR process in which infection probability depends on crisis magnitude $m$, bilateral economic tie strength, and the total strength of the target country (Garas et al., 2010).

The key structural device is the $k$-shell decomposition, described step by step following Pittel et al. 1996. Nodes of low current degree are iteratively removed, assigned shell indices $w_{ij}$0, and the highest shell defines the nucleus. In this framework, a CrisisLandMark is any country in that nucleus. The concept is therefore topological rather than merely size-based: spreading power is assigned by shell depth, not by GDP rank alone (Garas et al., 2010).

Network	Nucleus shell	Countries
CON	$w_{ij}$1	US, GB, FR, DE, NL, JP, SE, IT, CH, ES, BE, LU
ITN	$w_{ij}$2	CN, RU, JP, ES, GB, NL, IT, DE, BE, LU, US, FR

Several features make these countries “crisis landmarks” in the original sense. They remain in the deepest shell under broad weight-threshold variations; they have very high average degree; and they combine dense internal ties with strong connectivity to the rest of the world. A notable empirical result is that only six of the 12 CON nucleus countries are large economies, while the other six are medium or small economies that are heavily internationally invested. The paper highlights Belgium as a counterintuitive case: despite much lower GDP than the USA, it can initiate a global crisis in the model. In simulations, outbreaks starting in inner-shell countries reach larger global fractions at much smaller crisis magnitude, Belgium can infect more than 50% of countries for $w_{ij}$3 and up to 95% in worst-case runs, whereas outer-shell origins require $w_{ij}$4–15 to approach similar infection levels. Error bars across nodes in the same shell are reported to be small, implying that shell index is the dominant predictor of spreading power (Garas et al., 2010).

The limitations are equally explicit. The analysis uses static 2007 snapshots, assumes one-step infection with immediate recovery, omits nonlinear contagion effects beyond the specified infection rule, sets the directionality threshold $w_{ij}$5, and excludes financial loans, banking exposures, policy interventions, bailouts, and second-order cascade effects. This makes CrisisLandMark, in its original formulation, a structural contagion concept rather than a complete macro-financial crisis model (Garas et al., 2010).

2. Interbank and cross-border network indicators

A second usage treats CrisisLandMark as a monitoring architecture for systemic-risk topology. In the Dutch interbank network from 1998 to 2008, the relevant observables are link density, reciprocity, degree heterogeneity, and directed cycle counts. The analysis controls first for density with the Directed Random Graph (DRG), then for node-specific in- and out-degrees with the Directed Configuration Model (DCM), and then for out-only, in-only, and reciprocated dyads with the Reciprocal Configuration Model (RCM). Under these controls, the empirical evolution separates into three phases: a stationary phase from 1998 to 2004, a pre-crisis build-up from 2005 to 2007, and a crisis collapse in 2008. Starting in Q1 2005, dyadic $w_{ij}$6-scores under DCM depart smoothly from zero, reciprocated dyads become under-represented, single dyads become over-represented, and the unreciprocated 3-loop motif 9 under RCM swings from $w_{ij}$7 to $w_{ij}$8, signalling anomalous circular lending loops. In 2008, reciprocity drops by about 40% and moves almost $w_{ij}$9 below its long-run mean. A simple CUSUM or moving-window test identifies Q1 2005 as the first break and Q1 2008 as the second. A central conclusion is that reconstructions from partial bank-specific data miss these early signals because, under DCM or its weighted analog, the relevant anomalies are “washed out” (Squartini et al., 2013).

The same details propose a real-time “CrisisLandMark” blueprint: build and update the directed adjacency matrix $j$0, compute density, reciprocity, degrees and their variance, calibrate DCM and RCM parameters, evaluate motif counts and $j$1-scores, and apply change-point tests such as sequential CUSUM or Bayesian filters. Suggested alarm levels are $j$2 for a warning, $j$3 or simultaneous extreme dyadic $j$4-scores under DRG for escalation, with secondary triggers including sustained degree-variance increase or reciprocity $j$5. The policy implication is explicit: full disclosure of interbank exposures is essential, because aggregated or reconstructed data hide triadic risk loops (Squartini et al., 2013).

A related network-based formulation appears in the analysis of cross-border portfolio investment networks from 2002 to 2012. Here the first landmark is the algebraic connectivity of the equity securities network, taken as a measure of structural robustness. In the equity PIN, this quantity remains around $j$6–$j$7 from 2002 to 2004, falls to $j$8 in early 2005, and reaches an all-time low near $j$9 in early 2007. Fiedler-vector bi-section indicates that a small cluster comprising about 10–15% of the E-PIN can be cut by removing only weak links, and offshore financial centres are over-represented in the fragile subgraph with $i$0–2.0. The second landmark is the percolation edge density of the long-term debt PIN at a threshold of $i$1 million USD, which is found to describe and forecast the proliferation of OTC derivatives. For CDS notional outstanding amounts, the best-fit non-linear short-term memory model gives $i$2, $i$3, a lead time $i$4 months, and Pearson $i$5; for equity-linked derivatives gross-market value, $i$6, $i$7, $i$8, and $i$9. GDP-linked warning thresholds $E_{ij}$0 and $E_{ij}$1 are breached in early 2007 (Joseph et al., 2013).

