ORIGINAL RESEARCH article

Front. Sustain. Food Syst., 14 October 2026

Sec. Agricultural and Food Economics

Volume 10 - 2026 | https://doi.org/10.3389/fsufs.2026.1949662

Climate policy uncertainty and domestic–international grain-market risk spillovers: a TVP-VAR-BK analysis with machine learning early warning

  • School of Business Administration, Chongqing Vocational and Technical University of Mechatronics, Chongqing, China

Abstract

Background:

Climate-policy transition uncertainty may affect grain-market risk transmission.

Methods:

We use 2010–2025 daily data for Chinese and US corn and wheat, US soybean landed cost, and GCPU in a TVP-VAR-BK connectedness model and machine-learning warning exercise.

Results:

TCI averages 15.29% and peaks at 57.32%; short-horizon transmission dominates; XGBoost achieves PR-AUC of 0.986/0.949/0.781 at 1/5/20-period horizons.

Conclusion:

Climate-policy uncertainty is dynamically associated with grain-market risk and can support market-stability monitoring, but the evidence is associational rather than causal.

1 Introduction

Food security is a multidimensional outcome encompassing food availability, access, utilization, and stability. Climate-related threats to those dimensions arise through at least two conceptually distinct channels. Physical climate risk refers to heat, drought, floods, and other hazards that directly affect yields and supply. Transition risk refers to uncertainty created by climate-policy design and implementation, including carbon pricing, bioenergy mandates, and carbon-border measures. Grain-price risk is narrower still: it concerns volatility and the transmission of price shocks across markets. This study uses GCPU to measure transition-policy uncertainty, not physical climate hazards, and directly analyzes grain-price risk. Its food-security relevance therefore runs through the price-stability and import-cost channels; it does not estimate food availability, affordability, or nutritional outcomes themselves (Schlenker and Roberts, 2009; Wheeler and von Braun, 2013; Gavriilidis, 2021).

Grain markets have become more financialized and more integrated over the same period. That gives an international price shock more ways to reach a domestic market. China is a useful case. China grows almost all the corn and wheat it consumes, but it imports most of its soybeans, so the landed cost of soybeans tracks international prices closely. Corn is the less obvious case. After the procurement and storage system was reformed and prices were allowed to move with the market, domestic corn prices began responding to international prices much more than before (Ceballos et al., 2017; Qi et al., 2025). Several shocks then arrived in quick succession after 2020. Supply chains broke down during the pandemic, Black Sea trade stopped after the Russia–Ukraine conflict began in 2022, and extreme weather events became more frequent. International grain prices rose and fell sharply each time. Many countries also restricted grain exports, which made cross-market contagion easier (Falkendal et al., 2021; Behnassi and El Haiba, 2022; Mottaleb et al., 2022). For food security policy it therefore matters where risk enters this system, how fast it travels, and whether the pattern can be detected before prices move.

Three problems remain in this literature. First, GCPU is usually treated as an external explanatory factor rather than as a node whose dynamic association with domestic and international grain prices can be traced within one network. Second, existing studies rarely combine Chinese and US corn and wheat with a policy-relevant soybean import-cost measure and then separate transmission by frequency. Third, connectedness measurement and forward-looking warning remain disconnected. These gaps concern market-risk monitoring; any implications for food security must be interpreted through the stability of prices and import costs rather than as direct evidence on food-security outcomes.

We address these gaps with daily data from 2010 to 2025 on Chinese corn, US corn, Chinese wheat, US wheat, the landed cost of imported US soybeans, and GCPU. A TVP-VAR-BK framework measures total, directional, net, and frequency-specific connectedness. Rolling TCI quantiles then define high-risk states. We compare a TCI-only Logit benchmark with a full-feature Logit, Random Forest, and XGBoost at 1-, 5-, and 20-period horizons, using PR-AUC as the primary out-of-sample metric and calibration and confusion diagnostics as complements.

The paper makes three contributions. First, it places GCPU within a domestic–international grain-price network while explicitly limiting inference to climate-policy uncertainty and market-risk transmission. Second, the TVP-VAR-BK decomposition identifies whether connectedness is associated mainly with rapid contagion or more persistent adjustment. Third, the warning module links backward-looking connectedness measurement to out-of-sample classification, evaluates its incremental value against a TCI-only benchmark, and reports discrimination, calibration, and operating-point diagnostics.

The paper proceeds as follows. Section 2 reviews the literature and defines the research gap. Section 3 introduces the data. Section 4 sets out variable selection, the TVP-VAR-BK model, and the warning design. Section 5 reports the connectedness evidence, Section 6 presents the out-of-sample warning analysis, Section 7 reports robustness checks and limitations, and Section 8 concludes with qualified policy implications.

2 Literature review

2.1 Climate change, climate risk, and food security

Agricultural and environmental economists have worked on the climate and food security link for a long time. The early studies looked at production. Schlenker and Roberts (2009) show that the effect of temperature on US crop yields is nonlinear: yields hold up until a threshold is passed and then fall quickly. Lobell et al. (2011) estimate that climate trends since 1980 reduced global wheat output by about 5.5% and global corn output by about 3.8%. Food security depends on having enough grain and on supply being stable, so repeated yield losses affect both conditions. Later studies added other channels. Wheeler and von Braun (2013) argue that climate change reaches food security through falling production, higher prices and less stable supply, and that low-latitude developing countries suffer most because they depend more on agriculture and have less capacity to adapt.

Research then moved from yields to macroeconomic and price effects, and the definition of food security widened at the same time. It now covers whether food is available, whether people can afford it, and whether the system is stable. Two groups of results matter here. The first group is about quantities. Hasegawa et al. (2018, 2021) find that extreme climate events raise the probability of food insecurity worldwide and increase adaptation needs in the affected regions. They also find that mitigation policy can make things worse if it is poorly designed or too strict, because production costs and food prices rise. Kahn et al. (2021) show that sustained temperature deviations from historical norms lower long-run per capita output growth, so the cost of climate risk builds up over time instead of passing quickly. The second group is about prices. Kotz et al. (2024) estimate that warming and heat extremes will add between 0.9 and 3.2 percentage points to annual global food price growth by 2035, which makes affordability a channel of its own. Trade policy can amplify these effects. Falkendal et al. (2021) show that food export restrictions during the COVID-19 pandemic increased food insecurity in low- and middle-income countries. Taken as a whole, this work suggests that climate risk does not stay at the production stage. It moves across markets and regions, which is why a risk-spillover approach is appropriate here.

Recent research also identifies institutional and technological adaptation pathways. Using 2010–2024 panel data for 279 Chinese prefecture-level cities, Lu et al. (2026) find that the synergy between land consolidation and smart agriculture supports food security and agricultural carbon neutrality, with positive indirect and total spatial effects on food security. This supply-side evidence underscores the roles of production capacity, digital infrastructure, and low-carbon transformation. The present study complements that perspective by examining high-frequency price-risk transmission and market-stability warning.

