Organizations: Image Processing Laboratory, Universitat de València, València, Spain. · Artificial Intelligence Group, Wageningen University & Research, Wageningen, The Netherlands. · Joint Research Centre, European Commission, Ispra, Italy.
Climate variability influences whether a market disruption escalates into a food crisis, yet broad climate patterns like El Niño, tracked months before they alter hydro-climatic conditions, are still not incorporated as an early-warning component in food-security responses. We address this gap by introducing sensitivity regimes, a stratification of regions by the direction and strength of their vegetation response to the El Niño Southern Oscillation, and using them to estimate how food price spikes affect acute food insecurity across sub-Saharan Africa. Integrating remote sensing, socioeconomic data, and causal machine learning, we find that in regions where ENSO systematically suppresses vegetation, a price spike raises the share of the population at acute risk by 5.4 percentage points in the following month. In regions where vegetation is unaffected by or positively linked to ENSO, the estimated effect is smaller (around 2 percentage points) and statistically insignificant. These results demonstrate that climate context is critical for understanding food security vulnerabilities. Sensitivity regimes can be combined with operational price-spike triggers to stage anticipatory action: the ENSO state flags vulnerable regions months ahead, and a pre-positioned response in those regions to a price spike would avert the largest jump in acute food insecurity.
Figures & tables
Figure 1 : Proposed analytical pipeline. The workflow begins by extracting the dominant ENSO-linked vegetation signal, then stratifies regions by sensitivity regimes, and concludes with the estimation of causal effects averaged within each regime. The time-series plots show examples from Kenya, Mali, and Zimbabwe.
Figure 3 : Geographical distribution of sensitivity regimes based on GPCA stratification. (Left) First PC of the GPCA analysis on ENSO–NDVI, providing a continuous map of the causal influence. (Right) Regions classified as Positive, Negative, or Weak based on the strength and direction of ENSO–NDVI causal linkages.
Model
Regime
CATE (p.p.)
C.I.
p-value
Negative
6.00
(2.57, 8.73)
0.020
Linear DML
Weak
1.94
(-0.47, 4.99)
0.098
Positive
2.16
(-3.92, 4.21)
0.784
Negative
5.35
(2.16, 8.08)
0.020
Causal Forest DML
Weak
1.80
(-0.47, 4.75)
0.137
Positive
1.70
(-3.70, 4.15)
0.922
Table 1 : Conditional Average Treatment Effect (CATE) estimates for ALPS impacts on rCSI, stratified by sensitivity regimes. Confidence intervals are reported at the 95% level.
Figure 4 : Baseline risk of acute food insecurity under normal food price conditions and added risk when prices spike. Estimations provided by the Causal Forest DML CATE estimates stratified for the sensitivity regimes.
Figure 5 : Feature importance of the nuisance models. Estimations of the treatment (top) and outcome (bottom) for the Causal Forest DML estimator stratified for the sensitivity regimes.
βpfromNDVIp(t)=αp+βpΦ1(t)+εp,t,
Algorithm 1 GPCA for sensitivity Regime Classification
Appendix figures & tables17 assets
Supplementary material from the paper’s appendix.
Appendix
Domain
Variable
Description
Source
Outcome
rCSI
Reduced Coping Strategy Index crisis prevalence, monthly mean.
WFP
Treatment
ALPS
Alert for Price Spikes averaged for staple food commodities.
WFP Prices
Mean Temperature (Crop)
Spatial mean of the monthly average temperature over crop-covered areas.
ASAP
Mean FPAR (Crop)
Spatial mean of the monthly average fraction of absorbed photosynthetically active radiation (FPAR) over crop-covered areas.
ASAP
Appendix
Table 2 : Outcome, treatment, and confounders derived from the HFID and used in the causal analysis. All variables are aggregated at the ADM1-month level. Confounders are grouped into climate, economic, and social domains.
