- Crop insurance, corn, and climate change: This website serves as an interactive toolset where users can explore what the future looks like in the US Corn Belt and how policy may influence those outcomes. This complements our scientific paper (preprint) as interactive supplemental material.
+ Crop insurance, corn, and climate change: This website serves as an interactive exploration of what the future might look like in the US Corn Belt and how policy may influence those outcomes. This complements our scientific paper (preprint) as interactive supplemental material.
Introduction: Corn is in
Here's an outline of what is available on this website:
Table of contents
@@ -101,7 +101,7 @@ Introduction: Corn is in
Here's an outline of what is available on this website:
- - Introduction: This is the tab you are currently on which introduces this problem at a high level and hints at what the outcomes of the experiment might be.
+ - Introduction: This is the tab you are currently on which introduces this problem at a high level.
- Rates: Exploration of how rates are set within crop insurance and highlighting possible policy levers.
- ML & Hyperparameters: Discussion of machine learning hyperparameters considered with additional information on model performance.
- Predicted Distributions: Examination of Monte Carlo outcomes with predicted distributions under different conditions.
@@ -123,7 +123,7 @@ Introduction: Corn is in
Climate change threatens the future of corn. Studies expect planet-wide maize yields to drop by up to 24% by the end of this century (JΓ€germeyr et al. 2021). Why? Climate change alters growing conditions (Williams et al 2024), including a higher frequency and severity of stressful weather (Dai 2013) which can impact yields (Sinsawat et al 2004; Marouf et al 2013; Lobell et al 2020). Of course, these threats can endanger farmer revenue (Sajid et al 2023). However, these changes can also challenge the institutions tasked with protecting and helping those growers (Hanrahan 2024).
Introduction: Corn is in Climate change threatens the future of corn. Studies expect planet-wide maize yields to drop by up to 24% by the end of this century (JΓ€germeyr et al. 2021). Why? Climate change alters growing conditions (Williams et al 2024), including a higher frequency and severity of stressful weather (Dai 2013) which can impact yields (Sinsawat et al 2004; Marouf et al 2013; Lobell et al 2020). Of course, these threats can endanger farmer revenue (Sajid et al 2023). However, these changes can also challenge the institutions tasked with protecting and helping those growers (Hanrahan 2024).
- How might we use AI to understand this future and prepare for it? These explorables (Victor 2011) detail key aspects of our analysis of future maize yields from the perspective of one of these important institutions: the US Federal Crop Insurance Program. Climate change increasingly threatens this crucial resource for farmers and the broader US food system (Ray et al 2015). Specifically, this website houses tools which complement an academic paper (preprint), allowing for further exploration beyond the manuscript.
+ How might we use AI to understand this future and prepare for it? These explorables (Victor 2011) detail key aspects of our analysis of future maize yields. Our work specifically considers the future risk for one of these important institutions: the US Federal Crop Insurance Program. Climate change increasingly threatens this crucial resource for farmers and the broader US food system (Ray et al 2015). Specifically, this website houses tools which complement an academic paper (preprint), allowing for further exploration beyond the manuscript.
@@ -143,7 +143,7 @@
Introduction: Corn is in
- Let's get started. Each section of this website dives into one aspect of analysis. Each is a "mini-lab" in which you can operate analysis tools as described in our scientific paper (preprint).
+ Let's get started. Each section of this website dives into one aspect of analysis through a "mini-lab" in which you can operate analysis tools as described in our scientific paper (preprint).
@@ -193,7 +193,7 @@ Introduction: Corn is in
Introduction: Corn is in
Rates Explorer
- This first interactive tool explores the structure of Multi-Peril Crop Insurance's Yield Protection option. We specifically will walk through the rates setting aspects of MPCI YP in which a mixture of many variables determines the price a producer pays for insurance (Plastina and Edwards 2020). First, the grower chooses how much of their APH they want to insure (horizontal axis in the interactive). However, certain county-level data, the APH for a insured unit, and a series of adjustments also play a role. By combining everything together, this sets a price and a subsidy from the government (vertical axis in the interactive). See details about variables used for rate setting.
+ This first interactive tool explores the structure of Multi-Peril Crop Insurance's Yield Protection option. We specifically will walk through the rates setting aspects of MPCI YP in which a mixture of many variables determines the price a producer pays for insurance (Plastina and Edwards 2020). First, the grower chooses how much of their APH they want to insure (horizontal axis in the interactive). However, certain county-level data, the APH for a risk unit, and a series of adjustments also play a role. By combining everything together, this sets a price and a subsidy from the government (vertical axis in the interactive). See details about variables used for rate setting.
Details about variables used for rate setting
@@ -203,7 +203,7 @@ Rates Explorer
-
- Insured Unit APH: APH for the set of fields in the insured unit for which a policy is being created.
+ Insured Unit APH: APH for the set of fields in the risk unit for which a policy is being created.
-
County average yield: The average yield (APH equivalent) for all fields in the county.
@@ -218,7 +218,7 @@
Rates Explorer
Subsidy rate: What percent of the overall policy cost is covered by subsidy. A value of 55% means that the grower pays for 45% of the overall cost of the policy.
-
- Coverage amount: Up to what percent of APH will be covered. If one sees a yield that is only 75% of APH then the policy will pay out 75% - 70% or 5% of APH.
+ Coverage amount: Up to what percent of APH will be covered. If one sees a yield that is only 75% of APH then the policy will pay out 5% (75% minus 70%) of APH.
