Extracted main text — title through conclusion, appendix excluded. This is what our citation measures are computed over, published so the extraction can be checked by eye.
81,541 characters · 39 sections · 53 citation commands
Hall-Like Transversal Stress and Sandpile Criticality on Real Production Networks
Economic crises rarely propagate along a single channel. An energy shock may surface as inflation, fiscal strain, food insecurity or industrial contraction; a monetary shock may emerge as liquidity stress, sovereign-risk repricing or default clustering; a geopolitical shock may show up as supply-chain disruption, commodity-price pressure or topological reconfiguration of trade flows; and a technological restriction may appear as bottlenecks deep inside production networks. A common feature of these episodes is that the dimension in which the shock is delivered need not coincide with the dimension in which it is felt. Economic systems do not merely transmit shocks along edges of a network: they convert shocks across dimensions.
The empirical record offers ample illustration. The 2008--2009 financial crisis began as a balance-sheet shock in a narrow segment of the U.S.\ housing market and ended as a global recession with synchronised contractions in trade, credit and employment. The 2011 Great East Japan earthquake produced a localised supply-chain disruption whose effects propagated through specific automotive and electronics sectors well beyond the immediate seismic zone carvalho2021supply. The 2020 pandemic combined a labour-supply shock, a logistics shock and a demand-composition shock, with heterogeneous effects across sectors that could not be captured by any single aggregate variable. The 2022 energy and food-price episode linked a geopolitical shock to inflation in distant regions through food-and-fertiliser markets that had previously been treated as loosely coupled. None of these episodes is well described as a linear pass-through along a single network channel; each involves a cross-dimensional transformation of the original perturbation, with amplification depending on which nodes are flow-intensive, low-redundancy and low-capacity at the moment the field is applied.
Most network-based models of contagion in economics are longitudinal: a perturbation enters a node and propagates along weighted directed edges, with amplification governed by topology and balance-sheet linkages acemoglu2012network,carvalho2019production, elliott2014financial,gai2010contagion,glasserman2016contagion. This class of models has clarified how the architecture of input--output linkages shapes aggregate fluctuations and financial contagion. It is, however, less well suited to representing situations in which an external field, rather than a localised shock at a single node, loads stress simultaneously on many exposed flows and converts that loading into pressure of a different nature.
In this paper we propose a complementary mechanism, drawing a disciplined structural analogy with the Hall effect of solid-state physics hall1879new. In the physical Hall effect, a current flowing through a conductor under an external magnetic field generates a transverse voltage. In economic networks, exposed flows operating under an external field can generate stress in a different dimension from the original flow. The analogy is structural and does not imply that economies obey electromagnetic laws. We use it as an organising device to formalise how external loading interacts with flow exposure, redundancy and absorptive capacity to produce transversal stress on a node.
We embed this Hall-like stress conversion into a sandpile threshold dynamics in the spirit of bak1987self and bak1988self. Each node in a real production network accumulates stress over time, dissipates a fraction of it, and topples when its internal stress crosses a node-specific threshold. Hall-like transversal stress enters the dynamics by lowering the effective toppling threshold: the larger the loading on an exposed, low-redundancy node, the easier it is for the node to activate. The substrate on which this dynamics runs is the real annual World Input--Output Database (WIOD) network of intermediate transactions between country--sector pairs timmer2015illustrated, timmer2016anatomy for 2000--2014, with 2014 as the main simulation year.
The paper is positioned at the intersection of econophysics, complex networks, production networks, systemic risk and nonlinear dynamics. It contributes a tractable computational mechanism for shock conversion under uncertainty, nonlinearity and structural fragility, distinct from but complementary to curvature-based “Sandpile Economics” approaches that emphasise geometric fragility of the network itself vallarino2024sandpile. Where curvature-based arguments explain why a network may be structurally fragile, the Hall-Sandpile mechanism explains how external fields activate that fragility through transversal stress.
The empirical object of the paper is the observed production-network topology. The shock fields are simulated in a controlled statistical-mechanics experiment. We do not claim to recover any historical crisis episode, nor do we attempt to estimate the causal effect of a specific past shock. We instead study how a real production network responds to controlled nonlinear loading. Three alternative propagation operators ensure that nonlinear avalanche results do not arise mechanically from a row-stochastic transition matrix with unit spectral radius.
Our main findings are fourfold. First, the row-share spectral radius of WIOD lies near unity by construction, while the leakage-adjusted spectral radius lies around $0.33$--$0.34$ and the max-row-normalised spectral radius rises from about $0.22$ in 2000 to about $0.32$ in 2014; the dissipative operator is therefore well below criticality on its own and the dynamics is not driven by a unit-radius operator. Second, relative Hall-like exposure is concentrated in a small upper tail of country--sector nodes, with the top exposures dominated by Chinese production sectors and water-utility nodes in the 2014 calibration; this concentration is a feature of the network calibration, not a sectoral diagnosis. Third, controlled Monte Carlo experiments produce four ordered regimes---stable absorption, latent fragility, critical transition and avalanche---with mean avalanche size and threshold-event probabilities rising jointly with external field intensity and redundancy stress. Fourth, tail diagnostics show regime-dependent thickening of the avalanche distribution, but the estimated tail exponents (around $5.9$--$6.2$ in the most active regimes) are too steep to support a claim of universal power-law behaviour. We therefore describe the evidence as finite-size avalanche dynamics on a real economic network.
The remainder of the paper is organised as follows. Section (ref) positions the contribution within the literatures on self-organised criticality, econophysics, and economic networks. Section (ref) develops the Hall-Sandpile model. Section (ref) describes the real-network computational design. Section (ref) reports results across propagation normalisation, exposure concentration, baseline regimes, phase diagrams and tail behaviour. Section (ref) discusses implications and limitations. Section (ref) concludes.
This section positions the Hall-Sandpile model within four related but distinct strands of work: self-organised criticality (SOC) in physics and its generalisations; econophysics and the statistical physics of financial systems; network-origin and production-network theories of aggregate fluctuations; and curvature-based “Sandpile Economics” approaches to structural fragility. We close the section by stating what is genuinely new in our contribution relative to each of these strands.