From
To
Justification and references
ENSO
NDVI, Crop Drivers
ENSO alters regional rainfall and temperature patterns, which drive vegetation greenness and define crop growing conditions (Anyamba et al., 2018; Ropelewski and Halpert, 1987; FAO et al., 2023).
ENSO
GDP
Climate shocks associated with ENSO influence aggregate output and growth, particularly in agriculture-dependent economies (Dell et al., 2012; Burke et al., 2015).
NDVI
Crop Drivers
Vegetation condition indices are closely tied to local biophysical crop conditions such as canopy development and water stress (Anyamba et al., 2018; FAO et al., 2023).
Crop Drivers
GDP
Weather- and climate-sensitive crop conditions affect yields and agricultural value added, which is a key component of GDP in many low- and middle-income countries (Dell et al., 2012; Burke et al., 2015; FAO et al., 2023).
Crop Drivers
Conflicts
Agricultural prod. failures can exacerbate income loss, rural hardship, and competition over resources, recognized as structural and proximate drivers of conflict and instability (Burke et al., 2015; FSIN and Global Network Against Food Crises, 2024).
Crop Drivers
IDPs
Agroclimatic shocks contribute to livelihood collapse and displacement, especially in agriculture-dependent settings (FSIN and Global Network Against Food Crises, 2024; IDMC, 2024).
Appendix
Table 3 : Justification of the causal links in the DAG (Part I). For each directed edge we provide a short rationale and example supporting references.
From
To
Justification and references
GDP, Food Prices
rCSI
Household income and employment prospects modulated by GDP growth influence the need to adopt negative coping strategies (FAO et al., 2023; World Bank, 2020).
Conflicts
IDPs
Violent conflict is the leading global driver of internal displacement (IDMC, 2024; FSIN and Global Network Against Food Crises, 2024).
Conflicts
ALPS
Conflict disrupts markets, transport, and production, constraining supply and raising transaction costs (FAO et al., 2023; FSIN and Global Network Against Food Crises, 2024).
Conflicts
rCSI
Conflicts reduce incomes, destroy assets and constrain humanitarian access, all of which raise the prevalence of acute food insecurity and severe coping captured by rCSI (FSIN and Global Network Against Food Crises, 2024; FAO et al., 2023).
IDPs
ALPS
Inflows of displaced populations can increase local demand, strain markets and infrastructure, and contribute to localised food price spikes (FSIN and Global Network Against Food Crises, 2024; IDMC, 2024).
IDPs
rCSI
IDPs face greater exposure to shocks, have constrained access to livelihoods and services, and are consistently among the most food-insecure groups (FSIN and Global Network Against Food Crises, 2024; IDMC, 2024).
Appendix
Table 4 : Justification of the causal links in the DAG (Part II). For each directed edge we provide a short rationale and example supporting references.
Strong-pixel q
Min. Dominance
Min. Coverage
Unchanged
κ
Pos ↔ Neg
#Neg
#Weak/#Pos
0.20
0.20
0.10
1.000
1.000
0
158
204/203
0.20
0.20
0.20
1.000
1.000
0
158
204/203
0.10
0.15
0.10
0.993
0.989
0
159
200/206
0.10
0.15
0.20
0.993
0.989
0
159
200/206
Appendix
Table 5 : Sensitivity of ADM1 regime labels to threshold choices (study countries). Agreement metrics compare each configuration to the baseline labeling. Pos ↔ Neg counts direct Positive ↔ Negative sign reversals.
Metric
Value
Notes
# ADM1 units
222
Study countries only
Baseline counts (Neg/Weak/Pos)
71 / 64 / 87
Mean stability
0.941
Pr(Xr=Xrbase)
Share stability ≥0.9
0.794
Share stability =1.0
0.622
Mean Pos ↔ Neg switch prob.
1.39×10−4
Median = 0, 75th pct = 0
Appendix
Table 6 : Subsampling stability summary (study countries). Stability is the probability of retaining the baseline label when pixels are subsampled. The Pos ↔ Neg switch probability measures direct sign reversals.