@@ -281,7 +281,7 @@ Rates Explorer
- Note that the horizontal axis in the interactive representing desired coverage level is currently defined as a percent of APH. Other interactives explore what happens if this includes an understanding of yield volatility. More specifically, standard deviation (Hargrave 2024) gets smaller as a grower's yields become more stable, offering a way to quantify stability. Including std in APH-based coverage levels could happen by 508h or by changing the law.
+ Note that the horizontal axis in the interactive representing desired coverage level is currently defined as a percent of APH. Other interactives explore what happens if this includes an understanding of yield volatility. For example, standard deviation (Hargrave 2024) gets smaller as a grower's yields become more stable, offering a simple way to quantify stability. Including std or similar formulations in APH-based coverage levels could happen by 508h or by changing the law.
With that in mind, It's worth mentioning that some of these structures may be set by US law (CFR and USC). For example:
@@ -290,10 +290,10 @@ Rates Explorer
the level of coverage ... may be purchased at any level not to exceed 85 percent of the [expected] individual yield ... the yield for a crop shall be based on the actual production history for the crop
- Text specifics can dramatically change the outlook for both crop insurance and the farmers it serves. For example, our paper explores what might happen if that definition were to change to this:
+ Text specifics can dramatically change the outlook for both crop insurance and the farmers. For example, later explorables on this site consider what might happen if that definition were to change to this:
- the level of coverage ... may be purchased at any level not to exceed 2.1 standard deviations below the [expected] individual yield ... the yield for a crop shall be based on the actual production history for the crop
+ the level of coverage ... may be purchased at any level not to exceed 1.5 standard deviations below the [expected] individual yield ... the yield for a crop shall be based on the actual production history for the crop
- Hyper-Parameters Explorer
+ Hyper-Parameters Explorer
Our neural network learns from past growing conditions and their associated outcomes (Williams et al 2024; Lobell et al 2015) in order to predict future yields given climate projections (Williams et al 2024).
@@ -335,7 +335,7 @@ Hyper-Parameters Explorer<
More details about inputs, outputs, and climate variables
- The inputs to the neural network include climate variables, year, and baseline yield (mean and std of all historic yields for the cell). The output is mean and std of changes to that future yield for the grid cell. Yield data come from SCYM (Lobell et al 2015). Historic growing conditions and future climate projections come from CHC-CMIP6 (Williams et al 2024). The climate variables are as follows (Williams et al 2024) where all are daily:
+ The inputs to the neural network include climate variables, year, and baseline yield (mean and std of all historic yields for the cell). The output is mean and std of changes to that future yield for the grid cell. Note that our scientific paper (preprint) considers other non-normal distribution shapes. Anyway, yield data come from SCYM (Lobell et al 2015). Historic growing conditions and future climate projections come from CHC-CMIP6 (Williams et al 2024). The climate variables are as follows (Williams et al 2024) where all are daily:
- rhn: overall relative humidity
@@ -348,13 +348,13 @@ Hyper-Parameters Explorer<
- wbgt: wet bulb globe temperature
- These daily climate values are summarized to min, max, mean, and std per month per grid cell before going to the neural network. We use the dataset variable names for consistency.
+ These daily climate values are summarized to min, max, mean, and std of daily values per month per grid cell before going to the neural network. We use the dataset variable names for consistency.
- First, we divide up the US Corn Belt into a geographic grid, breaking up the problem so it becomes tractable for modeling. Next, our network goes year by year, forecasting the yields for each cell ("neighborhood") in that grid. As yield may vary within even those small areas, the network describes the range of expected yields as a mean and standard deviation per cell.
+ First, we divide up the US Corn Belt into a geographic grid, breaking up the problem so it becomes tractable for modeling. Next, our network goes year by year, forecasting the yields for each cell ("neighborhood") in that grid. Within these small areas, the network describes the range of expected changes to yield as a mean and standard deviation per cell. This is further discussed in our scientific paper (preprint).
More details about the grid
@@ -366,7 +366,7 @@ Hyper-Parameters Explorer<
- This interactive lets us try out different hyper-parameters: depths, L2s, dropouts, and blocks. Clicking the "Try Model" button will test a new model where we split the historic data such that the network trains from one piece (training set) before we grade it on the part it has not seen before (validation set). We score performance with mean absolute error (Acharya 2021). For example, if the model predicts that a neighborhood's yield will drop by 10% and it actually drops by 15%, the MAE would be 5%. Our goal: go left on the horizontal axis (minimize error predicting mean) and to the bottom on the vertical axis (minimize error predicting std).
+ This interactive lets us try out different hyper-parameters: depths, L2s, dropouts, and blocks. Clicking the "Try Model" button will test a new model where we split the historic data such that the network trains from one piece (training set) before we grade it on the part it has not seen before (validation set). We score performance with mean absolute error (Acharya 2021). For example, if the model predicts that a neighborhood's yield will drop by 10% and it actually drops by 15%, the error would be 5%. Our goal: go left on the horizontal axis (minimize error predicting mean) and to the bottom on the vertical axis (minimize error predicting std).
Details of hyper-parameters
@@ -464,7 +464,7 @@ Hyper-Parameters Explorer<
- The errors in the visualization reflect a smaller dataset than the one used to train the final model. After retraining with more information, the model sees an MAE of 5% for mean prediction and 2% for std prediction.
+ The errors in the visualization reflect a smaller dataset than the one used to train the final model. After retraining with more information, the model sees an MAE of 6% for mean prediction and 2% for std prediction.
More about model performance and uncertainty
@@ -478,23 +478,7 @@ Hyper-Parameters Explorer<
Only the training set teaches the model while we try to optimize parameters like L2, dropout, etc.