2.2 Climate policy uncertainty and commodity markets

Physical risk is not the only channel. Climate policy changes often, and those changes are themselves a source of risk. This is usually called transition risk. Carbon-neutrality targets, bioenergy subsidies, carbon pricing and carbon tariffs are all hard to predict, both when they are being designed and when they are being implemented. The uncertainty changes cost structures in energy and agriculture, and it then reaches commodity markets, grain prices and food security. Gavriilidis (2021) built a Climate Policy Uncertainty (CPU) index by counting relevant terms in major US newspapers, using the same news-based method that Baker et al. (2016) used for economic policy uncertainty. Bouri et al. (2022) find that this index helps explain and predict asset prices, and that the effect is strongest when markets are under stress. Related connectedness studies examine climate-policy uncertainty jointly with energy, green-bond and carbon markets (Wang et al., 2023b), and with coal and carbon prices in China (Yan and Cheung, 2023). These studies motivate examining cross-market channels while distinguishing the selected uncertainty measure from physical climate risk.

Evidence on agricultural and other commodity markets increasingly places climate policy uncertainty inside broader price-risk transmission systems. Wang et al. (2023a) find that connectedness among climate policy uncertainty, crude oil, and agricultural commodities rises under extreme market conditions, particularly for corn and soybeans. Zeng et al. (2025) find stronger grain-market spillovers in extreme quantiles than near the median and heterogeneous time–frequency relationships with US climate-policy uncertainty. Recent 2026 research extends this agenda: Shi et al. (2026) document links between climate policy uncertainty and agricultural green development in China. Zhang et al. (2026), using monthly data for 2008–2025, report changing connectedness across energy and agricultural markets and different short- and longer-horizon roles for policy uncertainty. Their monthly frequency and broader commodity coverage differ from the present aligned daily domestic–international grain network; their horizon labels should not be equated directly with ours. Together, these studies establish climate-policy uncertainty as a relevant market-risk factor, but they do not imply that it is equivalent to physical climate risk or that connectedness estimates identify causal food-security effects.

2.3 Risk spillovers in grain markets and international–domestic price transmission

Spillover research on grain markets began with energy. The worry was that oil price movements would reach food security through production costs and biofuel linkages, and the evidence bore out the concern: volatility spillovers between crude oil and agricultural commodity markets rose markedly after the 2008 global food crisis. Oil shocks of different origin affect farm prices differently, and the bidirectional spillovers linking crude oil, precious metals and agricultural futures grow larger in crises (Nazlioglu et al., 2013; Wang et al., 2014; Kang et al., 2017).

From a frequency-domain perspective, studies have found that the relationship between crude oil and agricultural commodities exhibits asymmetric characteristics across different frequency bands, with distinct short-term and long-term transmission mechanisms. This implies that the sources of grain price risks should be differentiated according to their frequency characteristics (Kang et al., 2019). In recent years, major risk events have become a central focus of spillover research. Compared with the pandemic period, volatility connectedness between fossil energy and agricultural commodities in the long-term frequency band increased significantly during the Russia–Ukraine conflict, as conflict-related uncertainty substantially amplified volatility spillovers between energy and agricultural markets (Wu et al., 2023; Vo and Tran, 2024). Geopolitical risks not only exert significant causal effects on the prices of key staple foods such as wheat (Saâdaoui et al., 2022), but also transmit risks to international agricultural markets through multiple channels, directly affecting global food security (Ren et al., 2025).

Evidence from the climate dimension has also accumulated rapidly. Climate risks significantly reshape the transmission structure of price risks within the water–energy–food nexus (Le et al., 2023). Their impacts on global food and agricultural markets are concentrated mainly in the short-term horizon and under extreme conditions, highlighting the sudden and tail-risk characteristics of climate shocks to food security (Khalfaoui et al., 2024). Guo et al. (2023) also examine the connections between climate-risk shocks and agricultural markets, reinforcing the need to distinguish the risk proxy and commodity coverage used in each study.

Another closely related strand of literature examines the transmission of international grain prices to domestic markets and the linkages between Chinese and international grain markets, which directly determines a country’s external exposure to food security risks. Cross-country evidence demonstrates significant heterogeneity in the transmission of international grain price fluctuations to domestic markets in developing countries (Ceballos et al., 2017). For China, the US crude oil market exhibits significant high-frequency volatility spillovers to Chinese agricultural futures markets (Luo and Ji, 2018), while extreme risk spillovers between Chinese and US agricultural futures markets during crisis periods display clear state-dependent characteristics (Hu et al., 2024).

The size of the effect of international prices on Chinese grain prices differs by commodity, and the difference follows how open each market is. Soybeans are affected most, because China imports most of its soybeans. Corn comes next; it has become more responsive to international markets since procurement and reserve reform. Wheat and rice are affected least, because self-sufficiency is higher and policy protection is stronger. Qi et al. (2025) describe this pattern as absolute security in staple grains combined with heavy dependence on soybean imports. Lu et al. (2025) find that better reserve adjustment mechanisms and more diversified import sources are the main ways to reduce the effect of external price shocks.

There is substantial evidence that risk linkages between domestic and international grain markets exist and vary over time. The remaining gap lies in coverage. Most studies focus on international markets or trace price transmission for one commodity at a time. Few studies combine Chinese and US corn and wheat with soybean import costs in a single frequency-domain spillover network, and fewer still examine how climate-policy uncertainty is associated with price-risk transmission in that network.

2.4 Risk connectedness measurement methods and machine learning-based early warning

Methods for measuring risk spillovers have changed in two ways. They moved from static to dynamic, and from the time domain to the frequency domain. Both changes make it easier to describe how risk in a food system travels. Spillover indices and connectedness networks based on generalized forecast error variance decomposition are now standard in cross-market risk research, because they show which markets send risk and which markets receive it (Pesaran and Shin, 1998; Diebold and Yılmaz, 2012, 2014).

Time-varying parameter VAR models solved two problems with rolling windows. They no longer discard part of the sample, and they no longer require the researcher to pick a window width. Connectedness estimated in this way follows changes in grain market linkages more closely (Primiceri, 2005; Antonakakis et al., 2020). Frequency-domain decomposition does something else. It separates short-horizon contagion from medium- and long-run structural dependence, so it can tell whether grain price risk is fast-moving or slow-moving (Baruník and Křehlík, 2018).

The TVP-VAR-BK framework combines these two ideas. It is now used widely for time-frequency spillover analysis in commodity and grain markets, and it suits the study of risk transmission in food systems under climate shocks (Kang et al., 2019; Ren et al., 2025; Zhang et al., 2026).

Work on early warning developed separately. Machine learning models deal well with nonlinearity and with interactions among many variables, so Random Forest and XGBoost are now common in classification-based warning problems (Breiman, 2001; Chen and Guestrin, 2016).

Machine learning has also been applied directly to food-security outcomes. Lentz et al. (2019), Westerveld et al. (2021), Martini et al. (2022), and Balashankar et al. (2023) show that data-driven models can improve crisis classification when household, satellite, or news information is available. Ferrianta and Yanti (2026) combine Random Forest and XGBoost with dynamic panel methods to forecast a multidimensional food-security vulnerability index in Asia-Pacific economies. Those studies predict food insecurity or vulnerability directly. By contrast, our target is a market-based systemic-connectedness state, which is narrower but can be updated at daily frequency.

Recent work also emphasizes the conditions under which predictive models can support decisions. Anderson et al. (2026) argue that artificial intelligence and machine learning should be used selectively in food-security early warning, supported by expert oversight and transparent governance. This reinforces the distinction between the present high-connectedness warning target and direct humanitarian food-insecurity assessment, and motivates reporting calibration and simple benchmarks alongside discrimination metrics.