Sandpile dynamics and SOC were introduced by bak1987self, bak1988self as a generic mechanism by which slowly driven dissipative systems generate scale-free avalanche distributions without parameter tuning. The original Bak--Tang--Wiesenfeld model and its many variants exhibit power-law statistics for avalanche size, duration and area, and have become a paradigmatic illustration of how heavy-tailed event distributions emerge in spatially extended systems. bak1996nature provides the textbook synthesis; turcotte1999self reviews applications across geophysics; and vespignani1998how offer a mean-field analytical framework that clarifies the relationship between dissipation, branching ratios and the onset of criticality.
Two features of the canonical SOC literature matter for our purposes. First, SOC is typically studied on regular lattices or on synthetic networks (Erd\H{o}s--R\'enyi, scale-free, modular), in which the substrate is an artefact of the modelling exercise. Our paper instead takes the substrate as observed: the WIOD input--output network is not an idealised topology, it is a description of the actual country--sector linkages of the world economy as recorded in timmer2015illustrated. Second, canonical SOC delivers tail exponents in the range $\alpha\in[1,2]$, often close to $\alpha=3/2$ in the Bak--Tang--Wiesenfeld model. We will see in Section (ref) that our estimated exponents ($\alpha\approx 5.9$--$6.2$) are markedly steeper than this canonical range, which is one of the reasons we refrain from claiming SOC for the global economy. Branching-process formulations of SOC lippiello2008dynamical provide a useful benchmark for finite-size diagnostics.
A parallel tradition uses statistical-physics methods to analyse financial and economic systems. mantegna2000introduction and bouchaud2003theory provide the textbook treatments of econophysics; schweitzer2009economic survey the network-physics view of economic systems; and sornette2002predictability discusses predictability of catastrophic events from material rupture to financial crashes. In banking and finance, may2008complex and haldane2011systemic draw analogies between ecological and banking ecosystems, while caldarelli2013reconstructing address the empirical problem of reconstructing credit networks from partial information. Network metrics of systemic risk such as DebtRank battiston2012debtrank measure how much distress propagates from a node to the rest of the network. Cascade dynamics on networks have been studied as binary-threshold contagion in watts2002simple, as load-redistribution attacks in motter2002cascade, and as interdependent network failures in buldyrev2010catastrophic. glasserman2016contagion provide a rigorous review of contagion in financial networks.
The Hall-Sandpile mechanism is related to but distinct from these approaches. Cascade-based attacks motter2002cascade,buldyrev2010catastrophic focus on topological amplification through load redistribution after a targeted node removal. Threshold-contagion models watts2002simple,gai2010contagion,elliott2014financial focus on linear propagation of binary distress along edges. DebtRank battiston2012debtrank is a static metric of vulnerability, not a dynamical model of activation. Our model adds two features: (i) an external field $B_t$ that loads stress simultaneously on many exposed nodes rather than starting from a single targeted node; and (ii) a multiplicative stress-conversion term (ref) that interacts with redundancy and capacity rather than a linear pass-through.
A third strand of literature, mostly within macroeconomics, has developed network-based theories of aggregate fluctuations. long1983real provided an early model in which sectoral shocks aggregate to business-cycle fluctuations through real linkages. gabaix2011granular showed how the granularity of firm-size distributions can give rise to non-trivial aggregate volatility, linking firm-level idiosyncratic shocks to business-cycle properties. acemoglu2012network formalised how the input--output structure of the economy shapes the propagation of microeconomic shocks into aggregate fluctuations. carvalho2019production provide a comprehensive primer on production networks, while baqaee2019networks,baqaee2019macroeconomic extend the analysis beyond Hulten's theorem to non-linear and disaggregated environments. On the empirical side, barrot2016input document how input specificity shapes the propagation of idiosyncratic shocks, and carvalho2021supply use the 2011 Great East Japan earthquake to identify supply-chain disruption effects on downstream firms.
This literature has clarified that the architecture of input--output linkages is consequential for aggregate dynamics. It is, however, predominantly longitudinal: a shock enters a node and propagates along weighted directed edges. The Hall-Sandpile model contributes a transversal mechanism: an external field $B_t$ loads stress on flow-intensive, low-redundancy nodes through (ref), and the resulting transversal stress lowers the effective activation threshold via Proposition (ref). Aggregate fluctuations in our framework therefore reflect not only the topological transmission of sectoral shocks but also the field-mediated activation of structural fragility.
vallarino2024sandpile introduces a curvature-based “Sandpile Economics” programme, in which production networks are characterised by Forman--Ricci or Ollivier--Ricci curvature concentrations that act as geometric markers of structural vulnerability. In this view, a network is fragile when its negative-curvature subgraphs are dense and its effective branching number is large; cascades emerge as power-law distributions on such substrates. The present paper shares the sandpile language and the network-substrate philosophy of vallarino2024sandpile but moves the analytical focus from geometric fragility to external loading: curvature describes the shape of the substrate, while the Hall-like mechanism describes the loading on it. The two perspectives are complementary. A high-curvature, low-redundancy node is structurally exposed; a high-$B_t$ field activates that exposure; and the avalanche size $S_t$ in (ref) reflects the joint action of structure and field.
Relative to the four strands above, the contribution of the present paper is fourfold. First, we formalise a Hall-like multiplicative stress-conversion mechanism (ref) that distinguishes external-field intensity from node-level exposure, redundancy and capacity, and we embed it into a sandpile threshold dynamics. Second, we calibrate the model on the observed WIOD production network rather than on a synthetic substrate, and we discipline the calibration with three alternative propagation operators to rule out mechanical near-criticality. Third, we map the regime structure of the model across a two-dimensional grid in $(\bar{B},\sigma_D)$ and document a positively sloped phase frontier between absorption and avalanche, consistent with Proposition (ref). Fourth, we report tail diagnostics that show regime-dependent thickening but not universal power-law behaviour, and we explicitly discuss how this finding fits within the broader debate on power laws in empirical data stumpf2012critical.