- That said, after determining our model preferred model configuration, we can have the model learn from the combined validation and training sets before we make a final estimation of error we expect in the future using the test set. In this final trial, we see an MAE of 5% for mean prediction and 2% for std prediction. We can later use an understanding of this model uncertainty in Monte Carlo.
-
-
- Finally, despite all of this work, we still need to determine if our model sees similar error across different parts of the US Corn Belt and types of yield changes.
-
-
- 
-
-
- Though not technically a test set, follow up visualizations suggest that:
-
-
- - Error remains generally stable across geography. The southwest corner has smaller sample size so slightly eleveated error makes sense. These are also areas where results are not typically significant.
- - The network appears reluctant to predict highly negative changes to yield. That said, these dramatic drops are overall less common and the error across the entire dataset is symmetric (+3% slightly too optimistic) so this likely does not impact system-wide findings. Furthermore, sampling error during Monte Carlo may address this concern.
-
-
- All things considered, this model offers sufficient performance for our Monte Carlo.
+ That said, after determining our model preferred model configuration, we can have the model learn from the combined validation and training sets before we make a final estimation of error we expect in the future using the test set. In this final trial, we see an MAE of 6% for mean prediction and 2% for std prediction. We can later use an understanding of this model uncertainty in Monte Carlo.
@@ -553,7 +537,7 @@ Hyper-Parameters Explorer<
- Predicted Distributions Explorer
+ Predicted Distributions Explorer
We made our neural network and Monte Carlo next lets us explore the results. This simulation tool examines what crop insurance outcomes might look like in the future given what our model learned. This lets us calculate loss frequency and severity, revealing how often we expect claims in the future. While running these trials, we can also incorporate neural network uncertainty.
@@ -562,7 +546,7 @@ Predicted Distributions E
Details about the simulations
- In our Monte Carlo, we sample the neural network's outputs. However, to account for uncertainty caused by our model, we also draw from the measurements of errors we've seen from our neural network (Yanai 2010). Finally, we also account for the size of insured units (RMA 2024) since policies cover collections of fields and not geohashes or individual fields. Altogether, this paints a picture of changes to yield relative to APH at the 2030 and 2050 timeframes. For more information, see our scientific paper (preprint).
+ In our Monte Carlo, we sample the neural network's outputs. However, to account for uncertainty caused by our model, we also draw from the measurements of errors we've seen from our neural network (Yanai 2010). Finally, we also account for the size of risk units (RMA 2024) since policies cover collections of fields and not geohashes or individual fields. Altogether, this paints a picture of changes to yield relative to APH at the 2030 and 2050 timeframes. For more information, see our scientific paper (preprint).
@@ -612,13 +596,10 @@ Predicted Distributions E
-
- Time: In general things get worse in terms of losses as climate change progresses during the 2030 and then 2050 series.
-
- -
- Coverage levels: The picture may look different for policies insuring 75% of yield versus 85% of yield due to the way that the edges of the distribution or "tails" are changing. Those insuring at higher levels see some very high claims rates under climate change.
+ Time: In general things get worse in terms of losses as climate change progresses during the 2030 and then 2050 series as warming progresses.
-
- Reference point: Things look different depending on what we compare against. Contrasting the continued climate change scenario (SSP245) to "historic" values in the recent past sees average yields stay about the same but the risk of loss increases. On the other hand, comparing to the "counterfactual" (a simulated future where climate change stops today) has our neural network actually expecting yields to increase relative to SSP245.
+ Coverage levels: The picture may look different for policies insuring 75% of yield versus 85% of yield. Those insuring at higher levels see some very high claims rates under climate change.
-
Unit size: The effects of climate change become less pronounced when simulating with the full insured unit where there may be more fields in a single policy compared to sub-unit which simulates at a smaller scale (simulates 5 character geohashes and near-individual field samples). This is due to the "portfolio effect" where larger units offer more chances for high performing fields to offset fields with worse outcomes (Knight et al 2024).
@@ -637,7 +618,7 @@
Predicted Distributions E
- Taken altogether, these results reveal a very specific threat to corn yields and crop insurance. Generally yields improve over time (Nielsen 2023) but climate change in our simulation wipes out most of the increases our neural network would otherwise expect if climate change stopped. This means that the average yield does not necessarily change dramatically from the historic APH but the ends of the distribution of yield outcomes or "tails" evolve. This represents less yield stability in the future such that the APH does not necessarily capture a risk increase even as the rate of claims for which YP pays out more than doubles in the 2050 series. This specific pairing of unchanged average yields but higher volatility is at the center of the risk posed to crop insurance.
+ Taken altogether, these results reveal a very specific threat to corn yields and crop insurance. In particular, the ends of the distribution of yield outcomes or "tails" evolve. This represents less yield stability in the future such that the APH does not necessarily capture all increasing risk even as the rate of claims for which YP pays out more than doubles in the 2050 series. This specific pairing of unchanged average yields but higher volatility is at the center of the risk posed to crop insurance.
More about simulation resolution
@@ -646,8 +627,8 @@ Predicted Distributions E
The resolution of these data could also influence results. In geohashing we currently use the 4 character size corresponding to 28 by 20 kilometers (Haugen 2020). We could also try:
- - Making the grid cells larger: Currently the data are finding significant results in 91% of grid cells so increasing the sample at the cost of predictive resolution offers very limited benefit.
- - Making the grid cells smaller: We attempt to approximate this in the sub-unit size results with simulation of 5 character geohashes. That said, this lowers the frequency with which a grid cell sees significant results by more than half though this number can vary from simulation to simulation. This makes sense as the 5 character level nears the resolution of the underlying data so increasing resolution introduces unacceptable noise.
