Eye in outer space: satellite imageries of container ports can predict world stock returns

Predicting stock returns is notoriously difficult, even with traditional economic data. These data are often released with lags, are subject to revisions, and are publicly available, quickly losing predictive power as they are incorporated into market prices. This paper proposes an alternative data source: satellite imagery of container ports. The number of containers at a port reflects real-time economic activity; increased congestion suggests decreased demand and potential economic downturn. Leveraging advancements in artificial intelligence and big data, the study aims to show that satellite-based container information can provide a real-time, previously untapped source of information for stock market prediction. The authors hypothesize that increases in the number of containers in ports will negatively correlate with future stock returns.

The paper reviews the challenges in using traditional macroeconomic indicators (like CPI and GDP) for stock return prediction due to lags, revisions, and public accessibility. It then examines existing literature on alternative data sources, highlighting the use of satellite imagery by hedge funds to gain insights into economic activity. Several studies are cited that use satellite data (e.g., nightlights, parking lot imagery) to measure economic variables, but these studies often lack the temporal resolution needed for short-term prediction. This paper differentiates itself by directly using publicly available satellite data to build its own database and model, instead of relying on commercial data.

The study utilized 83,672 Sentinel-2 satellite images (2017-2021) of the top 48 container ports globally. A U-Net model, a type of convolutional neural network (CNN) known for its effectiveness in semantic segmentation tasks, was trained using 3711 hand-labeled images from 2017 to identify container areas in the satellite images. The container coverage area was used as a proxy for the number of containers. Daily average change in the number of containers (GNC) was calculated for each port, and a combined GNC index was created by averaging across all ports. The GNC index was then used to predict daily stock returns for 33 major stock indices across 28 countries (18 developed, 12 emerging) using univariate predictive regression. The forecast combination technique, with equal weighting, aggregated the individual port predictions to generate a single forecast for each market. Out-of-sample R-squared (R²oos), success ratio (SR), and cumulative sum of squared prediction error difference (CSSED) were used to evaluate the model's performance against a benchmark of the historical average. The Clark and West (2007) test and a bootstrap-based Diebold-Mariano (1995) test were used for statistical significance. Investment strategies (tilted and untilted) were implemented to assess the economic significance of the findings.

The U-Net model demonstrated good performance in identifying container areas, achieving an accuracy of 93.20%, recall of 92.45%, and F-score of 92.81%. The combined GNC index showed significant out-of-sample predictive power for stock returns across most markets (27 out of 33) at horizons up to five days. The average out-of-sample R²oos was 0.0529% at a one-day horizon. Bootstrap-based Diebold-Mariano tests confirmed the statistical significance of the forecasting improvement. The success ratio was higher than 0.5 for 27 markets at the one-day horizon. The CSSED plots show that the GNC model consistently outperformed the historical average benchmark over time, with particularly strong performance during the COVID-19 period. Investment strategies based on the GNC forecasts generated substantial economic gains, outperforming the buy-and-hold strategy with annualized returns of 16.38% (tilted strategy) and 14.85% (untilted strategy). The GNC index was found to lead traditional shipping indicators like the Baltic Dry Index (BDI) and container throughput index by two months. Finally, a negative predictive relationship was observed between GNC and industrial production growth in 27 out of 28 countries at a four-month horizon.

The study's findings strongly support the hypothesis that satellite imagery of container ports provides valuable information for predicting stock returns. The significant and economically meaningful out-of-sample predictive power, particularly during the COVID-19 crisis, highlights the value of this previously underutilized data source. The lead-lag relationship with traditional shipping indicators demonstrates the forward-looking nature of the container data. The close link between GNC and industrial production growth provides a strong economic rationale for the observed return predictability. The results challenge the assumption of perfect market efficiency in the context of costly information acquisition, suggesting that those who invest in acquiring and processing this data can achieve an information advantage and earn excess returns. The enhanced predictive ability during the COVID-19 period may be attributable to the increased supply chain disruptions and their impact on economic activity.

This paper makes a significant contribution by demonstrating the predictive power of satellite-based container data for global stock market returns. This alternative data source offers real-time, high-frequency information, overcoming limitations of traditional economic data. The results highlight the economic significance of this approach and suggest avenues for future research into leveraging alternative data for improved investment decision-making. Future research could explore the use of this methodology with other types of satellite imagery, expanding the range of economic indicators that can be predicted.

The study's reliance on a proxy for container numbers (container coverage area) is a limitation. While the U-Net model performed well, it could not differentiate between stacks of containers, which might affect the accuracy of the container count. The focus on major container ports might limit the generalizability of the findings to smaller ports or regions with less developed infrastructure. Finally, while the authors accounted for time lags in data processing and information dissemination, unmodeled information channels could still affect the results.