Taken together, these studies make CrisisLandMark a network-theoretic vocabulary for structural fragility, anomalous reciprocity, cyclic debt loops, and market interdependence. The objects being tracked are not prices but mesoscopic and macroscopic graph observables.

3. Topological data analysis of financial time series

In the topological-data-analysis literature, Gidea and Katz recast crisis landmarks as transient topological patterns in sliding-window point clouds extracted from multidimensional return series. Let $E_{ij}$2 be the vector of returns on day $E_{ij}$3, and let $E_{ij}$4 be a rolling window of length $E_{ij}$5. For each $E_{ij}$6, one builds a Vietoris–Rips filtration, computes $E_{ij}$7, forms the persistence diagram $E_{ij}$8, converts it to persistence landscapes $E_{ij}$9, and evaluates the $i$0-norm $i$1. The resulting time series $i$2 is then processed by a second rolling window to compute variance, low-frequency spectral density, and lag-1 autocorrelation (Gidea et al., 2017).

The empirical claim is precise. For the daily returns of four major US stock market indices during the technology crash of 2000 and the financial crisis of 2007–2009, the $i$3-norms exhibit strong growth prior to the primary peak that ascends during a crash. Moreover, the average spectral density at low frequencies of the $i$4-norm series demonstrates a strong rising trend for 250 trading days prior to either the dotcom crash on 03/10/2000 or the Lehman bankruptcy on 09/15/2008. The reported Kendall $i$5 for rising variance and rising low-frequency power is approximately $i$6–$i$7. Robustness checks show that $i$8 and $i$9 both recover pre-crash growth, that $j$0 and $j$1 give qualitatively similar results, and that statistical windows $j$2–500 days are sufficient to detect monotonic trends. Synthetic tests on a noisy Hénon map with slowly drifting parameter, white noise with increasing variance, and Gaussian mixtures with Gamma-distributed precision also show sharp increases in $j$3 when the system becomes more chaotic or more volatile (Gidea et al., 2017).

Within this framework, “crisis landmarks” are not dates attached to single threshold rules but sharp, persistent rises in $j$4 and its low-frequency content. A plausible implication is that the method is designed to detect changes in the geometry of return clouds rather than only changes in marginal volatility.

4. Recurrence quantification and copula-based regime detection

Recurrence Quantification Analysis (RQA) offers a different operationalization. The relevant object is the recurrence plot built from reconstructed phase-space vectors $j$5, with recurrence matrix $j$6. For major equity indices, the reported empirical parameter choices are embedding $j$7, delay $j$8, threshold $j$9, and sliding window size $GDP_i$0 days. Among standard RQA measures, laminarity $GDP_i$1 is identified as the most sensitive to volatility changes and hence to bubble formation and crashes. Crisis landmarks are then defined as instability onset, critical point, crash onset, and full relaxation, all read from threshold crossings of the $GDP_i$2 curve (Piskun et al., 2011).

The case studies are highly explicit. For the 1929 DJI, the normal regime is $GDP_i$3, instability begins at $GDP_i$4 on 04 May 1928 with $GDP_i$5, the critical point occurs at $GDP_i$6 on 26 Mar 1929 with $GDP_i$7, the crash minimum is near $GDP_i$8 with $GDP_i$9, and full relaxation occurs at $m$0 on 18 Aug 1932. For the 1987 DJI, $m$1, instability begins about 250 trading days before the crash, the critical point about 25 days before, and crash $m$2. For the 2000 NASDAQ, $m$3, instability begins about 350 trading days before, the critical point about 11 trading days before, and crash $m$4. For the 2007–2010 global crisis, the paper reports no single-day crash; instead, $m$5 declines smoothly, with crisis onset for the DJI on 28 Aug 2007 and relaxation onset on 23 Apr 2009 (Piskun et al., 2011).

A separate line of work models crisis periods through empirical copulae built from rolling windows of asset returns. Daily prices $m$6 are converted to log-returns, rolling windows of length $m$7 are used to estimate sample means and covariances, and the joint law of $m$8 and $m$9 is approximated by drawing $k$0 uniformly from the simplex. After empirical-CDF transformation and binning on an $k$1 grid, one obtains a copula density matrix $k$2. The crisis indicator is the ratio of mass in a down-diagonal band to mass in an up-diagonal band; in normal times $k$3, while in crisis times $k$4. Entire copulae are then clustered using $k$5 or Earth-Mover’s distance, affinity $k$6, spectral embedding, and $k$7-medoids or $k$8-means. Clusters are ranked by median $k$9, with highest-median clusters treated as crisis states (Chalkis et al., 2021).