+ - Making the grid cells larger: Currently the data are finding significant results in most grid cells so increasing the sample at the cost of predictive resolution offers very limited benefit.
+ - Making the grid cells smaller: We attempt to approximate this in the sub-unit size results with simulation of 5 character geohashes. That said, this lowers the frequency with which a grid cell sees significant results substantially though this number can vary from simulation to simulation. This makes sense as the 5 character level nears the resolution of the underlying data so increasing resolution introduces unacceptable noise.
Altogether, we continue to suggest use of the 4 character geohash.
@@ -693,7 +674,7 @@ Predicted Distributions E
- Claims Visualization
+ Claims Visualization
We have a number of options to confront these increases in risk. For example, regenerative practices can reduce risks through diverse crop rotations (Bowles et al 2020) and improvements to soil health (Renwick et al 2021). This could provide output stability in addition to other environmental benefits (Hunt et al 2020).
@@ -718,10 +699,10 @@ Claims Visualization
- If we have the tools, what is the hold up? Recalling the rates explorer, the average-based APH may play a part in the delay. Despite valuable resilience offered by regenerative agriculture (given that our results see higher claims not through overall reduced averages but higher volatility), these important practices may not always improve mean yields or can even come at the cost of a slightly reduced average (Deines et al 2023). This may be part of why, even though they guard against elevations in the probability of loss events (Renwick et al 2021), crop insurance may discourage these resilience-building steps (Wang et al 2021; Chemeris et al 2022).
+ If we have the tools, what is the hold up? Recalling the rates explorer, the average-based APH may play a part in the delay. Despite valuable resilience offered by regenerative agriculture, these important practices may not always improve mean yields or can even come at the cost of a slightly reduced average (Deines et al 2023). This may be part of why, even though they guard against elevations in the probability of loss events (Renwick et al 2021), crop insurance may discourage these resilience-building steps (Wang et al 2021; Chemeris et al 2022).
- Instead of insuring up to a percentage below average yield, what if instead one insures up to a certain number of standard deviations below average? This next interactive explores this question by looking at how different risk units might respond to this kind of structural shift.
+ Alternative formulations require more reserach. However, for the sake of exploration, consider what if one insures up to a certain number of standard deviations below average instead of insuring up to a percentage below average yield. This next interactive explores this question by looking at how different risk units might respond to this kind of structural shift.
@@ -768,7 +749,7 @@ Claims Visualization
More about the policy
- Using recent data, the equivalent to insuring up to 25% below average using present-day risk is insuring up to 2.11 standard deviations below average though this does already have some small effects absent behavior change as units not typically making claims pull down that equivalency or, in other words, it already shifts exposure towards more stable yields. Here is how this would get formulated:
+ Using recent data, the equivalent to insuring up to 25% below average using present-day risk is insuring up to about 1.5 standard deviations below average though this does already have some small effects absent behavior change as units not typically making claims pull down that equivalency or, in other words, it already shifts exposure towards more stable yields. Here is how this would get formulated:
- Average-based (current): loss = max(coverage percentage * average yield - actual yield, 0)
@@ -812,7 +793,7 @@ Claims Visualization
Data downloads for discussion
- The data downloads for this discussion section ("how might we address these concerns") are as follows:
+ The data downloads section are as follows:
- export_claims.csv: The claims rate expected under different scenarios. Here threshold refers to coverage level (YP is 0.25 meaning 25% below average), thresholdStd refers to the translation of an average-based APH to std, and offsetBaseline refers to if this includes APH changes over time (we recommend using "always").
@@ -824,16 +805,16 @@ Claims Visualization
- Neighborhood Explorer
+ Neighborhood Explorer
- So far, we have explored data aggregated across many neighborhoods. This final graphic looks at those neighborhoods individually. Note that this refers to statistically significant results to cut through noise but, if desired, one can turn off that filter.
+ So far, we have explored data aggregated across many neighborhoods. This final graphic maps those neighborhoods individually. Note that this refers to statistically significant results to cut through noise but, if desired, one can turn off that filter.
More about the statistical tests
- Around 91% of maize-growing acerage examined in this study see statistically significant results (p < 0.05 / n) in at least one year of the 2050 series. We determine this by a Bonferroni-corrected (Bonferroni 1935; McDonald 2014) Mann Whitney U (Mann and Whitney 1947; McDonald 2014) per neighborhood per year as variance may differ between the two expected and counterfactual sets (McDonald 2014).
+ Around 95% of maize-growing acerage examined in this study see statistically significant results (p < 0.05 / n) in at least one year of the 2050 series. We determine this by a Bonferroni-corrected (Bonferroni 1935; McDonald 2014) Mann Whitney U (Mann and Whitney 1947; McDonald 2014) per neighborhood per year as variance may differ between the two expected and counterfactual sets (McDonald 2014).
@@ -890,7 +871,7 @@ Neighborhood Explorer
-
- Preciptiation may offer a protective benefit: an increased chirps value has a weak but significant rank correlation (Spearman 1904) to lower changes in risk. This would potentially agree with prior work (Sinsawat et al 2004; Marouf et al 2013).
+ Preciptiation may offer a protective benefit: an increased chirps value may be associated with less loss elevation. This would potentially agree with prior work (Sinsawat et al 2004; Marouf et al 2013).
-
Different geographies may see different outcomes with increased risk potentially concentrated in a band stretching through Iowa, Illinois, and Indiana.
@@ -900,10 +881,7 @@
Neighborhood Explorer
- Of course, the beauty of an explorable explanation is that you can dig beyond our original intent as you look through each of the variables that this model uses. That said, additional insights can be found in our scientific paper (preprint) including more details about the neural network itself and its performance in different tasks.
