Research ·
001 · Basis risk in parametric crop insurance: pre-registered evaluation
A pre-registration. The question, data, baselines, splits, metrics and decision thresholds below were fixed before any model was fitted. The result is reported in a follow-up note in January 2027, whichever direction it takes.
Question
Index insurance pays when an observable index I crosses a trigger, not when the insured incurs a loss L. The gap between P(I) and L is basis risk. The standard indices, cumulative rainfall and area yield, were designed before daily reanalysis weather, 250 m vegetation indices and cheap compute were available at scale.
The claim under test: indices built from public weather, vegetation and soil data, with index design treated as an optimisation problem, reduce county-level basis risk for drought losses in US corn and soybeans by a margin that survives held-out years, relative to the best index that can be computed by the payout date.
Anchors in the literature: measured basis risk of index-based livestock insurance in northern Kenya [Jensen, Barrett and Mude, 2016]; remote sensing and crop modelling for agricultural risk management, a review [Benami et al., 2021]; satellite-retrieved soil-moisture index insurance for wheat, maize and rapeseed in eastern Germany [Vroege et al., 2021]. Closest prior work: US county-yield index insurance evaluated against satellite-estimated yields of about 1.8 million fields [Stigler and Lobell, 2020]; the optimal zone index as the first principal component of yields, with its eigenvalue as an upper bound on the risk any index can insure [Stigler and Lobell, 2024]; a neural-network index for Illinois corn compared with linear indices on the same variables [Chen et al., 2024]. The note tests whether their reported basis-risk levels can be beaten on public US data under a fixed protocol, not whether their numbers reproduce.
Setup
Unit and scope. County × year. Corn and soybeans. Iowa, Illinois, Indiana, Ohio, Minnesota, Missouri, South Dakota. Seasons 2000–2025. Drought as the only loss cause. Payout dates: 31 August (corn), 15 September (soybeans). An index is admissible only if every input is available by the payout date.
Target. Drought-cause loss cost from USDA RMA Cause of Loss, L = indemnity / liability per county-year, with the coverage mix from the Summary of Business as a control. County yield shortfall from NASS is a secondary target, modelled separately, because insured loss reflects coverage levels, enrolment and price effects as well as agronomy.
Inputs.
| Source | Role |
|---|---|
| USDA RMA Summary of Business, Cause of Loss | Target L; coverage-mix control |
| USDA NASS Quick Stats | Secondary target; area-yield index |
| USDA Cropland Data Layer | Crop masks, year t−1 only |
| ERA5 / ERA5-Land | Weather features; weather-index baselines |
| MODIS MOD13Q1 | NDVI/EVI over the season |
| gSSURGO | Static soil covariates |
Baselines. No insurance. Naive rainfall floor. Tuned weather index (cumulative precipitation and extreme degree-days [Schlenker and Roberts, 2009], parameters fitted in-fold). NDVI/EVI area index. Area-yield index, reported as a reference bound because it is not payout-date-feasible. The bar for any proposed index is the best payout-date-feasible baseline.
Matched premium. For every index and every fold, the trigger and payout scale are calibrated on training years so that the expected payout E[P(I)] equals a common value. Comparisons are made only between indices with equal expected payout, so a reduction in false negatives bought with more false positives cannot register as a gain.
Splits. Outer: leave-one-year-out over 2000–2025, binding for every reported number, because in-sample gains of weather indices have reversed out of sample [Vedenov and Barnett, 2004] and random splits overstate crop-yield skill [Sweet et al., 2023]. Inner: folds within the training years for all design choices (features, detrending, trigger form). Spatial blocking is reported as a secondary check [Roberts et al., 2017]. Detrending and normalisation are fitted inside the fold.
Index families. Headline results use contractable families: linear and piecewise-linear functions of a small number of named features, with a single trigger and cap. Black-box models are evaluated and reported only as an upper bound on attainable basis-risk reduction, not as a product.
Measurement
For index payout P and loss L on held-out county-years, with residual R = L − P:
- False negative rate FNR = P(P = 0 | L > ℓ) and false positive rate FPR = P(P > 0 | L ≤ ℓ) at a fixed loss threshold ℓ.
- Hedging effectiveness HE = 1 − Var(R) / Var(L).
- Residual downside semivariance SV⁻(R) = E[min(R − E[R], 0)²], the primary metric.
- Certainty-equivalent gain under CRRA utility, u(w) = w^(1−γ) / (1−γ), at fair and at loaded premium, for γ ∈ {1, 2, 3}.