The Hall-Sandpile model has four building blocks: (i) a weighted directed network defined by a propagation operator on a real input--output topology; (ii) a Hall-like multiplicative stress-conversion function that maps an external field and node-level exposure into transversal stress; (iii) a discrete-time stress accumulation and toppling rule of sandpile type; and (iv) a finite-size avalanche statistic that aggregates toppling events at each period. The first two blocks are non-standard relative to the existing literature on contagion in production networks; the last two follow the canonical Bak--Tang--Wiesenfeld dynamics with explicit dissipation. This subsection summarises how the four blocks fit together; the rest of the section makes them precise.
The network block grounds the model in observed data. Where most SOC-style economic models use synthetic substrates, the present framework requires a concrete propagation operator derived from WIOD. We construct three such operators to discipline the analysis. The Hall-like block adds a multiplicative coupling between an external field $B_t$ and the inverse of node-level redundancy and capacity, producing transversal stress $H_{i,t}$ that depends on which nodes are exposed, not only on the magnitude of the field. The sandpile block converts accumulated stress into avalanches through a threshold rule. The avalanche statistic $S_t$ counts toppling nodes and provides the observable used in all phase-diagram and tail analyses below. The interplay between these four blocks is what makes the model tractable yet non-trivial: each block is simple in isolation, but their composition produces a regime structure that neither network propagation alone nor sandpile dynamics alone would generate.
Let $G_t = (V, E_t, A_t)$ denote a weighted directed network in year $t$, with node set $V$ of country--sector production units and edge weights $A_t = [a_{ij,t}]$ derived from the WIOD intermediate transaction matrix $Z_t = [z_{ij,t}]$, where $z_{ij,t}$ is the intermediate flow from node $i$ to node $j$. We construct three alternative propagation operators.
The row-share matrix is
which captures the distribution of outgoing intermediate flows. By construction, its rows sum to one (whenever the row total is positive) and its spectral radius is close to one. This near-unit radius is a property of the normalisation, not of the underlying economy.
The leakage-adjusted dissipative matrix is
where $\ell_{i,t}$ is the share of intermediate outflows in a gross row-use proxy for node $i$. This allows row sums to be strictly below one and embeds a controlled amount of dissipation at each step. We use $A^{\mathrm{leak}}_t$ as the main propagation operator in the simulation.
The max-row-normalised matrix is
which avoids imposing a stochastic transition structure altogether and preserves cross-node heterogeneity in row totals. We report it as a robustness operator.
Let $\rho(\cdot)$ denote the spectral radius. By construction $\rho(A^{\mathrm{share}}_t) \approx 1$, while $\rho(A^{\mathrm{leak}}_t)$ and $\rho(A^{\mathrm{max}}_t)$ are bounded away from one for our calibration. Reporting all three is essential to ensure that the avalanche dynamics analysed below is not a mechanical consequence of operating on a unit-radius operator.
Let $B_t \geq 0$ denote the intensity of an external field acting on the network at time $t$, and define for each node $i$: $I_{i,t}$ a measure of relative flow intensity (its share of total flows in the network), $D_{i,t}$ a measure of redundancy or output substitutability, and $C_{i,t}$ a measure of absorptive capacity. The Hall-like transversal stress on node $i$ is
where $\varepsilon > 0$ is a small regularisation constant. Equation (ref) has the structure of a stress-conversion function: an external field $B_t$ becomes transversal stress when it acts on high-flow, low-redundancy, low-capacity nodes. Defining the structural resistance
we can write $H_{i,t} = B_t\,I_{i,t}\,R_{i,t}$, i.e. a multiplicative combination of loading, exposure and structural resistance. The relative Hall-like exposure is $H^{\mathrm{rel}}_{i,t} = H_{i,t} / \sum_{j} H_{j,t}$.
The mathematical structure of (ref) mirrors that of the classical Hall coefficient in solid-state physics: the magnitude of the transverse response is proportional to the product of the field and the carrier flow, and inversely proportional to the carrier density. In our economic re-interpretation, the carrier density is replaced by the product $D_{i,t}\,C_{i,t}$ of redundancy and absorptive capacity, which plays the role of a structural buffer analogous to charge density in the physical case. This is a structural analogy, not a physical identity: there is no claim that economic flows obey Lorentz-force kinematics, nor that $H_{i,t}$ has a dimensional interpretation as a voltage. The analogy is useful only insofar as it organises the multiplicative interaction between external loading, node exposure and structural buffering in a single expression, and only insofar as it makes the comparative statics (ref) transparent: $\partial H/\partial B$ is the field-elasticity, $\partial H/\partial I$ is the exposure-elasticity, and $\partial H/\partial(D C)$ is the buffer-elasticity. All three are well defined and signed, and they reappear in Propositions (ref)--(ref) below.
Let $s_{i,t}$ denote the cumulative stress at node $i$ at time $t$ and $x_{i,t}$ a node-specific idiosyncratic shock. Stress evolves as
where $\delta\in(0,1)$ is the dissipation rate, $\alpha\geq0$ scales idiosyncratic shocks, $\beta\geq0$ scales network propagation through the leakage-adjusted operator, and $\gamma\geq0$ scales Hall-like loading. Node $i$ topples when
where $\theta_i$ is a node-specific threshold. The avalanche size at time $t$ is the number of toppling nodes,
When a node topples, a fraction of its excess stress is redistributed to its neighbours through $A^{\mathrm{leak}}_t$ and the remainder is dissipated. This dissipative redistribution prevents explosive trajectories and makes the dynamics finite-size and computationally well behaved.
We state three substantive hypotheses about the behaviour of the Hall-Sandpile model and three structural propositions implied by (ref) and (ref).
We complement these propositions with a lemma on the boundedness of the stress process, which guarantees that the simulated dynamics is well posed and finite-size in the sense used throughout the paper.
The structural propositions and the lemma jointly motivate the phase-diagram exercise reported in Section (ref): regime transitions are well posed, finite-size, and driven by the joint movement of $\bar{B}$ and $\sigma_D$.
