Learning Financial Networks with High-frequency Trade Data.

Area of Science:

• Quantitative Finance
• Machine Learning
• Financial Econometrics

Background:

• Traditional financial network analysis relies on low-frequency data, limiting insights into market dynamics.
• High-frequency trading data offers richer information but presents significant modeling challenges (asynchronicity, nonstationarity).
• Existing methods struggle to fully leverage the granularity of intraday trading data for network estimation.

Purpose of the Study:

• To develop a novel method for estimating financial networks using high-frequency intraday trading data.
• To apply machine learning, specifically random forests, to overcome the challenges of high-frequency data.
• To analyze the evolution of financial network connectivity leading up to the 2007-09 U.S. financial crisis.

Main Methods:

• Utilized random forests, a machine learning algorithm, for robust network estimation without extensive hyperparameter tuning.
• Defined network edges by forecasting the change in market measures (e.g., realized volatility) of one firm using microstructure data of another.
• Investigated network evolution and firm-level connectivity patterns.

Main Results:

• Financial network density was highest in 2007, preceding the U.S. financial crisis.
• Lehman Brothers exhibited high connectivity in 2006, indicating its central role.
• Larger firms demonstrated greater predictive power in network linkages, aligning with market microstructure theories.

Conclusions:

• Random forests provide an effective approach to modeling complex financial networks from high-frequency data.
• The study highlights dynamic network changes and key firm roles in the lead-up to the financial crisis.
• Findings underscore the importance of firm size in determining predictive influence within financial networks.