From 7b453d302bcd0eb98c5cbe7565b5bb6391b27a30 Mon Sep 17 00:00:00 2001 From: A Samuel Pottinger Date: Tue, 17 Dec 2024 19:48:37 -0800 Subject: [PATCH] Some additional proofreading. --- paper/paper.md | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) diff --git a/paper/paper.md b/paper/paper.md index e233602..a905645 100644 --- a/paper/paper.md +++ b/paper/paper.md @@ -46,12 +46,12 @@ Global warming threatens production of key staple crops, including maize [@rezae Within this context, the United States of America is the world's largest maize producer and exporter [@ates_feed_2023]. Its government-backed Federal Crop Insurance Program covers a large share of this growing risk [@tsiboe_crop_2023]. The costs of crop insurance in the U.S. have already increased by 500% since the early 2000s with annual indemnities reaching $19B in 2022 [@schechinger_crop_2023]. Furthermore, retrospective analysis attributes 19% of "national-level crop insurance losses" between 1991 and 2017 to climate warming, an estimate rising to 47% during the drought-stricken 2012 growing season [@diffenbaugh_historical_2021]. Looking forward, @li_impact_2022 show progressively higher U.S. maize loss rates as warming elevates. ## Prior work -Modeling possible changes in frequency and severity of crop loss events that trigger indemnity claims is an important step to prepare for the future impacts of global warming. Related studies have predicted changes in crop yields at broad scales such as the county-level [@leng_predicting_2020] and have estimated climate change impacts to U.S. maize within whole-sector or whole-economy analysis [@hsiang_estimating_2017]. These efforts include traditional statistical models [@lobell_statistical_2010] as well as an increasing body of work favoring machine learning approaches [@leng_predicting_2020]. Finally, the literature also considers how practice-specific insurance subsidies intersect with grower practices [@connor_crop_2022; @wang_warming_2021; @chemeris_insurance_2022] and observed fresilience [@renwick_long-term_2021; @manski_diversified_2024]. +Modeling possible changes in frequency and severity of crop loss events that trigger indemnity claims is an important step to prepare for the future impacts of global warming. Related studies have predicted changes in crop yields at broad scales such as the county-level [@leng_predicting_2020] and have estimated climate change impacts to U.S. maize within whole-sector or whole-economy analysis [@hsiang_estimating_2017]. These efforts include traditional statistical models [@lobell_statistical_2010] as well as an increasing body of work favoring machine learning approaches [@leng_predicting_2020]. Finally, the literature also consider how practice-specific insurance subsidies intersect with grower practices [@connor_crop_2022; @wang_warming_2021; @chemeris_insurance_2022] and observed resilience [@renwick_long-term_2021; @manski_diversified_2024]. Despite these prior contributions, important programs often include highly localized variables such as an individual farm's last ten years of yield for a specific crop [@rma_crop_2008]. Therefore, to inform policy, research must include more granular models than previous studies [@leng_predicting_2020] and, in addition to predicting yield [@lobell_scalable_2015; @jagermeyr_climate_2021; @ma_qdann_2024], need to simulate insurance instrument mechanics. Of particular interest, we fill a need for climate-aware simulations of loss probability and severity within a "risk" or "insured" unit, a geographic scale referring to a set of agricultural fields that are insured together [@fcic_common_2020]. ## Contribution -We address this need for institutionally-relevant granular future loss prediction through neural network Monte Carlo. We provide these projections at the policy-relevant risk unit scale, probabilistically forecasting institution-relevant outcome metrics under climate change. We focus on the important U.S. Corn Belt, a 9 state region within the United States essential to the nation's maize crop [@green_where_2018]. Within this agriculturally important area, we specifically model the Yield Protection plan, one of the options under the popular Multi-Peril Crop Insurance Program. Furthermore, by contrasting results to a "counterfactual" which does not include further climate warming, we quantitatively highlight the insurer-relevant effects of climate change. Trained on remote sensed maize yield estimations [@lobell_scalable_2015], these models project future insurance outcomes at approximately one and three decades [@williams_high_2024]. +We address this need for institutionally-relevant granular future loss prediction through neural network Monte Carlo. We provide these projections at the policy-relevant risk unit scale, probabilistically forecasting institution-relevant outcome metrics under climate change. We focus on the U.S. Corn Belt, a 9 state region within the United States essential to the nation's maize crop [@green_where_2018]. Within this agriculturally important area, we specifically model the Yield Protection plan, one of the options under the popular Multi-Peril Crop Insurance Program [@rma_statecountycrop_2024]. Furthermore, by contrasting results to a "counterfactual" which does not include further climate warming, we quantitatively highlight the insurer-relevant effects of climate change. Trained on remote sensed maize yield estimations [@lobell_scalable_2015], these models project future insurance outcomes at approximately one and three decades [@williams_high_2024]. # Methods We first build predictive models of maize yield distributions using a neural network at an insurer-relevant spatial scale before simulating changes to yield losses under different climate conditions with Monte Carlo. From these results, we calculate the probability and severity of indemnity claims. @@ -115,7 +115,7 @@ We "grid search" [@joseph_grid_2018] in order to find a suitable combination of We choose our model using each candidate's ability to predict into future years, a task representative of the Monte Carlo simulations [@brownlee_what_2020]: - **Training** on all data between 1999 to 2012 inclusive. -- **Validation** set comprised of 2014 and 2016 to compare candidates from grid search. +- **Validation** on 2014 and 2016 to compare candidates from grid search. - **Test** on 2013 and 2015 which serve as a fully hidden set, estimating how the chosen model may perform in the future. Having performed model selection, we further evaluate our chosen regressor through additional tests which more practically estimate performance in different ways one may consider using this method (see Table @tbl:posthoc). @@ -132,7 +132,7 @@ Table: Overview of trials after model selection. {#tbl:posthoc} These post-hoc trials use only training and test sets as we fully retrain models using unchanging sweep-chosen hyper-parameters as described in Table @tbl:sweepparam. Note that some of these tests use "regions" which we define as all geohashes sharing the same first three characters, creating a grid of 109 x 156 km cells [@haugen_geohash_2020] each including all neighborhoods (4 character geohashes) found within. ## Simulation -As described in Figure @fig:pipeline, neural network predictions of future yield delta distributions feed into Monte Carlo simulations [@metropolis_beginning_1987; @kwiatkowski_monte_2022] which estimate probabilities and severity of losses at the risk unit scale. This operation happens for 17 individual years sampled separately from both the 2030 and 2050 CHC-CMIP6 series [@williams_high_2024]. +As described in Figure @fig:pipeline, neural network predictions of future yield delta distributions feed into Monte Carlo simulations [@metropolis_beginning_1987; @kwiatkowski_monte_2022] which estimate probabilities and severity of losses at the risk unit scale. Though not predicting specific individual future years^[CHC-CMIP6 predicts years around 2030 and 2050 as conditions are co-correlated within a year. However, the product does not, for example, predict 2035 specifically.], this operation happens for each of the 17 years found within the 2030 and 2050 CHC-CMIP6 series [@williams_high_2024]. ![Model pipeline overview diagram. Code released as open source.](./img/pipeline.png "Model pipeline overview diagram. Code released as open source."){ width=80% #fig:pipeline } @@ -183,9 +183,9 @@ The claims rate elevates in the 2030 series and doubles in the 2050 timeframe wh We observe a number of policy-relevant dynamics when simulating insurance instrument mechanics under climate change. ## Yield expectations -Figure @fig:hist reveals possible challenges with using a simple average in crop insurance products. In current instruments, $y_{expected}$ captures changes to risk but simulations anticipate that higher yield volatility skews yield delta distributions such that simulated risk units see higher claims rates despite changes to their yield average. +Figure @fig:hist reveals possible challenges with using a simple average in crop insurance products. Current instruments expect $y_{expected}$ to capture changes to risk but simulations anticipate that higher yield volatility skews yield delta distributions such that simulated risk units see higher claims rates despite changes to their yield average. -![Interactive tool screenshot showing 2050 outcomes distribution as changes from $y_{expected}$, showing deltas and claims rates with climate change on the top and without further climate change (counterfactual) on bottom.](./img/hist.png "Interactive tool screenshot showing 2050 outcomes distribution. This graphic depicts changes from $y_{expected}$, showing deltas and claims rates with further climate change on the top and without climate change (counterfactual) on bottom."){#fig:hist} +![Interactive tool screenshot showing 2050 outcomes distribution as changes from $y_{expected}$, plotting deltas and claims rates with climate change on the top and without further climate change (counterfactual) on bottom.](./img/hist.png "Interactive tool screenshot showing 2050 outcomes distribution as changes from $y_{expected}$, plotting deltas and claims rates with climate change on the top and without further climate change (counterfactual) on bottom."){#fig:hist} Indeed, as further described in supplemental Table S5, 12.7% of neighborhoods and 9.8% of counties under SSP245 in the 2050 series report both increased claims rates and increased average yields. In other words, yield volatility could allow a sharp elevation in loss probability without necessarily decreasing overall mean yields substantially enough to reduce claims rates through $y_{expected}$. These results highlight a need for future research into alternative FCIP policy formulations, such as using historic yield variance when establishing production histories and $y_{expected}$.