Stakeholder priorities in model selection for managing basis risk in soybean index insurance in Paraná Brazil

Technical Summary: Stakeholder Priorities in Model Selection for Managing Basis Risk in Soybean Index Insurance in Paraná, Brazil

Problem Statement
Index insurance is a scalable mechanism for transferring weather risk in rain-fed agricultural systems, yet its efficacy is frequently undermined by basis risk—the divergence between index-triggered payouts and actual farm-level losses. While methodological advances have expanded yield-weather modeling (e.g., machine learning, hybrid designs), empirical research in South America remains limited. Furthermore, existing model comparisons often prioritize conventional predictive accuracy (e.g., RMSE) over the specific insurance consequences of prediction errors. These errors manifest asymmetrically: negative basis risk (uncompensated losses) erodes farmer protection, while positive basis risk (unnecessary payouts) increases insurer costs. The selection of a yield model is often treated as a purely statistical exercise, ignoring that the "optimal" model depends on the relative policy weight assigned to protecting farmers versus limiting insurer exposure. This study addresses these gaps by investigating how model selection for soybean index insurance in Paraná, Brazil, varies across agroclimatic clusters, maturity windows, and stakeholder priorities.

Methodology
The study employs a stakeholder-weighted framework to evaluate four yield-modeling specifications using municipal soybean yield data (2008–2022) and gridded extreme climate indices (0.1° resolution) for 399 municipalities in Paraná.

1. Data Preprocessing:

   • Yield: Municipal yields were detrended using a robust state-level trend model to isolate weather-driven variation from technological progress.
   • Climate Indices: A four-stage screening process (seasonal aggregation, variance/missingness screening, redundancy screening, and Lasso selection) reduced 48 monthly indices to a parsimonious set of predictors for early (Oct–Dec) and late (Oct–Jan) maturity windows.
   • Spatial Grouping: Municipalities were grouped into five exogenous agroclimatic clusters.

2. Model Specifications:

   • Panel Fixed-Effects Regression (PFER): Captures time-invariant heterogeneity (e.g., soil, topography) and minimizes symmetric squared error.
   • Quantile Regression (QR): Targets the lower tail of the yield distribution (baseline $\tau=0.3$) to minimize underprediction, using an asymmetric loss function.
   • Lasso Quantile Regression (LQR): Applies an $L_1$ penalty to QR to prevent overfitting and select variables.
   • Weighted Model Aggregation (WMA): An ensemble forecast combining the three individual models based on their out-of-sample RMSE.

3. Evaluation Framework:

   • Out-of-Sample Testing: An expanding window approach (training on $t-1$, testing on $t$) was used for years 2017–2022.
   • Basis Risk Decomposition: Total basis risk was decomposed into False Negative (FN) basis risk (payout < actual loss) and **False Positive (FP)** basis risk (payout > actual loss).
   • Stakeholder-Weighted Loss Function: A policy weight $\lambda \in [0, 1]$ was applied to FN risk, where $\lambda=1$ represents a farmer-oriented priority (minimizing uncompensated losses), $\lambda=0$ represents an insurer-oriented priority (minimizing unnecessary payouts), and $\lambda=0.5$ represents an equal weighting.

Key Results

• Climate Drivers: Extreme heat intensity (days >35°C, heat waves) and persistence were the most consistent suppressors of yield, with effects more pronounced in the late maturity window. Other significant drivers included long dry spells, low solar radiation/evapotranspiration, and high humidity persistence. Conversely, wider diurnal temperature ranges and cooler-than-average days supported yield gains.
• Model Performance & Trade-offs:
  • No Universal Superiority: No single model minimized total basis risk across all clusters and scenarios.
  • PFER Dominance: Under insurer-oriented ($\lambda=0$) and equally-weighted ($\lambda=0.5$) scenarios, PFER consistently minimized weighted basis risk. It effectively limited positive basis risk (unnecessary payouts) but resulted in higher negative basis risk (uncompensated losses) in several clusters.
  • Quantile Dominance: Under farmer-oriented scenarios ($\lambda=1$), QR and LQR models were preferred. These models significantly reduced negative basis risk (by 13–47% compared to PFER in various clusters) but at the cost of increased positive basis risk.
  • Spatial Heterogeneity: Cluster 2 exhibited low risk across all scenarios, while Cluster 3 showed a stark trade-off: excellent at avoiding excess payouts under PFER but poor at covering actual losses, necessitating tail-sensitive models for farmer protection.
• Ensembling: Weighted Model Aggregation (WMA) did not consistently outperform individual models, offering only marginal and inconsistent gains in specific clusters and quantile settings.
• Sensitivity: The baseline quantile $\tau=0.3$ provided the most stable balance. Lowering $\tau$ to 0.2 further reduced negative basis risk but drastically increased positive basis risk; raising it to 0.4 had the opposite effect.

Significance and Contributions
The paper makes three primary contributions to the literature on index insurance and agricultural risk management:

1. Empirical Extension: It provides evidence on soybean index insurance design in Paraná, a major but underexamined South American region, distinguishing between early and late maturity windows and utilizing a comprehensive set of extreme climate indicators.
2. Shift in Evaluation Criteria: The study moves beyond conventional predictive accuracy metrics to evaluate models based on their insurance consequences. It demonstrates that a model with the lowest prediction error may not be the most appropriate for insurance design if it fails to align with the specific risk transfer objectives (e.g., protecting farmers vs. controlling costs).
3. Policy-Dependent Model Selection: By introducing a transparent stakeholder-weighted loss function, the study formalizes the trade-off between negative and positive basis risk. It concludes that model selection is not a static statistical problem but a policy-dependent decision. The "best" model changes systematically based on whether the priority is minimizing uncompensated losses (favoring quantile models) or limiting unnecessary payouts (favoring panel fixed-effects models).

Limitations and Scope
The authors note that the findings are specific to rain-fed soybeans in Paraná and rely on historical climate stationarity, which may not hold under future climate regimes. The study focuses on basis risk decomposition and model selection rather than premium pricing, demand estimation, or welfare outcomes. Additionally, the limited gains from model ensembling suggest that combining correlated models (estimated on the same variable set) may not yield significant benefits in this context. The framework is presented as a transferable approach for evaluating model trade-offs rather than a prescriptive solution for all regions.