We use the World Input--Output Database timmer2015illustrated, timmer2016anatomy, restricted to the 2000--2014 release. WIOD is a publicly available multi-region input--output dataset that records intermediate transactions between country--sector production units. In our processed sample, each annual network has $n = 2{,}283$ country--sector nodes after harmonisation. The yearly panel is used to construct propagation operators and Hall-like exposure measures over time; the 2014 network is used as the substrate for the main avalanche simulation. No proprietary data are used.
For each year $t$ we compute the three propagation operators in (ref)--(ref), their spectral radii $\rho^{\mathrm{share}}_t$, $\rho^{\mathrm{leak}}_t$ and $\rho^{\mathrm{max}}_t$, and the leakage profile $\{\ell_{i,t}\}$. For each node $i$ we compute relative flow intensity $I_{i,t}$, redundancy $D_{i,t}$ from out-degree concentration, and absorptive capacity $C_{i,t}$ from a normalised redundancy proxy. We fix $\varepsilon$ at a small positive constant and compute $H_{i,t}$ and $H^{\mathrm{rel}}_{i,t}$ following Section (ref). Table (ref) summarises the resulting yearly indicators.
The Hall-Sandpile simulation runs on the 2014 WIOD network. For each configuration of external field intensity $B_t$ and redundancy stress $\sigma_D$, we draw idiosyncratic shocks $x_{i,t}$ from a node-specific distribution, apply Hall-like loading according to (ref), update stress via (ref) and determine avalanche size from (ref). Field intensity follows $B_t = \bar{B} + \xi_t$ with $\xi_t$ a mean-zero perturbation; redundancy stress is implemented as a controlled multiplicative reduction of effective redundancy, $D_{i,t} \mapsto D_{i,t}/\sigma_D$, which scales the contribution of $H_{i,t}$ in (ref).
We run two complementary experiments. Baseline scenarios fix four representative pairs $(\bar{B},\sigma_D)$ corresponding to stable absorption, latent fragility, critical transition and avalanche regime, with $15{,}000$ Monte Carlo periods each. The phase diagram experiment sweeps $\bar{B}$ over a regular grid in $[0.25,2.0]$ and $\sigma_D$ over $[0.5,2.5]$ and computes mean avalanche size, the share of non-zero events, the percentile statistics of $S_t$ and the threshold-event probabilities $\mathbb{P}(S\geq 5)$, $\mathbb{P}(S\geq 10)$ and $\mathbb{P}(S\geq 20)$. Throughout, $A^{\mathrm{leak}}$ is the main propagation operator and $A^{\mathrm{share}}$, $A^{\mathrm{max}}$ are reported only as spectral-radius robustness checks.
The Hall-Sandpile dynamics in (ref) depends on a small set of parameters: the dissipation rate $\delta$, the idiosyncratic shock scale $\alpha$, the propagation coefficient $\beta$, the Hall loading coefficient $\gamma$, the toppling thresholds $\{\theta_i\}_{i\in V}$, and the regularisation $\varepsilon$ in (ref). Table (ref) summarises the calibrated values used throughout the paper. The choices are guided by three principles: (i) the contraction condition of Lemma (ref), $\beta < \delta/\rho^{\mathrm{leak}}$, must hold to ensure boundedness of the stress process; (ii) the time scale of mean reversion, controlled by $\delta$, should be of the order of one to a few simulation periods, following standard practice in dissipative sandpile models vespignani1998how; and (iii) the magnitudes of $\alpha$, $\beta$, $\gamma$ should be of comparable order so that no single forcing term mechanically dominates (ref).
For our calibration, $\rho^{\mathrm{leak}} = 0.334$ in 2014 (Table (ref)); the contraction condition becomes $\beta < 0.20/0.334 \approx 0.599$, and our value $\beta = 0.40$ lies comfortably below this threshold. Sensitivity analyses with $\delta\in\{0.10, 0.15, 0.25\}$, $\beta\in\{0.30, 0.50\}$ and $\gamma\in\{0.30, 0.70\}$ leave the qualitative ordering of the four regimes unchanged; we report a subset of these robustness experiments in Appendix (ref).
The toppling thresholds are set uniformly at $\theta_i = 1$ across all nodes. Heterogeneous thresholds (e.g.\ scaled by node-level capacity) shift the location of the phase frontier in $(\bar{B},\sigma_D)$ space but preserve its shape; this aligns with Proposition (ref), which predicts a positively sloped frontier irrespective of the threshold profile. The regularisation $\varepsilon=10^{-6}$ prevents division by zero in (ref) for nodes with $D_{i,t}C_{i,t}=0$ (a small number of WIOD service-sector nodes have zero recorded outgoing intermediate flows and therefore zero raw redundancy in our construction); the resulting numerical stability does not affect the qualitative results.
Each cell of the phase diagram involves a Monte Carlo experiment with the following protocol. Burn-in. We discard the first $T_{\mathrm{burn}}=50$ simulation periods to remove the transient from the initial condition $s_{i,0}=0$. Figure (ref) shows that the burn-in is conservative: stationary levels of mean avalanche size are essentially achieved within the first ten periods. Stationary phase. We retain the next $T_{\mathrm{stat}}=150$ periods, during which avalanche sizes $S_t$ are recorded and percentile statistics computed. Replications. For the baseline scenarios we run $R = 100$ independent replications per scenario, yielding $T_{\mathrm{stat}} \times R = 15{,}000$ stationary observations of $S_t$. For the phase-diagram cells, we run $R = 50$ replications, yielding $7{,}500$ stationary observations per cell. Random seeds. Each replication uses an independent random seed drawn from a deterministic master seed, ensuring reproducibility. Convergence diagnostics. For each cell of the phase diagram we compute the standard error of the mean avalanche size across replications; standard errors of $\mathbb{E}[S_t]$ are below $5\%$ of the point estimate in all active cells of the grid, and below $0.02$ in all absorbing cells. The threshold-event probabilities $\mathbb{P}(S\geq k)$ are estimated as sample frequencies; for $k=5,10,20$, the binomial standard error is bounded by $\sqrt{p(1-p)/N_{\mathrm{eff}}}$ with $N_{\mathrm{eff}} \geq 7{,}500$.
The total number of simulated periods across the entire phase grid ($90$ cells, $50$ replications, $200$ periods including burn-in) is $9 \times 10^{5}$, which we found computationally tractable on a single workstation using a vectorised Python implementation.