These studies target food-insecure populations or crisis events, however, not market-based systemic risk. Connectedness measures, for their part, quantify spillovers accurately but say nothing about what comes next. The gap we set out to close lies between them: a framework that runs from spillover measurement through threshold identification to out-of-sample warning, and that gives food security monitoring something forward-looking to work with.

2.5 Literature review and research gaps

Four literatures bear on the question of this paper: physical climate risk and food security, climate-policy uncertainty and commodity markets, grain-market spillovers, and machine learning warning systems. Their outcome concepts differ, and the present study occupies the intersection concerned with climate-policy uncertainty, grain-price stability, and market-based warning. Three gaps remain.

First, climate-policy uncertainty is usually entered as an external regressor; treating GCPU as an endogenous node in a domestic–international grain-price network remains uncommon. This positioning permits measurement of dynamic association and transmission but not causal attribution.

Second, existing spillover studies on China’s food security rarely simultaneously cover major staple grains from both China and the United States together with soybean import costs. Consequently, they fail to fully characterize China’s external exposure to grain price risks and cannot adequately address the structural concern of “staple grain security versus soybean import security.”

Third, risk measurement and early warning remain largely disconnected. Connectedness studies often stop at ex post description, while food-security warning studies usually predict lower-frequency humanitarian outcomes rather than high-frequency market connectedness. A transparent bridge between the two requires time-ordered validation, a simple benchmark, and calibration diagnostics.

We therefore place GCPU and major domestic and international grain-price changes in one TVP-VAR-BK system and build a rolling-quantile warning model for high-TCI states. The resulting framework is designed for grain-market stability monitoring. It informs one component of food-security governance but should not be read as a complete measure of national food security.

3 Data sources, variable processing, and descriptive statistics

The daily sample contains six series: Chinese corn, US corn, Chinese wheat, US wheat, the landed cost of imported US soybeans, and GCPU. It runs from 7 May 2010 to 18 December 2025 and contains 3,052 observations. GCPU is obtained from the China Energy and Climate Finance Network, and the price series are obtained from Wind. All variables enter as log differences. The labels CN Corn, US Corn, CN Wheat, and US Wheat refer to the corresponding national corn and wheat price series. US Soy Cost abbreviates US soybean landed cost, the cost series for importing US-origin soybeans into China, rather than a US soybean futures quotation. GCPU is an uncertainty index, not a price. The transformed variables are dimensionless log changes.

Throughout the analysis, one period means one observation in the final aligned daily sample. It is an observation step, not necessarily one calendar day or one trading day on either national exchange. Consequently, all forecast horizons, rolling windows, and frequency bands are measured in aligned-observation periods; calendar gaps between retained observations need not be equal.

Log differencing makes the empirical scope explicit: the model measures connectedness among daily changes in prices and climate-policy uncertainty, not long-run equilibrium relations in levels and not physical climate shocks. This transformation mitigates nonstationary trends and matches the study’s focus on how market risk propagates and decays over short horizons.

Table 1 reports the descriptive statistics of the log-differenced series. Their sample means are close to zero, while their standard deviations differ substantially. GCPU has a standard deviation of 1.1033, which exceeds those of the price changes; this comparison concerns the variability of the transformed series rather than differences in their original measurement units. Kurtosis exceeds three for all six series, and the Jarque–Bera tests reject normality at the 1% level. The ADF tests reject the unit-root null at the 1% level for all six log-differenced variables. These results support modeling the transformed series, while the TVP-VAR allows their transmission relationships to change over time.

Table 1

VariableMeanMaxMinSDSkewnessKurtosisJ–BADF
CN Corn−0.00010.0245−0.12340.0037−11.4728402.81262.04 × 107***−10.794***
US Corn−0.00010.4782−0.49880.0267−1.3585118.46961.70 × 106***−14.463***
CN Wheat−0.00010.0485−0.04710.00310.235159.86654.11 × 105***−12.307***
US Wheat0.00000.3378−0.30490.0244−0.300435.91721.38 × 105***−14.795***
US Soy Cost−0.00010.6319−0.30560.02097.6360305.66981.17 × 107***−17.959***
GCPU0.00015.4316−5.37281.1033−0.03044.06451.45 × 102***−23.634***

Descriptive statistics.

Statistics refer to the log-differenced series. SD, standard deviation; J–B, Jarque–Bera test of the null of normality; ADF, augmented Dickey–Fuller test of the null of a unit root. ***, **, and * denote rejection at the 1%, 5%, and 10% significance levels, respectively. The J–B and ADF results for all six variables are marked *** to denote rejection at the 1% level.

4 Model construction

The model has three parts and we estimate them in order. First we estimate a TVP-VAR, which allows the relationships among the six variables to change over time. Then we apply BK frequency-domain connectedness, which splits the spillovers into frequency bands. Finally we convert the Total Connectedness Index (TCI) into a high-risk state, which is what the warning model predicts.

4.1 Variable selection and research objects

Variable selection follows three policy-relevant criteria. Chinese corn and wheat represent domestically important staple and feed-grain markets; US corn and wheat provide internationally traded benchmark prices with direct relevance to global price discovery; and the landed cost of imported US soybeans represents China’s import-cost exposure in a commodity for which external dependence is structurally high. Rice is excluded because its domestic market is more policy insulated and less directly comparable with the selected US benchmarks.

GCPU is included as a news-based measure of transition-policy uncertainty because climate-policy changes can alter energy, transport, biofuel, investment, and production-cost expectations before any physical climate event is observed. The six-variable network therefore covers domestic staple prices, international benchmarks, a bilateral soybean import-cost channel, and climate-policy uncertainty. It is deliberately parsimonious rather than exhaustive, and the limitations of omitted markets and alternative proxies are discussed in Section 7.4.

The 2018–2020 US–China trade war is an important competing explanation for movements in the bilateral soybean landed cost series. China’s 2018 retaliatory tariff sharply reduced purchases of US soybeans and redirected demand toward Brazil (Gale et al., 2019). Accordingly, connectedness during that period may combine tariff, supplier-substitution, exchange-rate, freight, and climate-policy-uncertainty channels. We retain the US soybean landed cost series because it directly captures the stressed bilateral route, but we do not attribute the episode to GCPU alone.

4.2 TVP-VAR model

Risk transmission through the grain-price system evolves over time, so we use a time-varying parameter vector autoregression (TVP-VAR) to track changing interdependence among the included markets and GCPU. Let y collect Chinese corn, US corn, Chinese wheat, US wheat, the landed cost of imported US soybeans, and GCPU. A TVP-VAR of order p then takes the following form, where A is the time-varying coefficient matrix, a the state vector obtained by vectorizing that matrix, and the disturbance covariance appears in the matrix notation below.

We set the lag order to 1, the forecast horizon for the variance decomposition to 100 periods, and both forgetting factors in the time-varying estimation to 0.99. That combination respects the sampling frequency, keeps the parameters stable, and still leaves enough information for the decomposition. The advantage over a fixed-parameter VAR is straightforward: letting the coefficient matrix and the shock covariance move with time makes structural change in transmission visible from one shock episode to the next.

