Abstract
Many scientific and economic applications involve the statistical learning of high-dimensional functional time series (FTS), where the number of functional variables is comparable to, or even greater than, the number of serially dependent functional observations. This paper introduces a robust framework for modeling observed functional time series subject to measurement errors. By leveraging the fact that the autocovariance function of observed FTS automatically filters out white noise, we propose a three-step procedure: autocovariance-based dimension reduction, a novel block regularized minimum distance (RMD) estimation framework to produce block sparse estimates, and final functional sparse reconstruction. We investigate the theoretical properties of these estimators and demonstrate, via simulated and real datasets, that our proposed framework significantly outperforms existing competitors.
Introduction
In fields ranging from finance (e.g., intraday stock price trajectories) to environmental science (e.g., daily pollution levels), data often arrives as Functional Time Series (FTS)—sequences of curves observed over time. Analyzing this data presents two significant challenges for researchers:
- High Dimensionality ($p > n$): Modern datasets often feature many functional variables (e.g., hundreds of different stocks) but a limited number of time points. Standard statistical tools frequently break down in this "large $p$, small $n$" regime.
- Measurement Error: Real-world curves are rarely "clean"; they are typically observed with noise. Traditional methods, such as Functional Principal Component Analysis (FPCA), rely on the covariance of the data at the same time point (lag 0). Because measurement noise is strongest at lag 0, these methods often yield biased results and poor predictive performance.
To address these hurdles, we introduce AUTO, a robust learning framework designed to be both high-dimension capable and naturally resistant to noise.
The AUTO Framework
The AUTO framework employs a three-step procedure to ensure accuracy and interpretability in high-dimensional settings:
1. Autocovariance-Based Dimension Reduction
Instead of using standard FPCA, which relies on lag-0 covariance, AUTO performs an eigenanalysis on an operator built from lagged autocovariances (where the time lag $h > 0$). Because white noise is uncorrelated over time, it naturally disappears at these lags. This allows the model to isolate the "clean" signal dynamics without the interference of measurement error.
2. Block Regularized Minimum Distance (RMD) Estimation
To manage high dimensionality, the problem is formulated as a set of moment equations. We utilize a convex optimization approach with a "block sparsity" penalty, similar to a Group Lasso. This ensures that the model selects only the most relevant functional variables. In the context of functional data, this means that if a variable is deemed irrelevant, the entire curve is excluded from the model, rather than just isolated points along the curve.
3. Functional Reconstruction
Finally, the sparse estimates are mapped back into the functional domain to produce the final curves or transition matrices, providing a clear and interpretable result for practitioners.
Key Insights
The "Free" Noise Filter
The most elegant aspect of the AUTO framework is that the autocovariance function of a noisy time series automatically ignores white noise. By shifting the analysis to lags $h \geq 1$, the underlying process is effectively "denoised" without requiring complex error modeling.
Non-Asymptotic Guarantees
Unlike many methods that rely on large-sample assumptions, the AUTO framework is supported by rigorous mathematical proofs (error bounds) that hold for finite sample sizes. This provides critical reliability for practitioners working with limited historical data.
Empirical Results
Simulation Studies
In extensive tests across three model types—Scalar-on-Function Linear Regression (SFLR), Function-on-Function Linear Regression (FFLR), and Vector Functional Autoregression (VFAR)—the AUTO method consistently achieved lower relative estimation errors than standard covariance-based (COV) methods. The performance gap widened as the sample size $n$ increased, highlighting the efficiency of the autocovariance approach.
Real-World Application: S&P 100 Stock Data
We applied the framework to Cumulative Intraday Return (CIDR) trajectories for 98 stocks in the S&P 100 index during 2017 ($n=251$ trading days). The goal was to predict the intraday return of the index based on the trajectories of its component stocks.
AUTO achieved the lowest Mean Squared Prediction Error (MSPE) across all tested time horizons. For instance, at $N=360$ minutes into the trading day, AUTO's MSPE was 3.892, significantly outperforming the standard COV method (4.976) and other benchmarks like AGMM.
Conclusion and Future Work
The AUTO framework provides a robust, sparse, and noise-resistant solution for high-dimensional functional time series analysis. While the method is robust to the choice of lags, future research may focus on automated parameter selection (e.g., refined AIC/BIC for functional data) and extending the theoretical framework to handle "heavy-tailed" non-Gaussian data, which is common in specific financial contexts. Despite these areas for growth, AUTO represents a significant step forward in filtering the noise of complex, high-dimensional data.
Summary updated