GID_0
Country
ADM1
Baseline
Stability
P(Neg)
P(Weak)
P(Pos)
MOZ
Mozambique
Maputo
Negative
0.575
0.575
0.425
0.000
NGA
Nigeria
Sokoto
Weak
0.585
0.000
0.585
0.415
KEN
Kenya
Kericho
Negative
0.585
0.585
0.380
0.035
NGA
Nigeria
Nasarawa
Weak
0.590
0.410
0.590
0.000
NGA
Nigeria
Rivers
Positive
0.600
0.005
0.395
0.600
KEN
Kenya
Turkana
Negative
0.615
0.615
0.385
0.000
Appendix
Table 7 : Lowest-stability ADM1 units in the subsampling test (study countries). Columns show baseline label, stability, and empirical label probabilities under subsampling.
Figure 6 : Kernel density estimates of the propensity scores. The density distributions illustrate the propensity score distributions for the treated and control groups. The vertical dashed lines indicate the lower and upper cutoffs for the propensity-score trimming.
Regime
Before trimming
After trimming
Control
Treated
Control
Treated
Overall
829
845
483
455
Negative
318
238
227
130
Weak
271
378
158
220
Positive
240
229
98
105
Appendix
Table 8 : Sample sizes before and after propensity-score trimming, overall and by sensitivity regime. We report the number of control ( T=0 ) and treated ( T=1 ) observations before trimming and after restricting to the common-support region 0.2≤e^(Wi)≤0.8 .
Confounder
∣ SMD ∣ (Before)
∣ SMD ∣ (After)
Conflict Stock Displacement
0.080
0.240
Mean Temperature (Crop)
0.115
0.118
Mean Soil Moisture (Crop)
0.202
0.064
Mean FPAR (Crop)
0.232
0.156
ACLED Fatalities
0.199
0.182
Rural Population
0.035
0.085
Appendix
Table 9 : Standardized mean differences (SMD) for confounders before and after propensity-score trimming (Overall sample).
Confounder
∣ SMD ∣ (Before)
∣ SMD ∣ (After)
Conflict Stock Displacement
0.441
0.274
Mean Temperature (Crop)
0.137
0.118
Mean Soil Moisture (Crop)
0.408
0.333
Mean FPAR (Crop)
0.406
0.325
ACLED Fatalities
0.302
0.165
Rural Population
0.183
0.102
Appendix
Table 10 : Standardized mean differences (SMD) for confounders before and after propensity-score trimming (Negative regime).
Confounder
∣ SMD ∣ (Before)
∣ SMD ∣ (After)
Conflict Stock Displacement
0.518
0.309
Mean Temperature (Crop)
0.177
0.061
Mean Soil Moisture (Crop)
0.123
0.336
Mean FPAR (Crop)
0.220
0.065
ACLED Fatalities
0.291
0.299
Rural Population
0.034
0.088
Appendix
Table 11 : Standardized mean differences (SMD) for confounders before and after propensity-score trimming (Weak regime).
Confounder
∣ SMD ∣ (Before)
∣ SMD ∣ (After)
Conflict Stock Displacement
0.248
0.089
Mean Temperature (Crop)
0.212
0.336
Mean Soil Moisture (Crop)
0.028
0.003
Mean FPAR (Crop)
0.213
0.018
ACLED Fatalities
0.086
0.083
Rural Population
0.071
0.049
Appendix
Table 12 : Standardized mean differences (SMD) for confounders before and after propensity-score trimming (Positive regime).
Model family
CV R2 for ( Y∣X,W ) (Regressor)
CV ROC–AUC for ( T∣X,W ) (Classifier)
Random Forest
0.214
0.613
Gradient Boosting
0.394
0.633
XGBoost
0.352
0.637
In-sample (selected models)
0.786
0.982
Appendix
Table 13 : Nuisance model selection and fit for the CATE experiments. Best cross-validated R2 and ROC–AUC for each candidate model family when predicting the outcome Y∣X,W and treatment T∣X,W . The final nuisance models used in the CATE estimators are highlighted in bold; the corresponding in-sample metrics are reported in the last row.
