Abstract In multivariate time series systems, lead-lag relationships reveal dependencies between time series when they are shifted in time relative to each other. Uncovering such relationships is valuable in downstream tasks, such as control, forecasting, and clustering. By understanding the temporal dependencies between different time series, one can better comprehend the complex interactions and patterns within the system. We develop a cluster-driven methodology based on dynamic time warping for robust detection of lead-lag relationships in lagged multi-factor models. We establish connections to the multireference alignment problem for both the homogeneous and heterogeneous settings. Since multivariate time series are ubiquitous in a wide range of domains, we demonstrate that our algorithm is able to robustly detect lead-lag relationships in financial markets, which can be subsequently leveraged in trading strategies with significant economic benefits.
Introduction
In complex systems like global financial markets or biological networks, time series often exhibit "lead-lag" relationships. This phenomenon occurs when one variable moves first, and another follows with a temporal delay. Identifying these patterns is crucial for forecasting and system control, yet it remains a significant challenge due to high noise levels in real-world data and temporal distortions where the "speed" of a lag varies over time.
Traditional methods, such as Cross-Correlation Functions (CCF), often struggle with these non-linear distortions and become computationally prohibitive as the number of variables increases. This paper introduces DTW_KMed, a robust and scalable framework designed to uncover these latent structures within lagged multi-factor models.
Methodology
The core of the proposed framework is the DTW_KMed algorithm, which combines Dynamic Time Warping (DTW) with K-Medoids clustering. The process follows three critical steps:
- Dynamic Time Warping (DTW): Unlike standard Euclidean distance, which compares points at identical timestamps, DTW "warps" the time axis to find the optimal alignment between two sequences. This allows the model to detect relationships even if one series is a stretched or compressed version of another.
- K-Medoids Clustering: To manage high dimensionality and filter out noise, the algorithm groups similar time series based on their DTW distances. This ensures that lag detection is focused on assets that are fundamentally related.
- Lag Estimation: For pairs within a cluster, the algorithm examines the "warping path" (the map of how the two series align). By calculating the index differences along this path and identifying the mode or median, the relative lag between assets can be precisely estimated.
Datasets and Validation
The researchers validated the DTW_KMed approach using a combination of synthetic and real-world datasets:
- Synthetic Data: Generated from lagged multi-factor models to test the algorithm against known ground-truth lags under varying noise conditions.
- US Equities: 679 stocks from Wharton’s CRSP (2000–2019).
- ETFs: 14 major Exchange Traded Funds (2006–2019).
- Futures: 52 different contracts across grains, metals, indices, and currencies (2000–2020).
Key Insights
- Clustering as a Denoising Filter: Using K-Medoids helps the algorithm ignore irrelevant noise from unrelated time series, focusing only on meaningful relationships within clusters.
- Window Size Constraints: For DTW to be effective, the "warping window" must be at least as large as the true lag. If the window is too small, performance degrades to that of standard Euclidean distance.
- Mode vs. Median: In noisy environments, using the mode (the most frequent lag value along the warping path) generally outperformed the median for detecting the true relationship.
- Economic Value: Lead-lag relationships represent "information leakage," where leaders incorporate news faster than laggers. Exploiting this can lead to superior trading performance.
Results
The DTW_KMed framework demonstrated significant improvements over traditional benchmarks:
- Efficiency: The algorithm is approximately 10 times faster than the CCF benchmark, making it viable for high-dimensional datasets.
- Robustness: In synthetic tests, the algorithm maintained a high Adjusted Rand Index (ARI) even as noise levels increased, whereas traditional methods saw performance plummet.
- Trading Performance: On the US Equity dataset, the DTW-based strategy achieved a Sharpe Ratio of 0.934, significantly outperforming the CCF benchmark of 0.67.
Limitations and Future Work
While the results are promising, the authors identify several areas for future research. The current model assumes "Single Membership," where each time series is driven by one factor; future work could explore "Mixed Membership" models. Additionally, while this study utilized daily data, applying the framework to intraday (minutely or tick-by-tick) data could reveal more granular dynamics that are lost over longer timeframes.
Summary updated