-
-
- With that, thank you for joining us on this journey through modeling the future of corn in the US!
+ In addition to digging in beyond our original findings as you look through each of the variables that this model uses, additional insights can be found in our scientific paper (preprint) including more details about the neural network itself and its performance in different tasks.
- This first interactive tool explores the structure of Multi-Peril Crop Insurance's Yield Protection option. We specifically will walk through the rates setting aspects of MPCI YP in which a mixture of many variables determines the price a producer pays for insurance (Plastina and Edwards 2020). First, the grower chooses how much of their APH they want to insure (horizontal axis in the interactive). However, certain county-level data, the APH for a insured unit, and a series of adjustments also play a role. By combining everything together, this sets a price and a subsidy from the government (vertical axis in the interactive). See details about variables used for rate setting.
+ This first interactive tool explores the structure of Multi-Peril Crop Insurance's Yield Protection option. We specifically will walk through the rates setting aspects of MPCI YP in which a mixture of many variables determines the price a producer pays for insurance (Plastina and Edwards 2020). First, the grower chooses how much of their APH they want to insure (horizontal axis in the interactive). However, certain county-level data, the APH for a risk unit, and a series of adjustments also play a role. By combining everything together, this sets a price and a subsidy from the government (vertical axis in the interactive). See details about variables used for rate setting.
Details about variables used for rate setting
@@ -203,7 +203,7 @@Rates Explorer
- - Insured Unit APH: APH for the set of fields in the insured unit for which a policy is being created. + Insured Unit APH: APH for the set of fields in the risk unit for which a policy is being created.
-
County average yield: The average yield (APH equivalent) for all fields in the county.
@@ -218,7 +218,7 @@
Rates Explorer
Subsidy rate: What percent of the overall policy cost is covered by subsidy. A value of 55% means that the grower pays for 45% of the overall cost of the policy. - - Coverage amount: Up to what percent of APH will be covered. If one sees a yield that is only 75% of APH then the policy will pay out 75% - 70% or 5% of APH. + Coverage amount: Up to what percent of APH will be covered. If one sees a yield that is only 75% of APH then the policy will pay out 5% (75% minus 70%) of APH.
Rates Explorer
- Note that the horizontal axis in the interactive representing desired coverage level is currently defined as a percent of APH. Other interactives explore what happens if this includes an understanding of yield volatility. More specifically, standard deviation (Hargrave 2024) gets smaller as a grower's yields become more stable, offering a way to quantify stability. Including std in APH-based coverage levels could happen by 508h or by changing the law.
+ Note that the horizontal axis in the interactive representing desired coverage level is currently defined as a percent of APH. Other interactives explore what happens if this includes an understanding of yield volatility. For example, standard deviation (Hargrave 2024) gets smaller as a grower's yields become more stable, offering a simple way to quantify stability. Including std or similar formulations in APH-based coverage levels could happen by 508h or by changing the law.
With that in mind, It's worth mentioning that some of these structures may be set by US law (CFR and USC). For example:
@@ -290,10 +290,10 @@
Rates Explorer
the level of coverage ... may be purchased at any level not to exceed 85 percent of the [expected] individual yield ... the yield for a crop shall be based on the actual production history for the crop
- Text specifics can dramatically change the outlook for both crop insurance and the farmers it serves. For example, our paper explores what might happen if that definition were to change to this:
+ Text specifics can dramatically change the outlook for both crop insurance and the farmers. For example, later explorables on this site consider what might happen if that definition were to change to this:
- the level of coverage ... may be purchased at any level not to exceed 2.1 standard deviations below the [expected] individual yield ... the yield for a crop shall be based on the actual production history for the crop
+ the level of coverage ... may be purchased at any level not to exceed 1.5 standard deviations below the [expected] individual yield ... the yield for a crop shall be based on the actual production history for the crop
Hyper-Parameters Explorer
+Hyper-Parameters Explorer
Our neural network learns from past growing conditions and their associated outcomes (Williams et al 2024; Lobell et al 2015) in order to predict future yields given climate projections (Williams et al 2024).
@@ -335,7 +335,7 @@ Hyper-Parameters Explorer<
Hyper-Parameters Explorer<
More details about inputs, outputs, and climate variables
- The inputs to the neural network include climate variables, year, and baseline yield (mean and std of all historic yields for the cell). The output is mean and std of changes to that future yield for the grid cell. Yield data come from SCYM (Lobell et al 2015). Historic growing conditions and future climate projections come from CHC-CMIP6 (Williams et al 2024). The climate variables are as follows (Williams et al 2024) where all are daily:
+ The inputs to the neural network include climate variables, year, and baseline yield (mean and std of all historic yields for the cell). The output is mean and std of changes to that future yield for the grid cell. Note that our scientific paper (preprint) considers other non-normal distribution shapes. Anyway, yield data come from SCYM (Lobell et al 2015). Historic growing conditions and future climate projections come from CHC-CMIP6 (Williams et al 2024). The climate variables are as follows (Williams et al 2024) where all are daily:
- rhn: overall relative humidity
@@ -348,13 +348,13 @@ Hyper-Parameters Explorer<
- wbgt: wet bulb globe temperature
- These daily climate values are summarized to min, max, mean, and std per month per grid cell before going to the neural network. We use the dataset variable names for consistency.
+ These daily climate values are summarized to min, max, mean, and std of daily values per month per grid cell before going to the neural network. We use the dataset variable names for consistency.