Model
Regime
CATE (p.p.)
C.I.
p-value
Negative
6.065
(3.257, 8.124)
0.020
Diff. in Means
Weak
-0.040
(-3.596, 3.114)
0.922
Positive
0.423
(-4.807, 5.344)
0.765
Negative
4.135
(2.051, 5.633)
0.020
Linear Regression
Weak
1.084
(-0.881, 4.204)
0.157
Positive
3.312
(-1.014, 5.671)
0.216
Appendix
Table 14 : Full CATE estimates for ALPS impacts on rCSI, stratified by sensitivity regimes. Baseline and DML models. Confidence intervals at 95%, and estimations are statistically significant if p -values ≤ 0.05.
Method
Regime
Placebo eff.
Placebo p
RCC eff.
RCC p
RSR eff.
RSR p
Negative
0.008
0.385
0.000
0.885
0.000
0.938
LinearDML
Weak
-0.004
0.807
0.002
0.808
0.002
0.500
Positive
0.012
0.615
-0.005
0.769
0.002
0.688
Negative
0.010
0.269
0.002
0.769
0.001
0.750
CausalForestDML
Weak
-0.004
0.731
0.002
0.731
0.002
0.500
Positive
0.012
0.615
-0.004
0.769
0.003
0.688
Appendix
Table 15 : Robustness diagnostics for regime-specific CATEs (DML estimators). For each estimator and sensitivity regime, we report the Placebo effect, the RCC mean absolute change, the RSR effect, and their respective p -values. Tests succeed if p -values ≥ 0.05.
Method
Lead ℓ
N (T/C)
Negative
Weak
Positive
0
823 (423/400)
0.0514 (0.0255, 0.0795) p=0.0196
0.0206 (-0.0030, 0.0558) p=0.1176
-0.0070 (-0.0475, 0.0317) p=0.6471
1
938 (455/483)
0.0599 (0.0257, 0.0873) p=0.0196
0.0194 (-0.0047, 0.0499) p=0.0980
0.0216 (-0.0392, 0.0421) p=0.7843
LinearDML
2
845 (430/415)
0.0550 (0.0328, 0.0814) p=0.0196
0.0024 (-0.0179, 0.0472) p=0.4314
-0.0472 (-0.0772, 0.0049) p=0.1373
3
840 (419/421)
0.0675 (0.0294, 0.0894) p=0.0196
0.0151 (-0.0112, 0.0370) p=0.2745
-0.0496 (-0.0908, -0.0056) p=0.0392
4
777 (414/363)
0.0638 (0.0412, 0.0923) p=0.0196
0.0193 (-0.0145, 0.0679) p=0.2941
-0.0738 (-0.1054, -0.0209) p=0.0196
0
823 (423/400)
0.0524 (0.0256, 0.0805) p=0.0196
0.0199 (-0.0060, 0.0510) p=0.1373
-0.0125 (-0.0465, 0.0398) p=0.6078
Appendix
Table 16 : Lead–lag robustness: regime-specific CATEs under outcome leads ℓ∈{0,…,4} . Rows report regime-specific CATEs for the Negative/Weak/Positive sensitivity regimes. Each cell shows the point estimate, the 95% confidence interval, and the p-value. N reports the post-trimming sample size (treated/control) under the common-support restriction 0.2≤e^(Wi)≤0.8 .
Method
Contrast
Δ (p.p.)
C.I.
p -value
Linear DML
Neg – Pos
3.844
(0.909, 11.738)
0.039
Neg – Weak
4.061
( −0.083 , 6.809)
0.098
Causal Forest DML
Neg – Pos
3.650
( −0.129 , 10.678)
0.078
Neg – Weak
3.549
( −0.446 , 6.572)
0.137
Appendix
Table 17 : Regime-contrast tests. Bootstrap distributions of pairwise differences between regime-specific CATEs for the DML estimators. Confidence intervals at 95%, and estimations are statistically significant if p -values ≤ 0.05.
