The research, led by Kaihao Liang and Yuqiu Chen together with Tinghong Guo, Wenfeng Zhang and Chaolong Zhang, focuses on the Chinese stock market and examines the topological consequences of two major geopolitical shocks: the Palestinian-Israeli conflict and the Russia-Ukraine war. Rather than simply comparing market behavior before and after these events, the team devised a cleverer control strategy. They trained a deep learning model on pre-conflict market data and used it to predict what stock prices and trading volumes would have looked like in a hypothetical world where the conflicts never happened. They then built two networks, one from the real post-conflict data and one from the predicted conflict-free data, and compared their structures directly. The difference between those two networks isolates, as cleanly as the authors argue is possible, the fingerprint of geopolitical conflict on market connectivity.

The foundation of the analysis lies in how the network itself is constructed. Each listed company’s stock serves as a node, and connections between nodes are established using the Pearson correlation coefficient computed from two financial time series: the logarithmic rate of return of each stock and its trading volume ratio. Logarithmic returns are the standard currency of financial network analysis because they normalize price changes across firms of very different sizes, making cross-company comparison meaningful. Trading volume adds a second, complementary dimension, capturing the intensity of market activity rather than just its direction. By applying a threshold method, the researchers kept only the correlations strong enough to count as genuine relationships, pruning away the noise that would otherwise make every stock appear connected to every other stock.

The predictive engine at the heart of the counterfactual experiment is a hybrid architecture the authors call CNN-BiLSTM-XGB. This combines three complementary machine learning approaches. A convolutional neural network, or CNN, excels at extracting local patterns and features from sequential data, essentially scanning windows of historical prices and volumes for recurring signatures. A bidirectional long short-term memory network, or BiLSTM, is a form of recurrent neural network that processes sequences in both forward and reverse directions, allowing it to capture temporal dependencies that unfold over longer horizons while retaining context from both past and future points in the training window. Finally, XGBoost, a gradient-boosted decision tree algorithm, takes the learned representations and refines the final prediction, often capturing nonlinear interactions that neural networks alone miss. According to the study, this combined model delivered the best prediction performance among the configurations tested, effectively capturing the market dynamics that prevailed before the conflicts broke out.

That predictive accuracy matters because the entire logic of the study rests on it. If the model could not faithfully reproduce pre-conflict market behavior, its extrapolations into the conflict period would be meaningless. With a well-validated forecaster in hand, the researchers could treat the predicted conflict-free network as a baseline: a picture of how the market would have evolved under ordinary conditions. Any structural divergence between this baseline and the network built from actual post-conflict data can then be attributed to the shock itself rather than to ordinary market evolution. This counterfactual design addresses a persistent weakness in earlier studies of financial contagion, which typically compared pre-crisis and post-crisis networks without accounting for the natural drift of market structure over time.

The comparison produced striking results. After the conflicts broke out, the real association network displayed substantially more negative connections than the predicted network, meaning that more pairs of stocks moved in opposite directions than the conflict-free model anticipated. The real network also showed weaker inter-firm correlations overall, indicating that the generalized comovement that normally binds markets together had been disrupted and reorganized. In plain terms, the conflicts did not simply make the market fall together; they reshuffled which companies rose and fell in tandem, introducing new patterns of divergence and opposition across the market. The authors interpret this as evidence that the conflicts genuinely reshaped the correlation structure of the Chinese stock market rather than merely amplifying pre-existing trends.

Equally significant were the changes in the network’s global topology. When the researchers compared the real post-conflict network against both the pre-conflict network and the predicted conflict-free network, they found that the average weighted degree and the graph density both increased. Average weighted degree measures how strongly, on average, each node is connected to its neighbors, weighting each link by the strength of the correlation it represents. Graph density measures the fraction of all possible connections that actually exist in the network. An increase in both quantities means that after the conflicts, more companies were linked to more other companies, and those links were on the whole stronger. The network became, in the authors’ words, more complex and tighter, a denser mesh of interdependence among listed firms.

Denser connectivity in a financial network is a double-edged sword, and the study is explicit about the dangerous edge. In network science, the structure of connections determines how disturbances propagate. A sparse network can localize a shock, letting it dissipate in a corner of the graph, whereas a dense network offers many parallel pathways along which distress can spread from one firm to the next until it reaches every part of the system. The observed tightening of the Chinese market’s corporate network under the influence of conflict-related shocks therefore signals an increase in risk linkages among listed companies, and the authors conclude that this heightened interconnection could potentially lead to increased systemic risk across the entire stock market. In other words, a war thousands of kilometers away can quietly convert a diversified equity market into a more fragile, more synchronized system in which trouble at one firm is more likely to become trouble everywhere.

The findings arrive at a moment when policymakers and investors are increasingly treating geopolitical risk as a first-order financial variable rather than a background nuisance. Previous research cited in the study has documented geopolitical risk effects on asset returns in Chinese markets, the impact of sudden shocks on stock market network structure during past crises, and the reconfiguration of financial networks during the COVID-19 pandemic and the Russia-Ukraine war. The present study extends this lineage by adding a rigorous counterfactual methodology and a state-of-the-art hybrid forecasting model, giving analysts a template they can apply to other markets and other shock types. The datasets generated and analyzed in the research have been made publicly available in the Zenodo repository, supporting reproducibility and further work by other teams.