Table (ref) summarises the correspondence between modelling objects and economic interpretations. The substrate is the observed WIOD network; the loading is a controlled simulation. We do not attempt to identify any historical shock.
Figure (ref) reports the three spectral radii $\rho^{\mathrm{share}}_t$, $\rho^{\mathrm{leak}}_t$ and $\rho^{\mathrm{max}}_t$ for the WIOD network over 2000--2014, and Table (ref) summarises the corresponding yearly indicators. The row-share spectral radius lies in the interval $[0.973,\,0.984]$ with little year-to-year variation: by the construction in (ref), $\rho^{\mathrm{share}}_t$ approximates a row-stochastic transition operator and is therefore close to one mechanically. The leakage-adjusted spectral radius is sharply lower, $\rho^{\mathrm{leak}}_t \in [0.334,\,0.344]$ across the entire panel, and the max-row-normalised spectral radius rises from $0.219$ in 2000 to $0.317$ in 2014. The leakage profile is stable around $\bar{\ell}_t \approx 0.373$.
The economic content of this exercise is methodological. The near-unit value of $\rho^{\mathrm{share}}_t$ is a property of the row-stochastic normalisation: it does not indicate that the world economy operates close to spectral criticality. Reporting $\rho^{\mathrm{leak}}_t$ and $\rho^{\mathrm{max}}_t$ alongside $\rho^{\mathrm{share}}_t$ documents that the dissipative operator we use in (ref) has spectral radius safely below unity throughout the sample. The avalanche dynamics analysed below therefore cannot be explained by a unit-radius operator alone; they arise from the interaction of dissipative propagation, idiosyncratic shocks and Hall-like loading.
Figure (ref) shows the yearly mean, median and 95th percentile of relative Hall-like exposure $H^{\mathrm{rel}}_{i,t}$ on a logarithmic scale. The mean stays close to $10^{-3}$ throughout the panel and the median is approximately one order of magnitude below the mean. The 95th percentile lies between $3.3\times10^{-3}$ and $4.2\times10^{-3}$, indicating that a small subset of nodes concentrates a disproportionate share of the loading potential. The maximum value of $H^{\mathrm{rel}}_{i,t}$ in 2014 reaches $0.317$, i.e. nearly one third of the entire network's Hall-like exposure is concentrated on a single node.
Table (ref) reports the top 15 nodes by $H^{\mathrm{rel}}$ in 2014 together with flow share and structural resistance. The top-ranked node is CHN_E36 (Chinese water collection and supply), with $H^{\mathrm{rel}} = 0.317$, followed by CHN_A02 (forestry and logging, $0.087$) and CHN_A03 (fishing and aquaculture, $0.054$). Manufacturing nodes such as CHN_C31_C32 (furniture and other manufacturing, $0.046$), CHN_C25 (fabricated metals, $0.044$), CHN_C17 (paper products, $0.043$) and \texttt{CHN_C26} (computer and electronic products, $0.037$) follow. Two patterns are worth noting. First, several of the top nodes are utility/service sectors (water collection, administrative services, telecommunications) that combine very high inflows with measured zero outgoing redundancy and a structural-resistance value of approximately $14.74$ in our calibration; this combination produces high $H^{\mathrm{rel}}$ even though their downstream propagation potential is limited. Second, the dominance of Chinese country--sector codes in the top ranks reflects the size and density of those sectors in the WIOD 2014 network and is a feature of the network calibration, not a sectoral diagnosis of national fragility. The interpretation is that, conditional on the external field intensity considered in the simulation, the loading of transversal stress is concentrated on a small set of country--sector nodes; whether and how that loading translates into realised cascades depends on the regime explored in Sections (ref)--(ref).
Figure (ref) provides a visualisation of the upper-tail concentration. The drop from CHN_E36 to the second-ranked node is sharp (from $0.317$ to $0.087$), and the long descending tail beyond the top dozen confirms that, in the 2014 calibration, the Hall-like loading potential is heavy-tailed. Figure (ref) plots structural resistance against $H^{\mathrm{rel}}$ in log scale: most nodes lie in the lower-left region with low structural resistance and low exposure, while a small number of high-resistance nodes (the utility/service sectors above) sit in the upper-right corner. The relationship is not monotonic because flow share also matters in (ref): a structurally resistant node with no inflows has zero exposure.
Table (ref) reports the four baseline scenarios on the 2014 WIOD network. Mean avalanche size grows monotonically across regimes, from $0.084$ in stable absorption ($\bar{B}=0.45, \sigma_D=0.7$) to $5.81$ in the avalanche regime ($\bar{B}=1.35, \sigma_D=2.3$). The share of non-zero events rises from $8.4\%$ in stable absorption to $95.3\%$ in the avalanche regime. The probability $\mathbb{P}(S \geq 5)$ is essentially zero in the first two regimes, equal to $9.9\%$ in critical transition and $58.5\%$ in the avalanche regime; $\mathbb{P}(S \geq 10)$ is non-trivial only in the last two regimes, while $\mathbb{P}(S\geq 20)$ remains below $0.1\%$ even in the avalanche regime, indicating that very large events are rare even under aggressive joint loading.
Figure (ref) shows the average avalanche size as a function of simulation time in each baseline regime. After a short transient, every regime converges to a distinct stationary level of avalanche activity: the stable-absorption path fluctuates near zero, latent fragility settles around $\bar{S}\approx 0.5$, the critical transition stabilises near $\bar{S}\approx 2$, and the avalanche regime fluctuates around $\bar{S}\approx 5.5$--$6.5$. The dynamics is not explosive: the dissipative redistribution prevents trajectories from running away to network-wide events. The four regimes are controlled stationary states of the same model under different loading intensities, not points on a divergent path.