- The inputs to the neural network include climate variables, year, and baseline yield (mean and std of all historic yields for the cell). The output is mean and std of changes to that future yield for the grid cell. Yield data come from SCYM (Lobell et al 2015). Historic growing conditions and future climate projections come from CHC-CMIP6 (Williams et al 2024). The climate variables are as follows (Williams et al 2024) where all are daily:
+ The inputs to the neural network include climate variables, year, and baseline yield (mean and std of all historic yields for the cell). The output is mean and std of changes to that future yield for the grid cell. Note that our scientific paper (preprint) considers other non-normal distribution shapes. Anyway, yield data come from SCYM (Lobell et al 2015). Historic growing conditions and future climate projections come from CHC-CMIP6 (Williams et al 2024). The climate variables are as follows (Williams et al 2024) where all are daily:
- rhn: overall relative humidity @@ -348,13 +348,13 @@
- wbgt: wet bulb globe temperature
Hyper-Parameters Explorer<
- First, we divide up the US Corn Belt into a geographic grid, breaking up the problem so it becomes tractable for modeling. Next, our network goes year by year, forecasting the yields for each cell ("neighborhood") in that grid. As yield may vary within even those small areas, the network describes the range of expected yields as a mean and standard deviation per cell.
+ First, we divide up the US Corn Belt into a geographic grid, breaking up the problem so it becomes tractable for modeling. Next, our network goes year by year, forecasting the yields for each cell ("neighborhood") in that grid. Within these small areas, the network describes the range of expected changes to yield as a mean and standard deviation per cell. This is further discussed in our scientific paper (preprint).
More details about the grid
@@ -366,7 +366,7 @@Hyper-Parameters Explorer<
- This interactive lets us try out different hyper-parameters: depths, L2s, dropouts, and blocks. Clicking the "Try Model" button will test a new model where we split the historic data such that the network trains from one piece (training set) before we grade it on the part it has not seen before (validation set). We score performance with mean absolute error (Acharya 2021). For example, if the model predicts that a neighborhood's yield will drop by 10% and it actually drops by 15%, the MAE would be 5%. Our goal: go left on the horizontal axis (minimize error predicting mean) and to the bottom on the vertical axis (minimize error predicting std).
+ This interactive lets us try out different hyper-parameters: depths, L2s, dropouts, and blocks. Clicking the "Try Model" button will test a new model where we split the historic data such that the network trains from one piece (training set) before we grade it on the part it has not seen before (validation set). We score performance with mean absolute error (Acharya 2021). For example, if the model predicts that a neighborhood's yield will drop by 10% and it actually drops by 15%, the error would be 5%. Our goal: go left on the horizontal axis (minimize error predicting mean) and to the bottom on the vertical axis (minimize error predicting std).
Details of hyper-parameters
@@ -464,7 +464,7 @@Hyper-Parameters Explorer<
- The errors in the visualization reflect a smaller dataset than the one used to train the final model. After retraining with more information, the model sees an MAE of 5% for mean prediction and 2% for std prediction.
+ The errors in the visualization reflect a smaller dataset than the one used to train the final model. After retraining with more information, the model sees an MAE of 6% for mean prediction and 2% for std prediction.
Only the training set teaches the model while we try to optimize parameters like L2, dropout, etc.
- That said, after determining our model preferred model configuration, we can have the model learn from the combined validation and training sets before we make a final estimation of error we expect in the future using the test set. In this final trial, we see an MAE of 5% for mean prediction and 2% for std prediction. We can later use an understanding of this model uncertainty in Monte Carlo.
-
- More about model performance and uncertainty
@@ -478,23 +478,7 @@Hyper-Parameters Explorer<
- Finally, despite all of this work, we still need to determine if our model sees similar error across different parts of the US Corn Belt and types of yield changes.
-
-
-
-
-
-
- Though not technically a test set, follow up visualizations suggest that:
-
-
-
-
- Error remains generally stable across geography. The southwest corner has smaller sample size so slightly eleveated error makes sense. These are also areas where results are not typically significant. -
- The network appears reluctant to predict highly negative changes to yield. That said, these dramatic drops are overall less common and the error across the entire dataset is symmetric (+3% slightly too optimistic) so this likely does not impact system-wide findings. Furthermore, sampling error during Monte Carlo may address this concern. -
Hyper-Parameters Explorer<
Predicted Distributions Explorer
+Predicted Distributions Explorer
We made our neural network and Monte Carlo next lets us explore the results. This simulation tool examines what crop insurance outcomes might look like in the future given what our model learned. This lets us calculate loss frequency and severity, revealing how often we expect claims in the future. While running these trials, we can also incorporate neural network uncertainty.
@@ -562,7 +546,7 @@ Predicted Distributions E
Predicted Distributions E
Details about the simulations
- In our Monte Carlo, we sample the neural network's outputs. However, to account for uncertainty caused by our model, we also draw from the measurements of errors we've seen from our neural network (Yanai 2010). Finally, we also account for the size of insured units (RMA 2024) since policies cover collections of fields and not geohashes or individual fields. Altogether, this paints a picture of changes to yield relative to APH at the 2030 and 2050 timeframes. For more information, see our scientific paper (preprint).
+ In our Monte Carlo, we sample the neural network's outputs. However, to account for uncertainty caused by our model, we also draw from the measurements of errors we've seen from our neural network (Yanai 2010). Finally, we also account for the size of risk units (RMA 2024) since policies cover collections of fields and not geohashes or individual fields. Altogether, this paints a picture of changes to yield relative to APH at the 2030 and 2050 timeframes. For more information, see our scientific paper (preprint).
