Causal Models For Time Series
Artificial Intelligence & Machine Learning
Advanced Causal Discovery in Non-Stationary Financial Markets: The CD-NOTS Framework
In the complex landscape of financial markets, understanding the causal relationships between different financial entities is crucial. A recent study led by Agathe Sadeghi from Stevens Institute of Technology, along with Achintya Gopal and Mohammad Fesanghary from Bloomberg, has introduced a novel framework that significantly advances our ability to discern these relationships within non-stationary time-series data. Their research, detailed in a paper that was a finalist for the Abu Dhabi Investment Authority (ADIA) Lab Award for Causal Research in Investment, presents a new method termed Constraint-based Causal Discovery from Heterogeneous Time-Series Data (CD-NOTS).
The CD-NOTS method is designed to identify causal dependencies in financial time-series data, which are inherently dynamic and complex.
This approach extends existing algorithms by incorporating the ability to account for lagged relationships, a feature critical in financial data analysis. The CD-NOTS framework is distinguished by its nonparametric nature, allowing it to uncover both linear and non-linear causal relationships as well as handle non-Gaussian distributions effectively.
The core of the CD-NOTS approach lies in its robust methodology that operates under realistic assumptions suitable for financial data. These include the capability to manage nonstationarity—a common characteristic of financial time-series data where statistical properties change over time. The framework also effectively detects both instantaneous and lagged causal effects, providing a comprehensive view of the causal dynamics within the financial markets.
The CD-NOTS algorithm requires several key assumptions to function effectively:
- All potential confounders are functions of time, allowing for dynamic adjustment in the analysis.
- The data must adhere to the principle of faithfulness, where observed independencies suggest causal independencies.
- The absence of selection bias ensures that the sampled data are representative of the broader dataset.
- Stability in the data-generating process over time is crucial for consistent causal inference.
Utilizing simulated data, the researchers applied the CD-NOTS method to various financial indices and factors to identify causal connections. These applications demonstrate the potential of causal discovery in enhancing factor-based investing strategies, diversifying portfolios, and deepening the understanding of market dynamics.
In comparative tests, CD-NOTS has shown superior performance over traditional causal discovery algorithms like PCMCI.
This is reflected in the significantly higher F1 scores achieved by CD-NOTS, indicating a more accurate identification of true causal relationships. The improved performance stems from the algorithm’s ability to handle the complex, non-linear, and non-stationary nature of financial data, which are often inadequately addressed by other methods.
In conclusion, the introduction of the CD-NOTS framework marks a significant advancement in the field of causal discovery within financial markets. It provides a more accurate, flexible, and comprehensive tool for analyzing the causal relationships in economic and financial time series. Looking ahead, the research team suggests potential avenues for further enhancing the robustness of causal discovery methods and extending their applicability to other complex datasets.
Causal+Discovery+in+Financial+Markets.pdf (squarespace.com)
For further reading see: Does Causality Predict Markets Movements Or Is It A Flawed Overhyped Tool?