Uncertainty by year-block bootstrap: seasons are resampled, counties within a season are not, since the dominant dependence is the common weather shock. Intervals for the selected index are deflated for selection: the choice of the best index family is repeated inside every bootstrap replicate, so the interval reflects having searched over K families rather than having specified one in advance [cf. Bailey and López de Prado, 2014; Bailey et al., 2017]. The exact procedure is in the methodology document released with the result note.
Decision thresholds, fixed 19 October 2026. At matched premium, relative to the best payout-date-feasible baseline, residual downside semivariance lower by at least 15% (corn) and 5% (soybeans); within 10 percentage points of the area-yield reference; deflated interval excluding zero; holding with 2012 removed; holding on the 2009–2025 subperiod.
Sanity checks. 2012 must be detected by any admissible index. A model trained on shuffled labels must score at baseline. Every reported gain is re-run on the held-out years by hand before it is written down.
Leakage audit. Crop masks from year t−1 only; detrending and normalisation fitted in-fold; data vintages recorded per source; every feature checked against the payout-date cut-off.
What this can and cannot show
Public US data has no farm-level losses. Every number here is county-aggregate basis risk, and hedging effectiveness measured on aggregated data overstates what individual farms obtain [Heimfarth, Finger and Musshoff, 2012]. The farm-level question, where basis risk is largest [Jensen, Barrett and Mude, 2016; Stigler and Lobell, 2024] and the product case is strongest, requires portfolio data and is out of scope for this note. Insured losses are not pure agronomy; the coverage-mix control and the separate yield model address this partially. The effective sample is about 25 seasons, and the tail metrics rest on a handful of drought years, hence the year-block inference and the drop-2012 check.
Result
Pending. Review in late December 2026. Outcomes as fixed in advance: thresholds met → calibration on partner data and a paper; strong result without a partner → preprint; gain within noise of the baselines, or visible only with leakage → kill, with a post-mortem published.
References
- Bailey and López de Prado (2014). The deflated Sharpe ratio. Journal of Portfolio Management 40(5): 94–107. SSRN 2460551
- Bailey, Borwein, López de Prado and Zhu (2017). The probability of backtest overfitting. Journal of Computational Finance 20(4): 39–69. doi:10.21314/JCF.2016.322
- Benami, Jin, Carter, Ghosh, Hijmans, Hobbs, Kenduiywo and Lobell (2021). Uniting remote sensing, crop modelling and economics for agricultural risk management. Nature Reviews Earth & Environment 2: 140–159. doi:10.1038/s43017-020-00122-y
- Chen, Lu, Zhang and Zhu (2024). Managing weather risk with a neural network-based index insurance. Management Science 70(7): 4306–4327. doi:10.1287/mnsc.2023.4902
- Heimfarth, Finger and Musshoff (2012). Hedging weather risk on aggregated and individual farm-level. Agricultural Finance Review 72(3): 471–487. doi:10.1108/00021461211277295
- Jensen, Barrett and Mude (2016). Index insurance quality and basis risk: evidence from northern Kenya. American Journal of Agricultural Economics 98(5): 1450–1469. doi:10.1093/ajae/aaw046
- Roberts et al. (2017). Cross-validation strategies for data with temporal, spatial, hierarchical, or phylogenetic structure. Ecography 40(8): 913–929. doi:10.1111/ecog.02881
- Schlenker and Roberts (2009). Nonlinear temperature effects indicate severe damages to U.S. crop yields under climate change. PNAS 106(37): 15594–15598. doi:10.1073/pnas.0906865106
- Stigler and Lobell (2020). On the benefits of index insurance in US agriculture: a large-scale analysis using satellite data. arXiv:2011.12544
- Stigler and Lobell (2024). Optimal index insurance and basis risk decomposition: an application to Kenya. American Journal of Agricultural Economics 106(1): 306–329. doi:10.1111/ajae.12375
- Sweet, Müller, Anand and Zscheischler (2023). Cross-validation strategy impacts the performance and interpretation of machine learning models. Artificial Intelligence for the Earth Systems 2(4): e230026. doi:10.1175/AIES-D-23-0026.1
- Vedenov and Barnett (2004). Efficiency of weather derivatives as primary crop insurance instruments. Journal of Agricultural and Resource Economics 29(3): 387–403. doi:10.22004/ag.econ.30916
- Vroege, Bucheli, Dalhaus, Hirschi and Finger (2021). Insuring crops from space: the potential of satellite-retrieved soil moisture to reduce farmers’ drought risk exposure. European Review of Agricultural Economics 48(2): 266–314. doi:10.1093/erae/jbab010