@@ -612,13 +596,10 @@ Predicted Distributions E
-
- Time: In general things get worse in terms of losses as climate change progresses during the 2030 and then 2050 series.
-
- -
- Coverage levels: The picture may look different for policies insuring 75% of yield versus 85% of yield due to the way that the edges of the distribution or "tails" are changing. Those insuring at higher levels see some very high claims rates under climate change.
+ Time: In general things get worse in terms of losses as climate change progresses during the 2030 and then 2050 series as warming progresses.
-
- Reference point: Things look different depending on what we compare against. Contrasting the continued climate change scenario (SSP245) to "historic" values in the recent past sees average yields stay about the same but the risk of loss increases. On the other hand, comparing to the "counterfactual" (a simulated future where climate change stops today) has our neural network actually expecting yields to increase relative to SSP245.
+ Coverage levels: The picture may look different for policies insuring 75% of yield versus 85% of yield. Those insuring at higher levels see some very high claims rates under climate change.
-
Unit size: The effects of climate change become less pronounced when simulating with the full insured unit where there may be more fields in a single policy compared to sub-unit which simulates at a smaller scale (simulates 5 character geohashes and near-individual field samples). This is due to the "portfolio effect" where larger units offer more chances for high performing fields to offset fields with worse outcomes (Knight et al 2024).
@@ -637,7 +618,7 @@
Predicted Distributions E
- In our Monte Carlo, we sample the neural network's outputs. However, to account for uncertainty caused by our model, we also draw from the measurements of errors we've seen from our neural network (Yanai 2010). Finally, we also account for the size of insured units (RMA 2024) since policies cover collections of fields and not geohashes or individual fields. Altogether, this paints a picture of changes to yield relative to APH at the 2030 and 2050 timeframes. For more information, see our scientific paper (preprint).
+ In our Monte Carlo, we sample the neural network's outputs. However, to account for uncertainty caused by our model, we also draw from the measurements of errors we've seen from our neural network (Yanai 2010). Finally, we also account for the size of risk units (RMA 2024) since policies cover collections of fields and not geohashes or individual fields. Altogether, this paints a picture of changes to yield relative to APH at the 2030 and 2050 timeframes. For more information, see our scientific paper (preprint).
- - Time: In general things get worse in terms of losses as climate change progresses during the 2030 and then 2050 series. - -
- - Coverage levels: The picture may look different for policies insuring 75% of yield versus 85% of yield due to the way that the edges of the distribution or "tails" are changing. Those insuring at higher levels see some very high claims rates under climate change. + Time: In general things get worse in terms of losses as climate change progresses during the 2030 and then 2050 series as warming progresses.
- - Reference point: Things look different depending on what we compare against. Contrasting the continued climate change scenario (SSP245) to "historic" values in the recent past sees average yields stay about the same but the risk of loss increases. On the other hand, comparing to the "counterfactual" (a simulated future where climate change stops today) has our neural network actually expecting yields to increase relative to SSP245. + Coverage levels: The picture may look different for policies insuring 75% of yield versus 85% of yield. Those insuring at higher levels see some very high claims rates under climate change.
-
Unit size: The effects of climate change become less pronounced when simulating with the full insured unit where there may be more fields in a single policy compared to sub-unit which simulates at a smaller scale (simulates 5 character geohashes and near-individual field samples). This is due to the "portfolio effect" where larger units offer more chances for high performing fields to offset fields with worse outcomes (Knight et al 2024).
@@ -637,7 +618,7 @@
Predicted Distributions E
- Taken altogether, these results reveal a very specific threat to corn yields and crop insurance. Generally yields improve over time (Nielsen 2023) but climate change in our simulation wipes out most of the increases our neural network would otherwise expect if climate change stopped. This means that the average yield does not necessarily change dramatically from the historic APH but the ends of the distribution of yield outcomes or "tails" evolve. This represents less yield stability in the future such that the APH does not necessarily capture a risk increase even as the rate of claims for which YP pays out more than doubles in the 2050 series. This specific pairing of unchanged average yields but higher volatility is at the center of the risk posed to crop insurance.
+ Taken altogether, these results reveal a very specific threat to corn yields and crop insurance. In particular, the ends of the distribution of yield outcomes or "tails" evolve. This represents less yield stability in the future such that the APH does not necessarily capture all increasing risk even as the rate of claims for which YP pays out more than doubles in the 2050 series. This specific pairing of unchanged average yields but higher volatility is at the center of the risk posed to crop insurance.
Predicted Distributions E
The resolution of these data could also influence results. In geohashing we currently use the 4 character size corresponding to 28 by 20 kilometers (Haugen 2020). We could also try:
More about simulation resolution
@@ -646,8 +627,8 @@ Predicted Distributions E
The resolution of these data could also influence results. In geohashing we currently use the 4 character size corresponding to 28 by 20 kilometers (Haugen 2020). We could also try:
- - Making the grid cells larger: Currently the data are finding significant results in 91% of grid cells so increasing the sample at the cost of predictive resolution offers very limited benefit.
- - Making the grid cells smaller: We attempt to approximate this in the sub-unit size results with simulation of 5 character geohashes. That said, this lowers the frequency with which a grid cell sees significant results by more than half though this number can vary from simulation to simulation. This makes sense as the 5 character level nears the resolution of the underlying data so increasing resolution introduces unacceptable noise.
+ - Making the grid cells larger: Currently the data are finding significant results in most grid cells so increasing the sample at the cost of predictive resolution offers very limited benefit.
+ - Making the grid cells smaller: We attempt to approximate this in the sub-unit size results with simulation of 5 character geohashes. That said, this lowers the frequency with which a grid cell sees significant results substantially though this number can vary from simulation to simulation. This makes sense as the 5 character level nears the resolution of the underlying data so increasing resolution introduces unacceptable noise.
