Statistics

Statistics is the discipline of learning from data through design, analysis and interpretation of experiments and observations. Its foundations lie in probability theory, enabling quantification of uncertainty, hypothesis testing and parameter estimation. Modern statistical science addresses challenges of large-scale and high-dimensional data by developing regularised estimators—such as shrinkage techniques for covariance matrices—and by integrating classical inference with computational methods like Monte Carlo, machine learning and Bayesian sampling. Key themes include reproducible workflows, scalable algorithms for complex and streaming data, and robust uncertainty quantification. Applications range from economic monitoring and genomics to environmental modelling and finance, where data-driven predictions and risk assessments guide decisions across research, industry and public policy.

Research from Nature Portfolio

New shrinkage estimators have been developed for large covariance matrices with Toeplitz structure, common in array-signal and time-series contexts. By deriving closed-form tuning parameters under a mean-squared-error criterion and estimating them unbiasedly from data, these methods yield well-conditioned covariance estimates that outperform prior approaches in high-dimension, low-sample-size regimes.

A global model for estimating household and district wealth has been introduced by fusing public satellite imagery, social-media marketing metrics and geospatial features. Trained on over 60 000 survey clusters across 59 countries, it explains a majority of variation in observed wealth levels and their changes, demonstrating the power of combining private and public data for real-time economic surveillance.

Exact Gaussian processes have been scaled to millions of observations by inventing ultra-flexible, non-stationary, sparsity-discovering covariance kernels. Rather than inducing sparsity via approximations, these kernels learn zero and non-zero covariances directly, preserving full uncertainty quantification while enabling tractable inference on massive spatial and temporal datasets.

Topic trend for the past 5 years

The graph below shows the article count in Nature Index journals for statistics.

Technical terms

High-dimensional data: Data in which the number of variables rivals or exceeds the number of observations, challenging classical inference.

Covariance matrix: A matrix of pairwise covariances quantifying linear dependence among multivariate data.

Shrinkage estimator: A regularised estimator blending sample statistics with structured targets to reduce variance and enforce stability.

Toeplitz structure: A covariance form with constant diagonals, reflecting stationarity in array-signal or time-series models.

Gaussian process: A distribution over functions where any finite collection of values follows a joint normal law, used for spatial-temporal modelling.

Non-stationary kernel: A covariance function whose properties vary over the input domain, allowing local adaptation in Gaussian processes.

Sufficient dimension reduction: Techniques that find low-dimensional projections of predictors retaining all information about the response.

Mutual information: An information-theoretic measure of dependence between random variables, guiding feature compression and selection.

James–Stein estimator: A shrinkage method that improves estimation accuracy by pulling sample eigenvectors toward a central target.

Notable articles in statistics

1. State estimation of a physical system with unknown governing equations. Nature (2023).
2. An additive Gaussian process regression model for interpretable non-parametric analysis of longitudinal data. Nature Communications (2019).
3. Self-learning Monte Carlo method: Continuous-time algorithm. Physical Review B (2017).
4. A Universal Operator Growth Hypothesis. Physical Review X (2019).
5. FEAST: fast expectation-maximization for microbial source tracking. Nature Methods (2019).
6. Spectral fluctuations in the Sachdev-Ye-Kitaev model. Journal of High Energy Physics (2020).
7. Guiding new physics searches with unsupervised learning. European Physical Journal C (2019).
8. James–Stein for the leading eigenvector. Proceedings of the National Academy of Sciences of the United States of America (2023).

Position of Statistics in Nature Index by Count

Top 5 leading collaborators in Statistics

Collaborating institutions