Machine Learning for Daily Return Direction Forecasting: Predicting Socio-Economic Development in Poland in 2020–2024 Using Satellite Imagery

Abstract

Access to up-to-date statistical data is crucial for understanding and responding to socio-economic trends. However, in many parts of the world, official statistics are often delayed, and in some regions, reliable data may be scarce or entirely unavailable. This makes it difficult for researchers, policymakers, and organizations to make informed and timely decisions. Traditionally, nighttime lights satellite imagery has been used as a proxy for economic activity, but this approach lacks precision. Recent advances in high-resolution satellite imagery, combined with increased computational power, offer new opportunities for near real-time socio-economic monitoring. Compared to traditional statistical data, high-resolution imagery provides two key advantages: it is available almost in real time, and predictions can be made for any area, even if it does not align with official administrative boundaries. In this interdisciplinary study, we bridge economics, machine learning, and remote sensing to predict socio-economic development in Poland between 2020 and 2024. We use two approaches for feature extraction from high-resolution satellite imagery, introduce a novel dataset designed to support transfer learning in computer vision, and develop an end-to-end pipeline for predicting regional socio-economic indicators. Our work demonstrates the potential of satellite-based methods to complement, and in some cases substitute, traditional data sources, enabling more timely and granular insights into socio-economic dynamics.