Altogether, we continue to suggest use of the 4 character geohash.
@@ -693,7 +674,7 @@ Predicted Distributions E
Claims Visualization
+Claims Visualization
We have a number of options to confront these increases in risk. For example, regenerative practices can reduce risks through diverse crop rotations (Bowles et al 2020) and improvements to soil health (Renwick et al 2021). This could provide output stability in addition to other environmental benefits (Hunt et al 2020).
@@ -718,10 +699,10 @@
Claims Visualization
- If we have the tools, what is the hold up? Recalling the rates explorer, the average-based APH may play a part in the delay. Despite valuable resilience offered by regenerative agriculture (given that our results see higher claims not through overall reduced averages but higher volatility), these important practices may not always improve mean yields or can even come at the cost of a slightly reduced average (Deines et al 2023). This may be part of why, even though they guard against elevations in the probability of loss events (Renwick et al 2021), crop insurance may discourage these resilience-building steps (Wang et al 2021; Chemeris et al 2022).
+ If we have the tools, what is the hold up? Recalling the rates explorer, the average-based APH may play a part in the delay. Despite valuable resilience offered by regenerative agriculture, these important practices may not always improve mean yields or can even come at the cost of a slightly reduced average (Deines et al 2023). This may be part of why, even though they guard against elevations in the probability of loss events (Renwick et al 2021), crop insurance may discourage these resilience-building steps (Wang et al 2021; Chemeris et al 2022).
- Instead of insuring up to a percentage below average yield, what if instead one insures up to a certain number of standard deviations below average? This next interactive explores this question by looking at how different risk units might respond to this kind of structural shift.
+ Alternative formulations require more reserach. However, for the sake of exploration, consider what if one insures up to a certain number of standard deviations below average instead of insuring up to a percentage below average yield. This next interactive explores this question by looking at how different risk units might respond to this kind of structural shift.
@@ -768,7 +749,7 @@ More about the policy
Claims Visualization
- Using recent data, the equivalent to insuring up to 25% below average using present-day risk is insuring up to 2.11 standard deviations below average though this does already have some small effects absent behavior change as units not typically making claims pull down that equivalency or, in other words, it already shifts exposure towards more stable yields. Here is how this would get formulated:
+ Using recent data, the equivalent to insuring up to 25% below average using present-day risk is insuring up to about 1.5 standard deviations below average though this does already have some small effects absent behavior change as units not typically making claims pull down that equivalency or, in other words, it already shifts exposure towards more stable yields. Here is how this would get formulated:
- Average-based (current): loss = max(coverage percentage * average yield - actual yield, 0) @@ -812,7 +793,7 @@
- export_claims.csv: The claims rate expected under different scenarios. Here threshold refers to coverage level (YP is 0.25 meaning 25% below average), thresholdStd refers to the translation of an average-based APH to std, and offsetBaseline refers to if this includes APH changes over time (we recommend using "always"). @@ -824,16 +805,16 @@
- - Preciptiation may offer a protective benefit: an increased chirps value has a weak but significant rank correlation (Spearman 1904) to lower changes in risk. This would potentially agree with prior work (Sinsawat et al 2004; Marouf et al 2013). + Preciptiation may offer a protective benefit: an increased chirps value may be associated with less loss elevation. This would potentially agree with prior work (Sinsawat et al 2004; Marouf et al 2013).
-
Different geographies may see different outcomes with increased risk potentially concentrated in a band stretching through Iowa, Illinois, and Indiana.
@@ -900,10 +881,7 @@
Neighborhood Explorer
- Of course, the beauty of an explorable explanation is that you can dig beyond our original intent as you look through each of the variables that this model uses. That said, additional insights can be found in our scientific paper (preprint) including more details about the neural network itself and its performance in different tasks. -
Claims Visualization
Data downloads for discussion
- The data downloads for this discussion section ("how might we address these concerns") are as follows:
+ The data downloads section are as follows:
Claims Visualization
Neighborhood Explorer
+Neighborhood Explorer
- So far, we have explored data aggregated across many neighborhoods. This final graphic looks at those neighborhoods individually. Note that this refers to statistically significant results to cut through noise but, if desired, one can turn off that filter.
+ So far, we have explored data aggregated across many neighborhoods. This final graphic maps those neighborhoods individually. Note that this refers to statistically significant results to cut through noise but, if desired, one can turn off that filter.
@@ -890,7 +871,7 @@
- More about the statistical tests
- Around 91% of maize-growing acerage examined in this study see statistically significant results (p < 0.05 / n) in at least one year of the 2050 series. We determine this by a Bonferroni-corrected (Bonferroni 1935; McDonald 2014) Mann Whitney U (Mann and Whitney 1947; McDonald 2014) per neighborhood per year as variance may differ between the two expected and counterfactual sets (McDonald 2014).
+ Around 95% of maize-growing acerage examined in this study see statistically significant results (p < 0.05 / n) in at least one year of the 2050 series. We determine this by a Bonferroni-corrected (Bonferroni 1935; McDonald 2014) Mann Whitney U (Mann and Whitney 1947; McDonald 2014) per neighborhood per year as variance may differ between the two expected and counterfactual sets (McDonald 2014).
Neighborhood Explorer
- With that, thank you for joining us on this journey through modeling the future of corn in the US!
+ In addition to digging in beyond our original findings as you look through each of the variables that this model uses, additional insights can be found in our scientific paper (preprint) including more details about the neural network itself and its performance in different tasks.