4.3 Generalized forecast error variance decomposition

Risk contributions are then measured by generalized forecast error variance decomposition (GFEVD). Unlike orthogonalized decomposition, GFEVD is invariant to the ordering of variables, so no result depends on how we happen to arrange them. Let θ be the share of the H-step forecast error variance of variable i attributable to a shock to variable j. Row sums under the generalized decomposition need not equal one, so we normalize each row to sum to one.

Read as a risk transmission matrix, the normalized decomposition is easy to interpret. Large off-diagonal entries mean strong cross-market transmission. Large diagonal entries mean a variable’s volatility is mostly self-generated, and that it spills comparatively little into the rest of the system.

4.4 Construction of the risk spillover indices

Following Diebold and Yılmaz, we build spillover measures at four levels: total, directional, net directional, and pairwise net directional. The total connectedness index (TCI) is the average cross-market risk linkage across all variables in the system. A higher TCI means the system as a whole is more contagious.

TO and FROM measure direction. TO is the risk a variable sends to the rest of the system, and FROM is the risk it receives. NET is TO minus FROM. A variable with positive NET is a net transmitter, and a variable with negative NET is a net receiver. NPDC measures direction between two specific variables.

4.5 BK frequency-domain decomposition of risk spillovers

We use the Baruník–Křehlík decomposition to find the cyclical origin of the spillovers. This method assigns forecast error variance contributions to different frequency bands, which tells us whether transmission comes mainly from short-term price shocks or from medium- and long-term structural links.

We split the spectrum into three bands, chosen to fit the observation frequency of the data and the way risk moves in these markets: 1–5 periods, 5–20 periods, and more than 20 periods. The first band is short-term contagion, which comes from unexpected events, market sentiment and daily price movement. The second band is medium-term adjustment, which comes from trade flows, inventories, costs and supply and demand expectations resetting. The third band is structural and reflects climate risk, the international supply and demand situation, and continuing food security pressure. Here, periods are aligned daily observation steps as defined in Section 3; the band labels do not imply fixed calendar-day durations.

4.6 Machine learning-based risk early warning model

On top of these spillover measures we build the warning model. Instead of forecasting the TCI as a point value, we pose a binary classification problem: standing at forecast origin t with only the information available then, will the system enter a high-risk connectedness state at some point over the next h observation periods?

At each forecast origin t, the preceding 252 TCI observations, excluding t, determine the 80th- and 90th-percentile thresholds for general and extreme high-risk states. A label equals one if TCI reaches or exceeds the time-t threshold at least once during the next h observations, where h = 1, 5, or 20. The warning exercise therefore predicts entry into a relative high-connectedness regime rather than a fixed TCI level. The TVP-VAR specification, generalized forecast error variance decomposition and normalization, connectedness indices, frequency-domain decomposition, and warning-label definitions are summarized in Equations 1–12.

5 Analysis of frequency-domain risk spillover effects

5.1 Total spillover effects and frequency-domain decomposition

We begin with the aggregate picture. Over the full sample the TCI averaged 15.29% and peaked at 57.32%. The domestic and international grain price system was therefore not tightly connected throughout; risk built up instead during particular shock episodes.

The peak occurs around 2021. Global shocks, supply chain breakdowns and changes in price expectations all happened in that period, and they probably reinforced each other.

Figure 1 and Table 2 show that most spillover activity is short-term. The 1–5-period band accounts for 73.27% of average total connectedness. The 5–20-period band accounts for 18.78%, and the band beyond 20 periods accounts for 7.96%. Risk in this system therefore moves fast, spreads fast, and comes in episodes. The off-diagonal entries of Table 2 are largest among international grain prices, import costs and domestic staple prices. We look at direction next.

Figure 1

Table 2

VariableCN CornUS CornCN WheatUS WheatUS Soy CostGCPUFROM
Short (1–5 periods)
CN Corn49.461.032.990.981.220.476.68
US Corn1.1562.830.8912.344.060.5619.00
CN Wheat3.040.7349.390.490.950.245.45
US Wheat1.2412.860.8763.073.020.6918.68
US Soy Cost1.486.01.314.2167.680.5513.57
GCPU0.571.00.480.980.889.353.84
TO7.4921.626.5519.0110.052.52TCI
NET0.812.621.090.32−3.52−1.3211.21
Medium (5–20 periods)
CN Corn26.480.542.080.420.570.133.74
US Corn0.459.750.42.070.60.063.58
CN Wheat1.850.3627.780.250.380.082.92
US Wheat0.462.170.399.740.480.093.59
US Soy Cost0.381.490.251.0210.560.063.20
GCPU0.020.050.030.050.034.870.20
TO3.174.613.163.812.070.41TCI
NET−0.571.030.240.22−1.140.222.87
Long (>20 periods)
CN Corn11.780.261.060.180.280.061.85
US Corn0.23.450.190.740.230.021.39
CN Wheat0.860.1713.110.120.170.041.35
US Wheat0.220.780.23.480.20.031.44
US Soy Cost0.160.550.110.373.770.021.21
GCPU0.010.020.010.020.011.680.07
TO1.451.781.571.430.90.17TCI
NET−0.40.40.21−0.01−0.310.11.22

Frequency-domain risk spillover matrix.

5.2 Directional spillover index (TO)

TO measures how much risk a variable sends to the rest of the system. Figure 2 shows that US corn and US wheat have high TO values through most of the sample, so international prices are a strong external shock to domestic grain markets. TO for Chinese corn and Chinese wheat rises sharply after 2020. Domestic markets therefore send price risk out in some periods and are not only receivers.

Figure 2

5.3 Directional spillover index (FROM)

FROM measures how much risk a variable receives. Figure 3 shows large FROM values for the landed cost of imported soybeans and for Chinese corn in some periods. Import costs and domestic prices are exposed at the same time to international market movement, to changes in climate risk, and to links with other commodities. GCPU has a low FROM value on average but it fluctuates periodically. Climate policy uncertainty is therefore not fully exogenous to the grain price system, and there may be feedback from market risk when price volatility is high.

Figure 3

5.4 NET directional spillover index (NET)

NET is TO minus FROM and it summarizes the role of each variable. A positive value means the variable is a net transmitter of risk. A negative value means it is a net receiver.

Figure 4 shows that these roles change over time. The averages in Table 3 are steadier. US corn, Chinese wheat and US wheat are net transmitters on average. US corn has the highest value at 4.05%. The landed cost of imported soybeans has the lowest at −4.97%.

Figure 4

Table 3

VariableAverage NETRole
US Corn4.0501Net transmitter
CN Wheat1.5432Net transmitter
US Wheat0.5385Net transmitter
CN Corn−0.1601Net receiver
GCPU−1.0031Net receiver
US Soy Cost−4.9685Net receiver

Average net directional spillover index.

Import costs therefore absorb most of the systemic shocks, and international prices, especially US corn, are the source of them.

5.5 Pairwise net spillovers and the spillover network

Average NET gives the overall role of each variable but not the direction between any two variables, so we also compute the pairwise net directional spillover index (NPDC). Climate risk is the focus of this paper, so Figure 5 shows NPDC between GCPU and each grain price variable.

Figure 5

Net spillovers between GCPU and US corn, US wheat and Chinese wheat are concentrated in the short-term band. Climate policy uncertainty reaches the grain price system through short-cycle shocks rather than through sustained pressure. It does not have to raise total system connectedness for a long time, but when uncertainty rises suddenly it reaches prices quickly through expectations and trading.

