Time-Series Analysis

Time-series analysis encompasses a spectrum of statistical and computational techniques for characterising, modelling and predicting data indexed by time. Core objectives include the decomposition of series into trend, seasonal and residual components; the assessment of stationarity and autocorrelation; and the identification of abrupt structural shifts or change points. Classical approaches such as autoregressive integrated moving average models rely on linear dependence and stationarity assumptions, while modern frameworks incorporate nonparametric learning, network methods and dimension‐reduction to handle nonstationarity, high dimensionality and complex dependencies. Advances in operator‐theoretic and wavelet‐based decompositions enable extraction of persistent cycles and evolving trends from single realisations, whereas graph‐based and factor‐adjusted models accommodate multivariate, non‐Euclidean and mixed‐frequency data. Applications span climate diagnostics, economic forecasting, anomaly detection in industrial systems and structural analysis of neural and genomic recordings. Together, these methods form an integrated toolkit for uncovering latent dynamics, signalling change in real time and making robust forecasts in domains where data are noisy, heterogeneous and nonstationary.

Research from Nature Portfolio

Operator‐theoretic methods have been extended to non‐autonomous time series, showing that spectral properties of evolution operators can disentangle slowly decorrelating trends and persistent cycles from a single trajectory. Applications to historical sea‐surface temperature records have revealed nonlinear shifts over the industrial era and characterised glacial transitions in the mid‐Pleistocene, offering a unified framework for trend and mode extraction without ensemble data.

Multichannel singular spectrum analysis has been used to jointly decompose sea‐surface temperature and sea‐level pressure fields in the South Atlantic, isolating a 13-year basin‐wide dipole mode and a 5-year interannual oscillation alongside a nonlinear trend. These modes explain a substantial fraction of regional variability and elucidate teleconnections with the Pacific Decadal Oscillation and El Niño–Southern Oscillation.

Research from all publishers

A nonparametric change‐point analysis constructs similarity graphs to detect both offline segmentation and online shifts in high‐dimensional or non‐Euclidean sequences. By defining graphs on pairwise dissimilarities and controlling the false discovery rate through analytic thresholds, this framework identifies structural breaks in imaging, network and genomic datasets without restrictive distributional assumptions.

An irregular‐sampling anomaly detection method employs neural controlled differential equations to capture joint spatial and temporal dependencies in multivariate streams. Coupled with a distribution‐based scoring scheme, it effectively handles missing values and flags outliers in applications ranging from power‐grid monitoring to industrial process control.

The FNETS factor‐adjusted network estimation framework first removes pervasive co-movements via dynamic principal component analysis, then fits a sparse vector autoregression to residuals. This yields directed and undirected networks of Granger‐causal and contemporaneous links, while providing consistent forecasts for both common factors and idiosyncratic dynamics in ultra-large panels.

Time-Series Analysis publication trend

The graph below shows the total number of articles in time-series analysis across all publications each year (not limited to Nature Index journals).

Technical terms

Stationarity: The property of a time series whose statistical moments (mean, variance, autocovariance) remain constant over time.

Autocorrelation: The correlation between observations of a time series at different lags, indicating serial dependence.

Change point: A time at which the underlying data‐generating process undergoes an abrupt shift in parameters or distribution.

Wavelet transform: A time–frequency decomposition that represents a series in terms of localized basis functions at multiple scales.

Factor model: A representation in which multivariate series are driven by a small number of latent common factors plus idiosyncratic noise.

Similarity graph: A graph whose nodes represent observations and whose edges encode pairwise similarity or dissimilarity for nonparametric analysis.

References

1. Revealing trends and persistent cycles of non-autonomous systems with autonomous operator-theoretic techniques. Nature Communications (2024).
2. The South Atlantic Dipole via multichannel singular spectrum analysis. Scientific Reports (2024).
3. Graph-Based Change-Point Analysis. Annual Review of Statistics and Its Application (2023).
4. Graph spatiotemporal process for multivariate time series anomaly detection with missing values. Information Fusion (2024).
5. FNETS: Factor-Adjusted Network Estimation and Forecasting for High-Dimensional Time Series. Journal of Business and Economic Statistics (2023).