This framework is also compressed to a low-dimensional regression approximation using at most 10% of copula bins. Using only a $w_{ij}$00 block, or 9 cells out of a $w_{ij}$01 copula, the linear model yields RMSE about $w_{ij}$02 and crisis-detection accuracy above 95%, with identical landmark calls except one mid-2018 warning. The empirical summary reports that all four known crises in French industrial stocks from 1929 to 2020 were detected with true positive rate 100%, false positive rate about 2% of all windows, and average lead time 10 trading days; all three major cryptocurrency crashes from 2014 to 2020 were also detected with true positive rate 100% and false positive rate about 0% (Chalkis et al., 2021).

These two approaches differ materially from the original network-contagion definition. RQA attaches landmarks to changes in recurrence structure and laminarity; copula clustering attaches landmarks to regime geometry in the dependence structure of returns and covariances.

5. Coupled sociopolitical events, planetary boundaries, and syndromes

In the 2025 Bertolami–Elísio framework, CrisisLandMark becomes part of a formal theory of coupled sociopolitical events (SPEs) and planetary boundaries (PBs). The PB variables $w_{ij}$03 evolve according to continuous logistic equations $w_{ij}$04, with explicit Verhulst solution, while a full-blown crisis in a single SPE component is modeled exponentially as $w_{ij}$05 with $w_{ij}$06. The coupling is introduced through a real $w_{ij}$07 matrix $w_{ij}$08, $w_{ij}$09, inside a Lagrangian and corresponding Hamiltonian. In second-order form, the coupled dynamics are $w_{ij}$10 and $w_{ij}$11. Positive $w_{ij}$12 means alignment in a “benign” direction, negative $w_{ij}$13 means misalignment that pushes a PB further out of safe limits, and $w_{ij}$14 gauges coupling strength (Bertolami et al., 22 Jul 2025).

The paper’s crisis logic is threshold-based and dynamical. If an SPE grows exponentially, then even a small $w_{ij}$15 and negative $w_{ij}$16 generate exponential runaway in $w_{ij}$17. Conversely, inserting logistic $w_{ij}$18 into the $w_{ij}$19 equation yields asymptotic $w_{ij}$20, so if $w_{ij}$21 the SPE component grows like $w_{ij}$22, producing what the paper calls a “runaway polycrisis.” Linear stability is tied to the eigenvalues of the block Jacobian $w_{ij}$23: if $w_{ij}$24 has a positive eigenvalue $w_{ij}$25, then $w_{ij}$26 has a real positive root $w_{ij}$27, implying instability, and the bifurcation threshold occurs when the largest eigenvalue crosses zero (Bertolami et al., 22 Jul 2025).

The paper then defines a taxonomy. A single-component crisis occurs when $w_{ij}$28. A polycrisis occurs when at least two SPE components grow simultaneously. A $w_{ij}$29-th order Earth System syndrome is the vector product $w_{ij}$30, whose norm is a hyper-volume index quantifying how entangled those $w_{ij}$31 crises are. Crisis landmarks are the logistic inflection point of a PB, the time when an SPE component crosses a crisis threshold $w_{ij}$32, the earliest simultaneous crossing of two crisis thresholds for polycrisis, and the crossing of an alarm level $w_{ij}$33 by $w_{ij}$34. In the illustrative parameter choice $w_{ij}$35, $w_{ij}$36, $w_{ij}$37, and $w_{ij}$38, the induced SPE satisfies $w_{ij}$39, implying asymptotic $w_{ij}$40, and a threshold $w_{ij}$41 yields $w_{ij}$42 years (Bertolami et al., 22 Jul 2025).

This is the most explicitly theoretical use of the term. Here CrisisLandMark is neither a country nor a data set, but a mathematical apparatus of inflections, threshold crossings, eigenvalue tests, and syndrome norms.

6. CrisisLandMark as a multimodal remote-sensing corpus

In remote sensing, CrisisLandMark names a large-scale corpus rather than an early-warning indicator. The dataset comprises approximately 647,000 image patches, evenly split between Sentinel-1 GRD SAR and Sentinel-2 L2A optical multispectral imagery. Sentinel-1 is represented by VV and VH channels resampled to 10 m ground sampling distance, while Sentinel-2 provides 12 atmospherically corrected bands, also at 10 m. The corpus draws from five repositories: re-BEN, CaBuAr, QuakeSet, MMFlood, and Sen12Flood. After uniformly tiling scenes into $w_{ij}$43 pixels and aligning spatial grids, the data are split by stratified multi-label sampling into 20% training images and 80% retrieval images (Cambrin et al., 14 Jul 2025).