The three network panels reveal horizon-specific structures (Figure 6; Table 4). In the 1–5-period network, relatively large edges connect US soybean landed cost with US corn and wheat, and the GCPU–US corn linkage is visible. These are associations within the generalized variance-decomposition network; they do not identify a structural causal effect of climate-policy uncertainty.

Figure 6

Table 4

BandFromToAverage weight
1–5 periodsUS Soy CostUS Corn1.9434
1–5 periodsUS Soy CostUS Wheat1.1956
1–5 periodsUS WheatUS Corn0.5213
1–5 periodsGCPUUS Corn0.4392
5–20 periodsUS Soy CostUS Corn0.8938
5–20 periodsUS Soy CostUS Wheat0.5383
5–20 periodsCN CornCN Wheat0.2353
5–20 periodsCN CornUS Soy Cost0.1847
>20 periodsUS Soy CostUS Corn0.3179
>20 periodsCN CornCN Wheat0.2037
>20 periodsUS Soy CostUS Wheat0.1703
>20 periodsCN CornUS Soy Cost0.1245

Major transmission edges of the spillover networks by frequency band.

In the 5–20-period network, linkages from soybean landed cost to US corn and wheat remain important, while connections involving Chinese corn, Chinese wheat, and soybean landed cost strengthen. Above 20 periods, soybean landed cost and Chinese corn remain relevant nodes, but average connectedness is much smaller than in the short band.

The network evidence supports joint monitoring of international grain prices, import costs, and GCPU. It should be interpreted as evidence of co-movement and risk transmission conditional on the included variables, not as proof that climate-policy uncertainty alone generated each episode.

6 Machine learning-based risk early warning analysis

6.1 Forecast target, rolling thresholds, and risk labels

The connectedness analysis identifies where market risk is concentrated ex post. The warning question is narrower and forward-looking: using only information dated at or before forecast origin t, will the system enter a high-TCI state over the next h observations? At each t, the previous 252 TCI observations define an 80th-percentile general-risk threshold and a 90th-percentile extreme-risk threshold. A high-risk label is assigned if the threshold available at t is crossed during the next h = 1, 5, or 20 observations (Figure 7).

Figure 7

6.2 Prediction features and model specification

The feature set contains 58 predictors in three groups. Market-state features comprise each of the six contemporaneous log differences, their first two lags, 5- and 20-period rolling standard deviations, and cross-sectional average absolute return, maximum absolute return, and dispersion. Connectedness-state features comprise total TCI, the short-band (1–5 periods), medium-band (5–20 periods), and long-band (more than 20 periods) TCI components, the rolling threshold, the TCI–threshold distance, and their ratio. Network features comprise the contemporaneous TO, FROM, and NET index for each variable.

We compare four specifications. The TCI-only Logit uses current TCI as a persistence benchmark. The full Logit uses all 58 features, while Random Forest and XGBoost allow nonlinearities and interactions. Random Forest uses 400 trees, mtry = floor(sqrt(58)) = 7, and inverse-frequency class weights. XGBoost uses 180 boosting rounds, learning rate 0.05, maximum depth 3, subsample and column-subsample ratios of 0.85, minimum child weight 3, and scale_pos_weight equal to the training ratio of negative to positive observations. These parsimonious settings were fixed before test evaluation; no post-test hyperparameter search was conducted.

For each horizon and risk threshold, the same fixed feature-construction and model settings are applied. Model selection is based on the prespecified primary metric, PR-AUC, and all reported test statistics come from the 2023–2025 block only. These specifications support reproducible evaluation without using information from the test period.

6.3 Out-of-sample early warning results

The split is chronological. Observations before 2022 form the training set, 2022 is the validation period, and 2023–2025 is the untouched test period; an h-observation embargo before each boundary prevents a future-horizon label from crossing into the next block. The validation period selects the operating cutoff by maximizing Youden’s J, which yields 0.50 in the reported runs. All rolling statistics and thresholds are backward-looking, and the threshold excludes the forecast-origin observation. Connectedness features are dated t and generated by the TVP-VAR filtering recursion. Rows lacking lag, volatility, or 252-period threshold information are removed; no future-data imputation, resampling, or test-set standardization is used. Random seeds are fixed at 20260727 and XGBoost is run with one thread. PR-AUC is primary because high-risk states are uncommon; ROC-AUC, Brier score, F1, accuracy, ECE, and confusion counts are reported as complementary diagnostics.

Table 5 shows that XGBoost has the highest PR-AUC at all three horizons under the 90th-percentile extreme-risk definition: 0.986, 0.949, and 0.781 at 1, 5, and 20 periods. The TCI-only benchmark reaches only 0.567, 0.527, and 0.340, so directional, frequency-specific, and market-state features add substantial discrimination. Random Forest slightly exceeds XGBoost’s F1 at the 1-period operating point and has marginally lower ECE at 20 periods, illustrating why no single metric should be interpreted in isolation (Figure 8).

Table 5

HorizonModelROC-AUCPR-AUCBrierF1ECE
1TCI-only Logit0.85890.56670.1055–0.0245
1Logit0.96430.79430.07190.62180.0715
1Random Forest0.99160.97530.01810.96970.0555
1XGBoost0.99640.98620.00890.94740.0120
5TCI-only Logit0.80880.52710.1221–0.0990
5Logit0.96190.83550.05510.75610.0390
5Random Forest0.95000.90620.03980.88440.0770
5XGBoost0.98120.94940.02270.88440.0206
20TCI-only Logit0.67520.34030.1410–0.0895
20Logit0.88750.58490.11400.52560.0962
20Random Forest0.91040.73780.08050.42860.0578
20XGBoost0.90000.78130.06440.72590.0621

Out-of-sample model comparison under the 90th-percentile extreme-risk threshold.

Horizons are measured in aligned-observation periods. ROC-AUC, area under the receiver operating characteristic curve; PR-AUC, area under the precision–recall curve; Brier, Brier score; F1, harmonic mean of precision and recall; ECE, expected calibration error. Classification metrics use a cutoff of 0.50. For TCI-only Logit, “–” indicates that no positive cases are predicted at this cutoff; precision is undefined and the reported F1 output is not defined under this reporting convention. It does not indicate unavailable model results. Under the alternative zero-division convention, F1 would be reported as zero.

Figure 8

At the selected 0.50 operating point, XGBoost accuracy is 0.987, 0.968, and 0.929 at 1, 5, and 20 periods, compared with no-event baselines of 0.875, 0.854, and 0.840. Accuracy is therefore reported only with its class-imbalance benchmark. The confusion counts in Table 6 show that missed events increase at 20 periods, consistent with the decline in PR-AUC.

Table 6

HorizonModelCutoffAccuracy/baselineTNFPFNTP
1XGBoost0.500.987/0.8754673463
5XGBoost0.500.968/0.85445141365
20XGBoost0.500.929/0.84043233449

XGBoost accuracy baselines and confusion counts.

TN, true negatives; FP, false positives; FN, false negatives; TP, true positives. The baseline always predicts the majority no-event class. Horizons are measured in aligned-observation periods. The test-sample totals, calculated as TN + FP + FN + TP, are 537, 533, and 518 for h = 1, 5, and 20, respectively. Labels require h subsequent observations, so longer horizons exclude more observations near the sample endpoint. The h-period embargo prevents labels from crossing split boundaries, while the lagged features, rolling volatility measures, and 252-period thresholds determine initial feature eligibility. Only observations with complete features and a valid horizon-specific label enter the evaluation; the horizons therefore need not have identical test-sample sizes.

