Abstract
Weather forecasts are an example of a public good that is often distributed at-scale without charging users. This feature can make measuring the benefits of weather forecast distribution challenging because information spillovers are likely; people like to talk about the weather and weather information is often available from a variety of sources. Despite the ubiquity of weather information, small-scale farmers often lack access to high-quality weather forecasts that are tailored to help them make production decisions. We implement a randomized experiment with 400,000 cotton growers in Pakistan and vary the share of farmers treated with large clusters (tehsils). We show that treated and untreated farmers in high-saturation clusters update their farming behavior in line with forecasts. Directly treated farmers in high saturation tehsils are 37-67\% more likely to avoid rain when irrigating and applying fertilizer and pesticides. Control farmers in highly saturated tehsils are 22-46\% more likely to avoid rain compared to controls in low-saturation tehsils. For heat avoidance, results follow a similar pattern but are statistically weaker. Direct information sharing is a plausible pathway - control farmers in high saturation areas were 8\% more likely than control farmers in low saturation areas to report discussing weather information with peers. At the end of the season, estimates for yields are positive but imprecise and there is evidence that input expenditure increased.