The label system is equally central. re-BEN uses the 43-category CORINE system, while the crisis-event repositories use the 9-category Dynamic World system. All land-cover labels are mapped to the 9 Dynamic World classes—Trees, Crops, Shrub & Scrub, Water, Grass, Built, Flooded Vegetation, Bare, Snow & Ice—through a deterministic lookup table, and crisis-event classes Flooded Area, Earthquake Damage, and Burned Area are retained. Each patch therefore carries a multi-hot vector of up to 12 labels. The authors enumerate 2,047 distinct textual queries formed by every non-empty subset of co-occurring labels and define graded relevance by label-set IoU, rounded to a 0–10 scale. A $w_{ij}$44 test reports statistical equivalence between the training split and retrieval corpus with $w_{ij}$45 (Cambrin et al., 14 Jul 2025).

The associated retrieval framework, CLOSP, is a CLIP-inspired three-tower architecture with a MiniLM text encoder, an optical encoder, and a SAR encoder. All outputs are linearly projected and $w_{ij}$46-normalized into a shared embedding space, and training uses symmetric InfoNCE image–text losses. Because Sentinel-1 and Sentinel-2 rarely share exact timestamps, no direct image–image loss is imposed; text acts as the bridge between modalities. GeoCLOSP adds a fourth tower, a location encoder built from a SIREN network with spherical-harmonic positional encoding, and two extra contrastive terms linking image and geographic embeddings (Cambrin et al., 14 Jul 2025).

The reported quantitative gains are substantial. The best RGB-only baseline, SkyCLIP-T, reaches nDCG@1000 of about 37.9%, while CLOSP-ResNet achieves 56.2%, an absolute gain of 18 percentage points or about 54% relative improvement. GeoCLOSP further raises nDCG@1000 to 57.8%. On Sentinel-1 only, CLOSP-ViT-Large reaches 55.7% against 37.4% for BiCLIP. GeoCLOSP also raises the Spearman correlation between image-embedding distance and geographic distance from 0.13 to 0.34. The paper’s interpretation is that CLOSP excels at general semantic tasks, whereas GeoCLOSP becomes a specialized expert for retrieving location-dependent crisis events and rare geographic features, with the trade-off that geographic specificity can slightly reduce performance on ubiquitous classes such as trees (Cambrin et al., 14 Jul 2025).

This usage differs categorically from the earlier ones. CrisisLandMark here is a machine-readable, crisis-aware Earth observation resource for text-to-image retrieval, not a crisis-propagation landmark or a threshold-crossing alarm.

7. Conceptual unity, divergence, and recurring limitations

Across these literatures, the term denotes different kinds of objects: nucleus countries in a weighted world-economy graph, motif or connectivity anomalies in financial networks, geometric or recurrence signatures in market data, threshold crossings in a dynamical SPE–PB model, and a multimodal satellite corpus with contrastive retrieval models. The commonality is methodological rather than ontological: each usage turns crisis salience into an observable that can be ranked, thresholded, or embedded (Garas et al., 2010, Squartini et al., 2013, Gidea et al., 2017, Cambrin et al., 14 Jul 2025).

The divergence is equally important. In the economic-network origin, CrisisLandMark is fundamentally about spreading power under SIR dynamics and $w_{ij}$47-shell centrality (Garas et al., 2010). In interbank and portfolio networks, it is about structural robustness, reciprocity, cycles, and interdependence (Squartini et al., 2013, Joseph et al., 2013). In TDA, RQA, and copula-based approaches, it is about topological or statistical state changes in financial time series (Gidea et al., 2017, Piskun et al., 2011, Chalkis et al., 2021). In the Bertolami–Elísio formalism, it is about dynamical thresholds and runaway criteria in a coupled nine-dimensional system (Bertolami et al., 22 Jul 2025). In remote sensing, it is a curated data substrate and retrieval benchmark (Cambrin et al., 14 Jul 2025).

Several recurrent limitations appear across these formulations. Static or weakly dynamic assumptions are common: the 2010 world-economy model uses static 2007 snapshots and excludes policy feedbacks (Garas et al., 2010); the interbank analysis shows that partial reconstructions erase early-warning content (Squartini et al., 2013); the remote-sensing corpus notes geographic bias and the absence of temporal modeling (Cambrin et al., 14 Jul 2025). More broadly, these works define operational crisis landmarks, but the operational variable depends entirely on the modeling layer chosen—network topology, persistence geometry, recurrence structure, copula shape, dynamical thresholds, or retrieval relevance. A plausible implication is that “CrisisLandMark” functions less as a single settled technical term than as a family resemblance across crisis analytics: a preference for explicit, monitorable markers of instability, spread, or crisis-related structure.