The intensification of climate change poses a growing threat to food security, especially in vulnerable communities. This study employs an observational machine-learning framework to estimate the causal association between access to credit and acute food insecurity in Somalia and across the Horn of Africa, drawing on a harmonized dataset spanning key environmental, socioeconomic, and conflict-related factors from 2015 to 2022. Results indicate that greater credit access is associated with a 2% reduction in acute food insecurity at the population level over the study period. Given that, on average, 16% of the population is in crisis, this effect represents a meaningful shift within the at-risk group. We interpret these estimates under explicit identification assumptions and complement them with robustness and refutation tests. The results provide context-specific evidence on how financial access correlates with food security outcomes in data-scarce, crisis-affected settings, and offer a transparent framework for integrating heterogeneous data sources when randomized evaluations are infeasible.
Jordi Cerdà-Bautista, Vasileios Sitokonstantinou, José Manuel Veiga López-Peña +3
Vegetable prices in Sri Lanka are highly volatile because the market is largely import-isolated, so supply disruptions quickly drive prices up. This study develops a machine learning framework to forecast such volatility by incorporating supply-chain-aware features and explicitly modelling the country's two cultivation seasons, Maha (October-April) and Yala (May-September). An integrated dataset was constructed by combining retail and farmer-gate prices with origin-aligned weather variables, diesel costs, and exchange rates across 12 vegetable varieties and 14 market centres from 2013 to 2019. A gradient-boosted ensemble model (XGBoost and LightGBM) was trained and optimised using Optuna, and unified and season-specific configurations were compared. Results show that season-specific models improve within-season fit, with the Yala-specific model achieving the highest R2 of 0.9420 (95% CI [0.690, 1.000]), while the unified model delivers the best overall predictive accuracy of 90.84% (95% CI [88.34%, 91.52%]) and an R2 of 0.9281 (95% CI [0.760, 1.000]). Notably, the unified model maintains 85.96% accuracy on a completely unseen 2024 hyperinflationary period without retraining, successfully tracking major price surges. These findings suggest that agricultural price movements in import-constrained markets are meaningfully predictable when models capture supply-chain dynamics, offering practical value for early warning and decision making by farmers, traders, and policymakers. Existing studies on Sri Lankan vegetable prices are confined to Autoregressive Integrated Moving Average (ARIMA) and Generalized Autoregressive Conditional Heteroskedasticity (GARCH) applied to single markets, with no supply-chain features, seasonal segmentation, or cross-regime validation.
Forecasting agricultural commodity prices in emerging economies is difficult due to high volatility, frequent supply disruptions, and strong cultural influences on demand. This study introduces the Kalimati Vegetable Price Index (KVPI), a new inverse-volatility weighted composite index that aggregates 135 daily wholesale commodities from Kathmandu over ten years (2013-2023). By creating a stable macro-level signal, the KVPI reduces the noise inherent in modelling individual crops. A rich set of 64 causally valid features was developed, including festival lead-lag effects, rolling statistics, and calendar variables. Fourteen forecasting models spanning statistical, tree-based, deep learning, hybrid, and transformer architectures were rigorously evaluated across short (7-day), medium (14- and 30-day), and long-term (90-day) horizons. Tree-based ensembles proved notably robust, while classical statistical models and complex transformers struggled with the noisy dataset. The proposed Momentum-Corrected Online Stacking Ensemble achieved the strongest performance, yielding a Root Mean Square Error (RMSE) of 1.771, an exceptionally low Mean Absolute Percentage Error (MAPE) of 0.68%, and explaining 84.5% of the variance (R-squared = 0.845) at the 90-day horizon. This open-source pipeline provides policymakers and supply chain actors in Nepal and similar markets with a practical, reliable tool for anticipating price movements and strengthening food security.