Figures (ref)--(ref) present the phase diagrams over the $(\bar{B},\sigma_D)$ grid for the 2014 substrate. Mean avalanche size (Figure (ref)) increases monotonically along both axes, reaching values near $13$ in the upper-right corner (high field, high redundancy stress). The probability $\mathbb{P}(S\geq 5)$ (Figure (ref)) sets in earliest, with non-trivial values from $\bar{B}\!\gtrsim\!1.0$ and $\sigma_D\!\gtrsim\!1.5$, and approaches one in the upper-right corner. The probability $\mathbb{P}(S\geq 10)$ (Figure (ref)) requires stronger joint loading and is the most informative threshold for the onset of systemic activity. The probability $\mathbb{P}(S\geq 20)$ (Figure (ref)) is small everywhere except in the upper-right corner, never exceeding $\sim\!7\%$. The model therefore does not produce a permanent supercritical state under reasonable loading.
A compact summary of the four phase regions is reported in Table (ref). The progression from absorption to avalanche is smooth and ordered, consistent with Proposition (ref): the avalanche regime emerges only when both $\bar{B}$ and $\sigma_D$ are large; high values of either variable in isolation are insufficient.
Figure (ref) shows the complementary cumulative distribution function (CCDF) of avalanche size by regime in log--log scale. The distribution shifts to the right and develops a thicker tail as one moves from stable absorption (essentially no events) through latent fragility, critical transition and avalanche regime. The avalanche regime CCDF reaches values near $S=60$ at probabilities of order $10^{-4}$, while the critical transition regime extends to $S\approx20$ at similar probabilities. Figure (ref) shows the corresponding histograms of avalanche size: stable absorption and latent fragility concentrate mass near zero or one, while critical transition and avalanche shift mass to larger sizes.
We assess tail behaviour by maximum-likelihood estimation of a power-law exponent $\alpha$ above an estimated lower cut-off $x_{\min}$ clauset2009power. Table (ref) reports the results. For stable absorption the distribution is essentially degenerate at zero and the estimated tail is uninformative (no statistical power for power-law inference). For latent fragility the estimated $\alpha$ is approximately $29$, corresponding to a thin tail. The two active regimes critical transition and avalanche regime return $\alpha\approx 6.20$ and $\alpha\approx 5.94$, respectively, with $x_{\min}\in\{4,9\}$. These exponents are well above the values typically reported in the SOC literature bak1987self,turcotte1999self, where canonical sandpile models deliver exponents in the range $1$--$2$. The tails we observe are thicker than in the inactive regimes but still much thinner than would be expected under genuine SOC.
We therefore describe the finding as regime-dependent tail thickening in a finite real network. The sandpile-type dynamics on a real, dissipative production-network substrate amplify avalanche-size heterogeneity in the active regimes, but do not produce evidence of a universal power-law exponent. This caveat aligns with the broader critique that empirical heavy-tailed data should not be too quickly labelled as power laws stumpf2012critical.
The estimated tail exponents $\alpha\approx 5.94$--$6.20$ in the critical-transition and avalanche regimes are conspicuously above the canonical SOC range. Three observations help to interpret this gap.
First, $|V| = 2{,}283$ is large by the standards of empirical production networks but small by the standards of statistical-physics SOC simulations, where lattice sizes routinely exceed $10^{4}$--$10^{6}$ vespignani1998how. Finite-size cutoffs in SOC scale typically appear around $S^* \propto |V|^{1/D}$ for some network dimension $D$, and any power-law signal is mechanically truncated above this cutoff. In our active regimes, the largest recorded avalanches are $S\approx 30$, which is two orders of magnitude below $|V|$ but still close enough to the system size that finite-size corrections can plausibly steepen the apparent slope of the CCDF.
Second, the dissipative operator $A^{\mathrm{leak}}$ has spectral radius $\rho^{\mathrm{leak}}\approx 0.334$, which is well below one. SOC arguments typically require operators close to the percolation or epidemic threshold boguna2003absence; with $\rho^{\mathrm{leak}}$ comfortably subcritical, the system is intrinsically dissipative and cannot sustain the long-range spatial correlations needed for a true critical point. Reporting $\rho^{\mathrm{share}}\approx 0.97$ alongside is a robustness check, not a claim that the economy operates near percolation.
Third, the Bak--Tang--Wiesenfeld model attains its $\alpha\approx 3/2$ exponent in the limit of infinite separation of time scales between slow driving and fast relaxation bak1987self, turcotte1999self. In our calibration, the field $B_t$ is applied at the same time scale as the relaxation steps; the system is closer to a noisy driven dissipative system than to a strictly slow-driven one. This too tends to produce steeper apparent tails.
Taken together, the estimated $\alpha\approx 6$ should be interpreted as a finite-size, finite-driving-rate diagnostic on a real network, not as a falsification of SOC. We do not claim the underlying dynamics is critical; we claim that loading-induced regimes thicken the tail relative to the absorbing baseline, and that even in the most active regime the system remains subcritical and finite-size-bounded. A formal investigation of how $\alpha$ scales with $|V|$ and with the driving-rate ratio is left for future work.
We re-ran the entire simulation suite under the row-share operator $A^{\mathrm{share}}$ and the max-row-normalised operator $A^{\mathrm{max}}$. Three findings emerge.
First, the four-regime ordering is preserved across all three operators. Mean avalanche size is monotone in $(\bar{B},\sigma_D)$ under each normalisation, and the threshold-event probabilities $\mathbb{P}(S\geq k)$ retain their qualitative shape on the phase grid.
Second, the location of the phase frontier shifts. Under $A^{\mathrm{share}}$, the near-unit spectral radius pushes the dynamics towards more frequent small events even in low-loading regimes; the absorption zone is correspondingly smaller. Under $A^{\mathrm{max}}$, which preserves cross-node heterogeneity, the regime ordering is again preserved with slightly stronger amplification near the upper-right corner of the phase diagram, reflecting the differential weight given to flow-rich nodes.
Third, the tail exponents in the active regimes shift only modestly: $\alpha$ moves between $5.4$ and $6.4$ across the three operators in the avalanche regime. This robustness reinforces our cautious reading of the tail behaviour: the regime-dependent thickening of the avalanche distribution is a feature of the Hall-Sandpile dynamics, not a feature of the row-stochastic normalisation. The substantive findings of Sections (ref)--(ref) therefore do not depend on the choice of normalisation. We report the leakage-adjusted operator in the main text because it embeds dissipation in a way that is economically more defensible than imposing a stochastic transition structure.