For practitioners, the practical implications are concrete. Risk managers who monitor correlation matrices as a gauge of portfolio diversification should be alert to the structural shifts the study documents: during geopolitical crises, historical correlations become unreliable guides, negative co-movements proliferate, and the effective level of diversification inside an equity portfolio can silently erode as the underlying network densifies. Regulators concerned with systemic risk gain a new diagnostic lens, one that watches not just the volatility of individual stocks but the evolving architecture of their relationships. And for the broader field of complex systems science, the study demonstrates how deep learning can serve as a bridge between prediction and network analysis, generating the counterfactual baselines needed to attribute causal structural change in systems, like financial markets, where controlled experiments are impossible. As regional conflicts continue to punctuate the global landscape, the invisible wiring beneath market prices has become both measurable and, increasingly, a matter of financial stability.

Subject of Research: Topological evolution of stock market complex networks under regional geopolitical conflicts

Article Title: Dynamic evolution of complex network topology of listed companies under the Influence of local conflicts

Article References: Liang, K., Chen, Y., Guo, T., Zhang, W., & Zhang, C. (2026). Dynamic evolution of complex network topology of listed companies under the Influence of local conflicts. Cluster Computing, 29(15), Article 847. https://doi.org/10.1007/s10586-026-06664-y

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06664-y

Keywords: complex networks, stock market, systemic risk, geopolitical conflict, deep learning, CNN-BiLSTM-XGBoost, network topology, Chinese stock market, financial risk, Pearson correlation, risk propagation, time series forecasting

The research team, led by Xuena Liu and Haibin Liu, constructed a multilayer supply chain network centered on nine of the world’s dominant battery manufacturers, including CATL, BYD, LG Energy Solution, Panasonic, SK Innovation, CALB, Samsung SDI, Gotion, and EVE Energy. Using the FactSet Revere database of interfirm supply relationships, supplemented by corporate disclosures and stock-market volatility data, they built a directed, weighted network spanning mineral extraction, intermediate materials, battery assembly, electric vehicle manufacturing, and recycling. The timing could hardly be more pressing: global electric vehicle battery demand reached 950 gigawatt-hours in 2024 and is projected to exceed 3 terawatt-hours by 2030, while the International Energy Agency expects electric vehicles and energy storage to account for more than 90 percent of future lithium demand.

The network’s anatomy alone tells a striking story. All nine manufacturer-centered networks follow power-law degree distributions, meaning a few hub firms dominate connectivity while most firms maintain only a handful of relationships. The commodity trader Glencore, the Australian lithium producer Pilbara Minerals, and the Chinese materials firm CNGR each maintain business relationships with seven to nine core battery makers, forming what the authors call cross-layer super-hubs. Meanwhile, most shared nodes connect to only two manufacturers. The networks led by Samsung SDI and LG Energy Solution are strongly correlated with each other and with Panasonic’s, implying that a shock at a shared supplier could transmit simultaneous damage across multiple corporate ecosystems. By contrast, the SK Innovation- and Gotion-led networks show slightly negative correlations, hinting at potential substitutability at certain supply nodes.

When the researchers simulated shortages of increasing intensity, from 10 percent to a complete cutoff of supply, lithium emerged as the most damaging mineral by a clear margin, with network losses ranking lithium above cobalt above nickel under every scenario. This asymmetry reflects chemistry as much as geology. Lithium is an indispensable element in every lithium-ion battery, whereas cobalt is progressively being engineered out of newer cathode designs, and nickel benefits from abundant resources and diversified supply. At a shortage intensity of 0.9, the CATL-dominated network suffered losses of 1171.42 under lithium scarcity, compared with 909.05 for cobalt and 918.45 for nickel. The BYD-led network fared worst of all under lithium stress, reaching 2587.49, a consequence of its enormous lithium iron phosphate production volumes and vertically integrated business model.

Manufacturer-level losses varied dramatically, defying any simple ranking of corporate might. CATL, the world’s largest battery maker, recorded the highest losses across all three minerals, ranging from roughly 130 to 160 under lithium shortages, a pattern the authors attribute to its simultaneous operation of both LFP and high-nickel NMC pathways, which broadens its mineral exposure and concentrates its supplier base. SK Innovation’s network proved exceptionally vulnerable to nickel because its product portfolio is dominated by high-nickel NCM811 cells, while EVE Energy’s cobalt-containing cathode materials and stringent certification requirements concentrated its dependence on upstream refining hubs. Panasonic, by contrast, recorded the lowest losses and responded only weakly to changes in shortage intensity, reflecting its comparatively low structural exposure.

