Title: Using maps to predict economic activity

Abstract: We introduce a novel machine learning approach to leverage historical and contemporary maps and systematically predict economic statistics. Our simple algorithm extracts meaningful features from the maps based on their color compositions for predictions. We apply our method to grid-level population levels in Sub-Saharan Africa in the 1950s and South Korea in 1930, 1970, and 2015. Our results show that maps can reliably predict population density in the mid-20th century Sub-Saharan Africa using 9,886 map grids (5km by 5 km). Similarly, contemporary South Korean maps can generate robust predictions on income, consumption, employment, population density, and electric consumption. In addition, our method is capable of predicting historical South Korean population growth over a century.