Figure 9 reports 10-bin reliability diagrams for every model and horizon. XGBoost ECE is 0.012, 0.021, and 0.062 across the three horizons. The first two horizons are comparatively well calibrated, while the 20-period curve is less stable and should not be converted mechanically into policy action. The TCI-only benchmark can display modest ECE while failing to identify positive events at the 0.50 cutoff, which illustrates why calibration, discrimination, and confusion counts must be read together.

Figure 9

The risk-state definition is then relaxed to the rolling 80th percentile. XGBoost remains the best model by PR-AUC at all horizons, reaching 0.987, 0.955, and 0.926 at 1, 5, and 20 periods (Table 7). Performance therefore does not depend on the extreme-state threshold alone, although the task is easier when positive observations are more common.

Table 7

HorizonBest modelROC-AUCPR-AUCBrierF1Positive rate
1XGBoost0.99560.98750.01640.95540.2104
5XGBoost0.97880.95520.03790.89800.2420
20XGBoost0.96030.92590.07060.78880.2857

Best forecasting results under the 80th-percentile high-risk threshold.

Taken together, Tables 5–7 show that the warning model is most useful for classifying market-connectedness states; it does not replace the TVP-VAR-BK analysis of transmission mechanisms or constitute a direct forecast of food insecurity.

Figure 10 reports mean absolute SHAP (SHapley Additive exPlanations) values for the XGBoost models under the rolling 90th-percentile threshold. For each feature, the absolute values of its observation-level SHAP attributions are averaged over the observations being explained. A larger mean absolute SHAP value indicates a larger average contribution to the fitted model output relative to its baseline; it is not XGBoost split frequency or gain. SHAP summaries are reported separately for the three horizon-specific models; the existing panels are ordered h = 1, 20, and 5 periods from left to right. The TCI–threshold distance and ratio dominate the displayed rankings, with the rolling threshold, selected volatility measures, and directional connectedness measures also represented. Taking absolute values removes the direction of an attribution, so these rankings measure predictive importance rather than signed economic effects or causal contributions.

Figure 10

The prominence of the TCI–threshold distance and ratio also reflects the construction of a threshold-based forecast target. Their importance should therefore be assessed together with the TCI-only benchmark and the chronological test design. The “(Intercept)” entry denotes the model baseline and should not be interpreted as an explanatory variable.

Using 18 December 2025 as the forecast origin, XGBoost assigns probabilities of 0.004, 0.013, and 0.024 to entry into the 90th-percentile high-connectedness state within the next 1, 5, and 20 observations, respectively (Table 8). These low values are model-based conditional probabilities rather than assurances that grain prices or food-security outcomes will remain stable.

Table 8

ThresholdHorizonBest modelForecast originHigh-risk probability
0.90001XGBoost2025-12-180.0041
0.90005XGBoost2025-12-180.0125
0.900020XGBoost2025-12-180.0240

Predicted high-risk probabilities at the end of the sample period.

This result represents only a simulated forecast based on the current sample endpoint. In practical early warning applications, rolling estimation should be conducted after updating the latest grain price and climate risk data.

7 Robustness tests

Robustness is assessed along three dimensions: the TVP-VAR lag order, the forecast-error variance-decomposition horizon, and the rolling quantile used to define warning labels. The benchmark uses one lag, a 100-period decomposition horizon, and 0.99 forgetting factors. Because TVP-VAR-BK does not require a rolling estimation window, lag length and decomposition horizon are the relevant connectedness specifications to vary.

7.1 Robustness of the lag order

We re-estimate the dynamic TCI with lag orders 1, 2, and 3. Figure 11 and Table 9 show correlations of 0.984 and 0.919 between the benchmark and the two longer-lag series. Average TCI rises from 15.29 to 18.69, but the principal peaks and their timing remain. Economically, this means that the level of measured interconnectedness is somewhat lag-sensitive, whereas the identification of high-risk windows is stable.

Figure 11

Table 9

SettingMean TCIMax TCIP80 thresholdCorrelation with baseline
Lag = 115.294357.323022.30481.0000
Lag = 217.009458.098923.49170.9840
Lag = 318.690863.355524.88480.9190

Robustness test of the lag order.

7.2 Robustness of the forecast horizon

The variance-decomposition horizon is then set to 50, 100, and 150 periods. The resulting TCI series overlap almost exactly (Figure 12; Table 10): correlations with the benchmark are effectively one, and means and maxima change only in the second decimal place. The dominance of short-horizon transmission is therefore not an artifact of the 100-period benchmark.

Figure 12

Table 10

SettingMean TCIMax TCIP80 thresholdCorrelation with baseline
Nfore = 5015.308657.374922.29511.0000
Nfore = 10015.294357.323022.30481.0000
Nfore = 15015.289557.306422.30171.0000

Robustness test of the forecast horizon.

7.3 Robustness of the risk thresholds

Finally, the warning models are re-estimated under the rolling 80th-percentile definition. XGBoost remains best by PR-AUC at all three horizons, with values of 0.987, 0.955, and 0.926. The predictive ranking is therefore robust to whether the event represents a general or extreme high-connectedness state.

Across these checks, the timing of connectedness peaks, the dominance of short-band transmission, and XGBoost’s PR-AUC ranking remain stable. The robustness results support the empirical pattern, while the limitations below bound its interpretation.

7.4 Scope and limitations

Four limitations qualify the findings. First, GCPU is a proxy for climate-policy uncertainty and does not measure temperature, precipitation, drought, floods, or crop-yield shocks. Second, the network is limited to six series. It omits rice, Brazilian soybean prices, inventories, exchange rates, freight, tariffs, energy costs, and other exporters; the bilateral US soybean landed cost series is especially sensitive to trade diversion after 2018. Third, generalized connectedness identifies dynamic association and forecast-error variance transmission, not structural causality.

Fourth, the warning target is a sample-relative high-TCI state rather than food insecurity itself. The test block covers 2023–2025, and long-horizon positive events remain limited; calibration accordingly deteriorates at 20 periods. External validation with physical-climate indicators, broader commodity and trade coverage, alternative event definitions, and genuinely real-time vintages is needed before operational deployment.

8 Conclusions and policy recommendations

Using daily data from 2010 to 2025, this study measures dynamic connectedness among selected Chinese and US grain-price changes, US soybean landed cost changes, and GCPU. The empirical object is climate-policy uncertainty and grain-market risk. Broader food-security relevance is inferred through price stability and import-cost exposure rather than measured directly.

First, connectedness is episodic. TCI rises sharply from late 2020 into early 2021 and approaches its sample maximum of 57.32%. This timing coincides with pandemic-related supply-chain disruption, export restrictions, and major weather disturbances, but the connectedness design cannot assign the increase to any one event. The practical implication is conditional: monitoring should pay particular attention when several market and policy stressors overlap.

Second, 73.27% of average connectedness lies in the 1–5-period band. High-frequency price, landed-cost, and policy-uncertainty indicators can therefore complement monthly and quarterly food-system monitoring. They should be used as a market-stability layer, not as substitutes for production, inventory, affordability, and nutrition indicators.

