Abstract:Forecasting the evolution of complex systems is one of the grand challenges of modern data science. The fundamental difficulty lies in understanding the structure of the observed stochastic process. In this paper, we show that every uniformly-positive-definite-in-covariance and sufficiently short-range dependent non-stationary and nonlinear time series can be well approximated globally by an auto-regressive process of slowly diverging order. When linear prediction with ${\cal L}^2$ loss is concerned, the latter result facilitates a unified globally-optimal short-term forecasting theory for a wide class of locally stationary time series asymptotically. A nonparametric sieve method is proposed to globally and adaptively estimate the optimal forecasting coefficient functions and the associated mean squared error of forecast. An adaptive stability test is proposed to check whether the optimal forecasting coefficients are time-varying, a frequently-encountered question for practitioners and researchers of time series. Furthermore, partial auto-correlation functions (PACF) of general non-stationary time series are studied and used as a visual tool to explore the linear dependence structure of such series. We use extensive numerical simulations and two real data examples to illustrate the usefulness of our results.

What problem does this paper attempt to address?

The main problems that this paper attempts to solve are the globally optimal and adaptive short - term prediction problems for locally stationary time series, as well as testing the stability of such predictions. Specifically: 1. **Globally Optimal Short - term Prediction**: The paper shows that every non - stationary, nonlinear time series with short - range dependence and uniformly positive definite covariance (UPDC) can be globally approximated by an autoregressive (AR) process of slowly diverging order. When considering linear prediction using the $L_2$ loss function, this result provides a unified globally optimal short - term prediction theory for a wide class of locally stationary time series. 2. **Estimation of Prediction Coefficients**: A non - parametric screening method is proposed to globally and adaptively estimate the optimal prediction coefficient function and its associated mean squared error (MSE). This method uses all available time series observations to determine the optimal prediction coefficients, thereby improving prediction efficiency. 3. **Stability Test of Prediction Coefficients**: An adaptive stability test is proposed to check whether the optimal prediction coefficients change over time. This is very important for time series prediction in practice, because many practitioners tend to use the classical ARMA model with constant coefficients for prediction. 4. **Research on Partial Autocorrelation Function (PACF)**: The PACF of generalized non - stationary time series is studied and used as a visualization tool to explore the linear dependence structure of these sequences and preliminarily determine the order of the AR model required for optimal prediction. ### Formula Summary - **AR Approximation**: \[ x_i=\phi_{i0}+\sum_{j = 1}^{\min\{b, i - 1\}}\phi_{ij}x_{i - j}+\epsilon_i+O_P(n^{-1+(2 + \epsilon)/\tau}) \] where $\epsilon_i$ is a time - white - noise process, that is, $E[\epsilon_i]=0$ and $Cov(\epsilon_i,\epsilon_j)=\mathbb{1}(i = j)\sigma_i^2$. - **Smooth Approximation of Optimal Prediction Coefficients**: \[ x_i=\phi_0\left(\frac{i}{n}\right)+\sum_{j = 1}^{b}\phi_j\left(\frac{i}{n}\right)x_{i - j}+\epsilon_i+O_P(n^{-1+5(1 + \epsilon)/(2\tau)}) \] - **Smooth Approximation of PACF**: \[ \rho_j(t)\in C^d([0,1]) \] and for some constant $C>0$, \[ \sup_{i>b}\left|\rho_j\left(\frac{i}{n}\right)-\rho_{i,j}\right|\leq Cn^{-1+3(1 + \epsilon)/(2\tau)} \] Through these methods, the paper provides a theoretical basis and practical tools for short - term prediction of locally stationary time series, especially in non - stationary cases.

Globally Optimal And Adaptive Short-Term Forecast of Locally Stationary Time Series And A Test for Its Stability

Specification Testing in Nonstationary Time Series Models

Auto-Regressive Approximations to Non-stationary Time Series, with Inference and Applications

On the partial autocorrelation function for locally stationary time series: characterization, estimation and inference

Testing Strict Stationarity With Applications To Macroeconomic Time Series

Approach to Testing Nonlinear Time Series Models

A Stability Principle for Learning under Non-Stationarity

Short-term Load Forecasting Utilizing Wavelet Transform and Time Series Considering Accuracy Feedback

An Improved Generalized Spectral Test for Conditional Mean Models in Time Series with Conditional Heteroskedasticity of Unknown Form

Generalized Spectral Tests for Conditional Mean Models in Time Series with Conditional Heteroscedasticity of Unknown Form

Diagnostic Checking for the Adequacy of Nonlinear Time Series Models

TimeBridge: Non-Stationarity Matters for Long-term Time Series Forecasting

Theory and Algorithms for Forecasting Time Series

NonSTOP: A NonSTationary Online Prediction Method for Time Series

Non-parametric Probabilistic Time Series Forecasting via Innovations Representation

Prediction of Locally Stationary Data Using Expert Advice

Channel-Aware Low-Rank Adaptation in Time Series Forecasting

Time Series Forecasting for Structures Subjected to Nonstationary Inputs

Predictive Inference for Locally Stationary Time Series With an Application to Climate Data

Optimal Change-point Testing for High-dimensional Linear Models with Temporal Dependence

Uniform reliability tests for forecasting systems with small lead time