The five preceding subsections deliver an internally consistent picture. The propagation operators of Section (ref) establish that the leakage-adjusted dissipative substrate is subcritical throughout the WIOD panel, with $\rho^{\mathrm{leak}}$ stable around $0.33$--$0.34$. The exposure analysis of Section (ref) shows that, on this subcritical substrate, the relative Hall-like exposure is nonetheless heavily concentrated: the upper tail of $H^{\mathrm{rel}}$ is dominated by a small set of country--sector nodes, with the top-ranked node accounting for nearly one third of total relative exposure in 2014. The baseline scenarios of Section (ref) show that the same substrate, under controlled $(\bar{B},\sigma_D)$ pairs, produces four qualitatively distinct stationary regimes whose mean avalanche size spans nearly two orders of magnitude. The phase diagrams of Section (ref) document a positively sloped frontier between absorption and avalanche regimes, consistent with Proposition (ref); the threshold-event probabilities $\mathbb{P}(S\geq k)$ activate sequentially as loading increases, with $\mathbb{P}(S\geq 5)$ entering first, $\mathbb{P}(S\geq 10)$ requiring stronger joint loading and $\mathbb{P}(S\geq 20)$ confined to the upper-right corner. The tail diagnostics and finite-size discussion of Sections (ref)--(ref) caution that, while loading produces regime-dependent thickening of the avalanche distribution, the steepness of the estimated $\alpha$ leaves no room for a claim of universal SOC at this network size. Across all five findings, the dynamics is well posed, finite-size and bounded, and the choice of propagation operator does not alter the substantive picture.
The Hall-Sandpile model formalises a simple but consequential idea: economic systems do not only transmit shocks along edges of a network; they also convert shocks across dimensions. The Hall-like stress term (ref) captures this conversion as a multiplicative interaction between external field intensity $B_t$, relative flow exposure $I_{i,t}$, and the inverse of redundancy and absorptive capacity. By embedding this loading into a sandpile threshold dynamics, we obtain a framework in which the same external shock can remain absorbed in a high-redundancy network and trigger an avalanche regime in a low-redundancy one. Hypothesis (ref) and Proposition (ref) are concrete statements of this contingency.
The conversion logic helps to organise heterogeneous crisis episodes under a common formal lens. An energy shock raises $B_t$ for nodes heavily exposed to fuel inputs; the same field generates inflation in food-and-beverage sectors, fiscal strain in energy-subsidising governments, and industrial contraction in energy-intensive manufacturing, all because the underlying $H_{i,t}$ acts on different flow profiles. A monetary shock raises $B_t$ for nodes exposed to short-term funding, with the same field producing liquidity stress in banking, sovereign-risk repricing in fiscal authorities and default clustering in leveraged corporates. A geopolitical shock raises $B_t$ for nodes exposed to specific cross-border flows, generating supply-chain disruption, commodity-price pressure and topological reconfiguration of trade. In each case, the field-mediated activation is heterogeneous across nodes, and the systemic outcome depends on the upper tail of $H^{\mathrm{rel}}_{i,t}$ rather than the mean.
Three diagnostic implications follow. First, systemic risk is not a property of the average node. The mean of $H^{\mathrm{rel}}_{i,t}$ is small and stable across the WIOD panel, while its upper tail concentrates large loading potential on a small subset of country--sector nodes. Average connectivity statistics therefore systematically understate exposure to transversal stress; the diagnostically informative quantity is the upper tail of the loading distribution.
Second, redundancy and absorptive capacity matter not because they eliminate shocks, but because they reduce the conversion of shocks into systemic stress (Hypothesis (ref) and Proposition (ref)). A node with abundant input substitutes and slack capacity translates a given external field into a smaller increment of internal stress and is correspondingly less likely to topple.
Third, the joint nature of the regime transition (Proposition (ref)) cautions against single-axis interpretations of fragility. Neither field intensity alone nor redundancy stress alone is sufficient to push the system into an avalanche regime; both must rise jointly. This has practical implications for early-warning systems based on a single indicator, which can fail to detect the regime transition simply because the orthogonal axis has not moved.
A useful corollary of Proposition (ref) is that the leverage available to a policy designer operates on the multiplicative structure of (ref) rather than on the additive shock $x_{i,t}$. Policies that aim to prevent shocks ($B_t \to 0$) are typically beyond domestic control, since external fields originate elsewhere in the global system. Policies that target stress-conversion coefficients, by contrast, act on the denominator of (ref) and are within the policy space of national or sectoral authorities. Concretely, an increase in $D_{i,t}$ at a vulnerable node, achieved through alternative-supplier registries, modular standards or inventory pre-positioning, lowers $H_{i,t}$ for any given $B_t$. An increase in $C_{i,t}$, achieved through liquidity buffers, working capital pre-funding or absorptive insurance, has the same effect. Because $D_{i,t}$ and $C_{i,t}$ enter (ref) multiplicatively, small simultaneous improvements on both can yield disproportionately large reductions in transversal stress.
A second policy implication concerns the targeting of interventions. The upper tail of $H^{\mathrm{rel}}_{i,t}$ in WIOD 2014 is highly concentrated: the top fifteen nodes account for nearly $90\%$ of relative Hall-like exposure under our calibration. A policy that improves redundancy or capacity at the very top of the exposure distribution therefore produces a disproportionate reduction in expected avalanche size, relative to a policy that operates on the median node. This is the network-amplification analogue of granular interventions in gabaix2011granular.
Finally, the phase-frontier geometry of Proposition (ref) suggests that resilience policies are most effective when they are anticipatory, i.e.\ when they act on $D_{i,t}$ and $C_{i,t}$ before $\bar{B}$ rises. A network designed for the absorption regime under benign $\bar{B}$ may sit close to the frontier under elevated $\bar{B}$, and policy lags can become consequential precisely along the diagonal in $(\bar{B},\sigma_D)$ space along which the avalanche regime opens.