Third, US corn, Chinese wheat, and US wheat are net transmitters on average, while US soybean landed cost, GCPU, and Chinese corn are net receivers. These roles indicate conditional transmission positions, not structural causality. The results support monitoring international benchmark prices and import costs together and considering greater supplier diversification, but the model does not establish that a particular policy action will reduce food insecurity.

Fourth, the warning exercise has meaningful out-of-sample discrimination. XGBoost PR-AUC is 0.986, 0.949, and 0.781 across the three horizons. Its accuracies of 0.987, 0.968, and 0.929 exceed no-event baselines of 0.875, 0.854, and 0.840, respectively, but PR-AUC remains the headline metric because accuracy is inflated by class imbalance. ECE rises from 0.012 to 0.062 as the horizon lengthens, so long-horizon probability signals require additional caution and periodic recalibration.

Policy use should therefore be conservative and layered. A threshold crossing can trigger analyst review, cross-checking against inventories, trade flows, weather, freight, exchange rates, and physical supply indicators before intervention is considered. Rolling thresholds and the model should be re-estimated as new observations arrive. International information sharing can improve interpretation, but the evidence reported here supports these actions as risk-management implications rather than as causal policy prescriptions.

Statements

Data availability statement

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

Author contributions

JY: Writing – original draft, Writing – review & editing.

Funding

The author(s) declared that financial support was not received for this work and/or its publication.

Conflict of interest

The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The author(s) declared that Generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

References

  • 1

    AndersonW.Becker-ReshefI.BodanacN.CarneiroB.EnglishD.HoellA.et al. (2026). Responsible use of artificial intelligence and machine learning for food security early warning systems. Nat. Food7, 736–739. doi: 10.1038/s43016-026-01400-6,

  • 2

    AntonakakisN.ChatziantoniouI.GabauerD. (2020). Refined measures of dynamic connectedness based on time-varying parameter vector autoregressions. J. Risk Financ. Manag.13:84. doi: 10.3390/jrfm13040084

  • 3

    BakerS. R.BloomN.DavisS. J. (2016). Measuring economic policy uncertainty. Q. J. Econ.131, 1593–1636. doi: 10.1093/qje/qjw024

  • 4

    BalashankarA.SubramanianL.FraibergerS. P. (2023). Predicting food crises using news streams. Sci. Adv.9:eabm3449. doi: 10.1126/sciadv.abm3449,

  • 5

    BaruníkJ.KřehlíkT. (2018). Measuring the frequency dynamics of financial connectedness and systemic risk. J. Financ. Econ.16, 271–296. doi: 10.1093/jjfinec/nby001

  • 6

    BehnassiM.El HaibaM. (2022). Implications of the Russia–Ukraine war for global food security. Nat. Hum. Behav.6, 754–755. doi: 10.1038/s41562-022-01391-x,

  • 7

    BouriE.IqbalN.KleinT. (2022). Climate policy uncertainty and the price dynamics of green and brown energy stocks. Financ. Res. Lett.47:102740. doi: 10.1016/j.frl.2022.102740

  • 8

    BreimanL. (2001). Random forests. Mach. Learn.45, 5–32. doi: 10.1023/A:1010933404324

  • 9

    CeballosF.HernandezM. A.MinotN.RoblesM. (2017). Grain price and volatility transmission from international to domestic markets in developing countries. World Dev.94, 305–320. doi: 10.1016/j.worlddev.2017.01.015

  • 10

    ChenT.GuestrinC. (2016). XGBoost: a scalable tree boosting system. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. pp. 785–794

  • 11

    DieboldF. X.YılmazK. (2012). Better to give than to receive: predictive directional measurement of volatility spillovers. Int. J. Forecast.28, 57–66. doi: 10.1016/j.ijforecast.2011.02.006

  • 12

    DieboldF. X.YılmazK. (2014). On the network topology of variance decompositions: measuring the connectedness of financial firms. J. Econom.182, 119–134. doi: 10.1016/j.jeconom.2014.04.012

  • 13

    FalkendalT.OttoC.ScheweJ.JägermeyrJ.KonarM.KummuM.et al. (2021). Grain export restrictions during COVID-19 risk food insecurity in many low- and middle-income countries. Nat. Food2, 11–14. doi: 10.1038/s43016-020-00211-7,

  • 14

    FerriantaY.YantiN. D. (2026). Vulnerability of agricultural food systems to climate and economic shocks: a hybrid machine learning and dynamic panel analysis of Asia-Pacific. Agric. Food Secur.15:52. doi: 10.1186/s40066-026-00639-3

  • 15

    GaleF.ValdesC.AshM. (2019). Interdependence of China, United States, and Brazil in soybean trade. OCS-19F-01. Washington, DC: U.S. Department of Agriculture, Economic Research Service. Available online at: https://ers.usda.gov/publications/93389

  • 16

    GavriilidisK. (2021). Measuring climate policy uncertainty. SSRN [Preprint]. doi: 10.2139/ssrn.3847388

  • 17

    GuoK.LiY.ZhangY.JiQ.ZhaoW. (2023). How are climate risk shocks connected to agricultural markets?J. Commod. Mark.32:100367. doi: 10.1016/j.jcomm.2023.100367

  • 18

    HasegawaT.FujimoriS.HavlíkP.ValinH.BodirskyB. L.DoelmanJ. C.et al. (2018). Risk of increased food insecurity under stringent global climate change mitigation policy. Nat. Clim. Chang.8, 699–703. doi: 10.1038/s41558-018-0230-x

  • 19

    HasegawaT.SakuraiG.FujimoriS.TakahashiK.HijiokaY.MasuiT. (2021). Extreme climate events increase risk of global food insecurity and adaptation needs. Nat. Food2, 587–595. doi: 10.1038/s43016-021-00335-4,

  • 20

    HuX.ZhuB.ZhangB.ZengL. (2024). Extreme risk spillovers between US and Chinese agricultural futures markets in crises: a dependence-switching copula-CoVaR model. PLoS One19:e0299237. doi: 10.1371/journal.pone.0299237,

  • 21

    KahnM. E.MohaddesK.NgR. N. C.PesaranM. H.RaissiM.YangJ.-C. (2021). Long-term macroeconomic effects of climate change: a cross-country analysis. Energy Econ.104:105624. doi: 10.1016/j.eneco.2021.105624

  • 22

    KangS. H.McIverR.YoonS.-M. (2017). Dynamic spillover effects among crude oil, precious metal, and agricultural commodity futures markets. Energy Econ.62, 19–32. doi: 10.1016/j.eneco.2016.12.011

  • 23

    KangS. H.TiwariA. K.AlbulescuC. T.YoonS.-M. (2019). Exploring the time-frequency connectedness and network among crude oil and agriculture commodities. Energy Econ.84:104543. doi: 10.1016/j.eneco.2019.104543

  • 24

    KhalfaouiR.GoodellJ. W.Mefteh-WaliS.ChishtiM. Z.GozgorG. (2024). Impact of climate risk shocks on global food and agricultural markets: a multiscale and tail connectedness analysis. Int. Rev. Financ. Anal.93:103206. doi: 10.1016/j.irfa.2024.103206

  • 25

    KotzM.KuikF.LisE.NickelC. (2024). Global warming and heat extremes to enhance inflationary pressures. Commun. Earth Environ.5:116. doi: 10.1038/s43247-023-01173-x

  • 26

    LeT.-H.PhamL.DoH. X. (2023). Price risk transmissions in the water-energy-food nexus: impacts of climate risks and portfolio implications. Energy Econ.124:106787. doi: 10.1016/j.eneco.2023.106787

  • 27

    LentzE. C.MichelsonH.BaylisK.ZhouY. (2019). A data-driven approach improves food insecurity crisis prediction. World Dev.122, 399–409.

