The Potential Implications of 'Big Ag Data' for USDA Forecasts

Recent advances in precision agriculture technology have increased the potential to capture near-real time data such as planting and yield information. It is well established that information on crop acreage and yield can have value in commodity markets. That is why the USDA conducts farm surveys and freely reports such information. In this study we use a large sample (just over 1.5 million observations) of farm-level corn yield data to consider if it is possible to use non-random farm yield data (such as might be available to providers of precision agriculture services) to accurately predict the national corn yield. Specifically, we examine scenarios where a forecasting agent has data that is not representative either because it is from a limited region or because it consists primarily of large farms. In general, we conclude that large volumes of data can, to some degree, overcome forecasting bias caused by non-representative samples. Moreover, if the forecaster can benchmark against an unbiased estimator, it may be possible to remove much of the bias from estimates generated by non-representative samples