The framework does not identify any historical episode but offers a language in which several documented features of recent crises can be re-described. The 2008--2009 financial crisis combined elevated $B_t$ in financial nodes with reduced absorptive capacity and the removal of input substitutability in short-term funding; the 2020 pandemic combined a labour-supply shock with cross-border supply-chain disruptions, raising $H^{\mathrm{rel}}$ on nodes with concentrated upstream suppliers; the 2022 energy shock raised $B_t$ for nodes with high energy-intensity exposure under simultaneously low spare capacity in energy markets. In each case, the joint activation of field intensity and redundancy stress is consistent with the avalanche regime in our phase diagram. We emphasise that this is a re-description, not an identification: causal claims would require linking measured shock proxies to specific cells of the phase grid, which is beyond the scope of the present paper.
We do not interpret our results as evidence of self-organised criticality in the global economy. Three reasons argue for caution. The estimated tail exponents in the active regimes ($\alpha \approx 5.9$--$6.2$) are well above the canonical SOC range. The avalanche dynamics is finite-size and dissipative by construction, not driven to a true critical point through endogenous self-tuning. And the fields used in the simulation are externally controlled, not estimated from a historical episode. The contribution is therefore better described as a controlled real-network experiment: we study how a real production-network substrate responds to a class of nonlinear loadings, and we document the resulting regime structure. This is complementary to, but distinct from, the empirical identification of historical contagion episodes barrot2016input,carvalho2021supply.
\paragraph{Limitations} The results are conditional on the WIOD coverage up to 2014 and on the harmonisation choices used to construct the country--sector node set; expanding to more recent vintages and finer sectoral disaggregation is a natural extension. The Hall-like exposure $H^{\mathrm{rel}}_{i,t}$ is an index, not a directly observed physical variable; its concentration on certain utility/service sectors with measured zero outgoing redundancy reflects how those sectors enter the WIOD accounting structure (large inflows, no recorded intermediate outflows) and should be read as a calibration indicator, not a sectoral vulnerability diagnosis. Tail-exponent estimates are finite-size diagnostics: with a panel of $n=2{,}283$ nodes and $15{,}000$ Monte Carlo periods per regime, very large events are rare and their statistics are correspondingly noisy. The simulated fields are controlled experiments rather than measured shocks, and the results say nothing about the probability with which any specific external field will materialise in practice. The redundancy and capacity proxies of Appendix (ref) use only network-level information and would benefit from refinement using inventory, liquidity and balance-sheet data; we leave such refinements for future work.
\paragraph{Avenues for future research} Three extensions stand out. First, embedding measured shock proxies (energy-price changes, monetary-policy surprises, supply-chain disruption indices) in the external field $B_t$ would link the controlled-loading exercises performed here to historical episodes. Second, integrating the Hall-Sandpile mechanism with curvature-based geometric fragility approaches vallarino2024sandpile would produce a unified framework in which structural vulnerability and external loading interact. Third, scaling the framework to firm-level production-network data, where $|V|$ can be of order $10^5$--$10^6$, would allow a sharper investigation of finite-size scaling and a direct test of whether the apparent steepening of $\alpha$ in our calibration is genuinely a finite-size artefact.
We have introduced a Hall-Sandpile model that combines a Hall-like transversal stress mechanism with sandpile threshold dynamics on a real production-network substrate. The model formalises how external fields convert exposure into transversal stress and lower effective activation thresholds, and it uses the WIOD 2000--2014 input--output network as the empirical topology. Three propagation operators guarantee that nonlinear avalanche behaviour does not arise from a unit-radius row-stochastic operator. Controlled Monte Carlo experiments produce four ordered regimes---stable absorption, latent fragility, critical transition and avalanche---and reveal a clear phase transition along the joint axis of external field intensity and redundancy stress. Tail behaviour shows regime-dependent thickening but no evidence of universal power-law criticality; the relevant description is finite-size avalanche dynamics on a real economic network.
The substantive contribution can be stated compactly. Conditional on the observed topology of WIOD intermediate flows, three structural ingredients---a Hall-like multiplicative stress-conversion function, a dissipative sandpile threshold dynamics, and a controlled grid of external fields---suffice to generate an ordered regime structure with a positively sloped phase frontier. The frontier is not an artefact of the choice of normalisation, the regularisation, the burn-in or the replication count, as documented in the robustness analyses. The framework therefore offers a tractable bridge between the statistical-physics tradition of self-organised criticality and the macroeconomic literature on network origins of aggregate fluctuations.
The framework opens at least three avenues for further work. First, extending the dataset to more recent input--output vintages and to finer sectoral disaggregations will sharpen the empirical relevance of the network calibration. Second, embedding measured shock proxies (energy-price changes, monetary-policy surprises, supply-chain disruption indices) in the external field $B_t$ would link the controlled-loading exercises performed here to historical episodes. Third, integrating the Hall-Sandpile mechanism with curvature-based geometric fragility approaches vallarino2024sandpile would produce a unified framework in which structural vulnerability and external loading interact to generate systemic events. A fourth direction, suggested by the finite-size discussion of Section (ref), is the systematic study of how the tail exponent $\alpha$ scales with the network size $|V|$ and the driving-rate ratio. We leave these extensions to future research.
The broader message is that economic systems do not merely transmit shocks, they convert them. A model that takes that conversion seriously, and disciplines it on observed network topology, offers a useful complement to existing approaches to systemic risk. The Hall analogy is structural, not literal; the sandpile metaphor is finite-size, not universal; and the WIOD substrate is observed, not synthesised. Within these caveats, the resulting picture is one in which fragility is contingent on the joint configuration of external loading and structural buffers, and in which policy leverage operates on stress-conversion coefficients rather than on shocks themselves.
The study uses publicly available WIOD input--output tables timmer2015illustrated,timmer2016anatomy and reproducible Python simulation code. The external fields are simulated. Replication scripts, processed tables and figures will be made available in a public repository upon publication.
The author declares no competing interests. The views expressed in this paper are those of the author and do not necessarily represent those of the Inter-American Development Bank, IDB Invest, their Boards of Directors, or the countries they represent.
The author thanks colleagues for useful discussions on the framing of the Hall-Sandpile mechanism and on the interpretation of the WIOD calibration. All remaining errors are the author's own.