  • 28

    LobellD. B.SchlenkerW.Costa-RobertsJ. (2011). Climate trends and global crop production since 1980. Science333, 616–620. doi: 10.1126/science.1204531,

  • 29

    LuZ.DongJ.LiN.FengH.GouD.XuM. (2026). Land consolidation and smart agriculture synergy for food security: pathways toward agricultural carbon neutrality in China. Agriculture16:287. doi: 10.3390/agriculture16030287

  • 30

    LuG.MiaoY.DongX.DongZ.AnH.LiX. (2025). Strategies for improving food security under the impact of external grain price fluctuations: evidence from China. Appl. Econ.57, 887–903. doi: 10.1080/00036846.2024.2309465

  • 31

    LuoJ.JiQ. (2018). High-frequency volatility connectedness between the US crude oil market and China’s agricultural commodity markets. Energy Econ.76, 424–438. doi: 10.1016/j.eneco.2018.10.031

  • 32

    MartiniG.BracciA.RichesL.JaiswalS.CoreaM.RiversJ.et al. (2022). Machine learning can guide food security efforts when primary data are not available. Nat. Food3, 716–728. doi: 10.1038/s43016-022-00587-8,

  • 33

    MottalebK. A.KrusemanG.SnappS. (2022). Potential impacts of Ukraine-Russia armed conflict on global wheat food security: a quantitative exploration. Glob. Food Secur.35:100659. doi: 10.1016/j.gfs.2022.100659

  • 34

    NazliogluS.ErdemC.SoytasU. (2013). Volatility spillover between oil and agricultural commodity markets. Energy Econ.36, 658–665. doi: 10.1016/j.eneco.2012.11.009

  • 35

    PesaranH. H.ShinY. (1998). Generalized impulse response analysis in linear multivariate models. Econ. Lett.58, 17–29. doi: 10.1016/S0165-1765(97)00214-0

  • 36

    PrimiceriG. E. (2005). Time varying structural vector autoregressions and monetary policy. Rev. Econ. Stud.72, 821–852. doi: 10.1111/j.1467-937X.2005.00353.x

  • 37

    QiC.MengC.YanS. (2025). Transmission effect of international grain prices on China’s grain prices. Front. Sustain. Food Syst.9:1483424. doi: 10.3389/fsufs.2025.1483424

  • 38

    RenX.WangT.LiuZ.XuH. (2025). The risk transmission mechanism between geopolitical risks and the international agricultural product market: an analysis based on the cross-quantilogram and TVP-VAR-BK models. Humanit. Soc. Sci. Commun.12:1799. doi: 10.1057/s41599-025-06072-4

  • 39

    SaâdaouiF.Ben JabeurS.GoodellJ. W. (2022). Causality of geopolitical risk on food prices: considering the Russo–Ukrainian conflict. Financ. Res. Lett.49:103103. doi: 10.1016/j.frl.2022.103103

  • 40

    SchlenkerW.RobertsM. J. (2009). Nonlinear temperature effects indicate severe damages to U.S. crop yields under climate change. Proc. Natl. Acad. Sci.106, 15594–15598. doi: 10.1073/pnas.0906865106,

  • 41

    ShiY.LuW.LinL.LiZ.ChenQ. (2026). How does climate policy uncertainty impact green development of agriculture in China? A perspective from heterogeneous technological innovation. J. Clean. Prod.554:148050. doi: 10.1016/j.jclepro.2026.148050

  • 42

    VoD. H.TranM. P. B. (2024). Volatility spillovers between energy and agriculture markets during the ongoing food & energy crisis: does uncertainty from the Russo-Ukrainian conflict matter?Technol. Forecast. Soc. Change208:123723. doi: 10.1016/j.techfore.2024.123723

  • 43

    WangK.-H.KanJ.-M.QiuL.XuS. (2023a). Climate policy uncertainty, oil price and agricultural commodity: from quantile and time perspective. Econ. Anal. Policy78, 256–272. doi: 10.1016/j.eap.2023.03.013

  • 44

    WangK.-H.WangZ.-S.YunisM.KchouriB. (2023b). Spillovers and connectedness among climate policy uncertainty, energy, green bond and carbon markets: a global perspective. Energy Econ.128:107170. doi: 10.1016/j.eneco.2023.107170

  • 45

    WangY.WuC.YangL. (2014). Oil price shocks and agricultural commodity prices. Energy Econ.44, 22–35. doi: 10.1016/j.eneco.2014.03.016

  • 46

    WesterveldJ. J. L.van den HombergM. J. C.NobreG. G.van den BergD. L. J.TeklesadikA. D.StuitS. M. (2021). Forecasting transitions in the state of food security with machine learning using transferable features. Sci. Total Environ.786:147366. doi: 10.1016/j.scitotenv.2021.147366,

  • 47

    WheelerT.von BraunJ. (2013). Climate change impacts on global food security. Science341, 508–513. doi: 10.1126/science.1239402,

  • 48

    WuY.RenW.WanJ.LiuX. (2023). Time-frequency volatility connectedness between fossil energy and agricultural commodities: comparing the COVID-19 pandemic with the Russia-Ukraine conflict. Financ. Res. Lett.55:103866. doi: 10.1016/j.frl.2023.103866

  • 49

    YanW.-L.CheungA. W. K. (2023). The dynamic spillover effects of climate policy uncertainty and coal price on carbon price: evidence from China. Financ. Res. Lett.53:103400. doi: 10.1016/j.frl.2022.103400

  • 50

    ZengH.AbedinM. Z.AhmedA. D.LuceyB. (2025). Quantile and time–frequency risk spillover between climate policy uncertainty and grains commodity markets. J. Futures Mark.45, 659–682. doi: 10.1002/fut.22583

  • 51

    ZhangZ.LiuJ.WangX.MaoS.ChenL. (2026). Research on spillover effects of climate policy uncertainty on energy and agricultural product markets from a time-frequency perspective. Agriculture16:1019. doi: 10.3390/agriculture16101019

Summary

Keywords

climate policy uncertainty, grain-price risk, machine learning early warning, systemic connectedness, TVP-VAR-BK

Citation

You J (2026) Climate policy uncertainty and domestic–international grain-market risk spillovers: a TVP-VAR-BK analysis with machine learning early warning. Front. Sustain. Food Syst. 10:1949662. doi: 10.3389/fsufs.2026.1949662

Received

27 July 2026

Revised

18 September 2026

Accepted

18 September 2026

Published

14 October 2026

Volume

10 - 2026

Edited by

Zhaoyang Lu, Southwest University of Political Science & Law, China

Reviewed by

Qiang He, Yunnan Normal University, China

Hanjie Wang, Southwest University, China

Updates

Copyright

*Correspondence: Jia You,

Disclaimer

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.

Outline

Figures

Cite article

Copy to clipboard


Export citation file


Share article

Article metrics