Tracing exactly where the damage came from produced some of the study’s most actionable findings. Under cobalt shortages, a single firm, Glencore, accounted for between 73.43 and 92.75 percent of CATL’s attributed losses, a near-single-source structure in which one disruption could rapidly amplify risk downstream. Nickel losses for CATL were dominated by Glencore and the Indonesian producer Trimegah Bangun Persada, the latter reflecting the localized supply concentration fostered by Indonesia’s nickel ore export ban. Lithium losses were far more dispersed, spread across Tianqi Lithium, Ganfeng Lithium, Yahua, and Albemarle. Path analysis of a total Tianqi Lithium shortage showed that two pathways, a direct cross-chain link to BYD and a long cascade through the midstream materials firm BTR, together accounted for roughly 70 percent of BYD’s total impact, with the cathode makers Hunan Yuneng and Ronbay emerging as the most frequently traversed intermediary nodes.

Perhaps the most surprising result concerns how the industry recovers. The researchers decomposed recovery into three mechanisms: direct restoration through existing supply relationships, indirect recovery through shared nodes that connect multiple manufacturer networks, and recovery through newly established supply links. Indirect recovery dominated overwhelmingly, contributing 70 to 74 percent of total recovery, followed by new links at 15 to 18 percent, with direct recovery contributing a mere 10 to 12 percent. In other words, when a mineral shortage strikes, battery makers do not primarily heal through their own contracted suppliers; they heal through the network’s hidden cross-connections, as shared suppliers reallocate resources across tiers. EVE Energy relied most heavily on forging new relationships, at 35 to 40 percent, while Gotion addressed roughly 40 percent of its shortfalls through existing ties. Recovery pathways themselves were highly concentrated, repeatedly routing through a small group of firms including Hunan Zhongke Electric, Hunan Yuneng, and Tianqi Lithium.

To quantify resilience across the full disruption-and-recovery cycle, the team borrowed the marginal-response logic of price elasticity of demand, regressing each manufacturer’s cumulative loss rate on shortage intensity to obtain a resilience response coefficient. The results exposed a crucial disconnect between loss magnitude and loss sensitivity. Under lithium shortages, CATL combined the highest cumulative losses with the longest recovery requirement, roughly 88 to 90 cycles, and a significantly positive response coefficient of 0.071. Under cobalt shortages, CALB showed the strongest marginal response at 0.420, while under nickel shortages all nine manufacturers responded significantly, with Samsung SDI and EVE Energy posting the highest coefficients at 0.209 and 0.138 despite suffering relatively low cumulative losses. Samsung SDI even recorded a negative coefficient under lithium stress, indicating attenuating losses as intensity grew. High loss, the study concludes, does not necessarily coincide with high sensitivity, and loss rankings alone cannot capture corporate vulnerability.

The findings survived an unusually thorough battery of robustness tests. Inventory buffering of 90 days cut high-end loss rates by up to 24.7 percent under cobalt stress but left manufacturer rankings essentially unchanged, with Spearman correlations above 0.93. Convex nonlinear propagation, meant to mimic order-batching and capacity constraints, raised median cumulative losses by 51 percent yet preserved both the lithium-over-cobalt-over-nickel ordering and the vulnerability rankings. Across 1,500 Monte Carlo parameter combinations, indirect recovery remained dominant with average contributions near 79 percent. Parameterized adaptation, simulating strategic inventories, material substitution, chemistry switching, and contract renegotiation, cut mean cumulative losses by up to 34.7 percent and recovery cycles by 47.4 percent. Recycling networks built around GEM and Umicore already reduced system losses by 10.35 percent under cobalt and 9.69 percent under nickel at full shortage intensity, though strengthening recycling paradoxically increased cobalt losses in some scenarios by rerouting propagation pathways.

The implications reach well beyond corporate boardrooms. The authors argue that critical mineral governance must extend beyond counting reserves and import dependence to monitor midstream processors, shared suppliers, and the deep interfirm relationships through which risk actually travels. Because a small number of pathways contribute disproportionately to downstream losses, targeted interventions such as dual sourcing, pre-qualified backup suppliers, and inventory buffers along dominant routes offer more leverage than blanket diversification. For the energy transition as a whole, the message is sobering: the feasibility of electrifying transport depends not only on how much lithium, cobalt, and nickel the world can extract, but on the resilience of a network whose most important nodes are often invisible until the moment they fail.

Subject of Research: Resilience of lithium-ion battery supply chains to critical mineral shortages

Article Title: Resilience assessment of lithium ion battery supply chains under critical mineral shortages

Article References: Liu, X., Huang, H., An, F., Dong, X., Fang, W., Wang, Y., Zhang, H., Liu, Y., & Liu, H. (2026). Resilience assessment of lithium ion battery supply chains under critical mineral shortages. iScience, 29(10), Article 117703. https://doi.org/10.1016/j.isci.2026.117703

Image Credits: AI Generated

DOI: 10.1016/j.isci.2026.117703

Keywords: lithium-ion batteries, supply chain resilience, critical minerals, lithium, cobalt, nickel, electric vehicles, risk propagation, network analysis, CATL, BYD, energy transition