Researcher leverages RavenPack data to test neural networks

In his research paper, "Learning Deep News Sentiment Representations for Macro Finance," Axel Groß-Klußmann of Quoniam Asset Management GmbH employs innovative neural network techniques designed to tackle the challenge of extracting latent economic factors from vast volumes of news analytics data.

Groß-Klußmann's research leverages two comprehensive datasets of daily news sentiment scores, encompassing over 1,000 topics with a focus on macroeconomic themes: Ravenpack News Analytics and GDELT. The two datasets were chosen to complement each other: while the Ravenpack dataset is finance-focused, the GDELT data mainly focuses on politics. The author then goes on to employ neural networks to leverage high-dimensional sentiment data (involving a large number of features or variables) to extract low-dimensional hidden insights (simpler, more condensed data points with fewer variables or features).

The power of Supervised Autoencoders

The study introduces supervised autoencoders, a sophisticated computational framework (neural network) that learns from examples and leverages feedback to improve its outputs, notably with tasks related to how different investments perform (common asset class returns) and traits of the market. These neural network-based representations work better than other methods like PCA (Principal Component Analysis) and PLS (Partial Least Squares) in predicting future GDP growth, understanding why different investments perform as they do, and making investments based on the direction of trends over time.

Capturing fine market dynamics through a multifaceted approach

The study stands out by focusing on high-frequency, daily sentiment scores rather than the usual monthly data aggregation, allowing for a more detailed view of market dynamics. This daily approach not only captures finer nuances but also benefits from a larger dataset, enhancing the strength of statistical models used for analysis.

Additionally, the research adopts a comprehensive strategy, merging macroeconomic data, asset-specific time series, and macro-financial news sentiment. Groß-Klußmann explains how they collected Real-time historical economic data from organizations like OECD and ECB, cleaned it by eliminating trends and seasonal patterns, and then condensed it into vital economic factors. This multifaceted approach includes examining stock and bond returns, FX forward returns, and commodity futures data across specific regions.

Key Takeaways

• Neural network-encoded sentiment insights, especially supervised autoencoders, show clear connections with stock market, government bond, and foreign exchange returns;
• Short-term trading strategies based on daily sentiment insights outperform traditional methods, indicating potential opportunities for investors to capitalize on market momentum;
• Adding a reconstruction loss to the architecture (a measure that checks how well the model rebuilds data) enhances stability of the model, providing investors with more confidence in the reliability and interpretability of sentiment-based predictions;
• Sentiment data proves effective in capturing economic trends, including the impact of the Covid-19 pandemic, offering insights for investors navigating uncertain market conditions.
• Daily sentiment predictors outperform monthly aggregations in forecasting real GDP growth, showcasing the potential for investors to gain an early assessment of economic trends using high-frequency sentiment data.

By harnessing the capabilities of neural networks, specifically supervised autoencoders, the study demonstrates the potential of deep news sentiment representations to enhance our understanding of economic dynamics. As the financial landscape continues to evolve, the insights from this research pave the way for more sophisticated and nuanced approaches to news analytics in the pursuit of informed decision-making