Modelling Nonstationary Time Series using Trend-Stationary Hypothesis

Zhandos Abdikhadir Department of MathematicsHong Kong Baptist University · 2025

Read Modelling Nonstationary Time Series using Trend-Stationary Hypothesis by Zhandos Abdikhadir Department of MathematicsHong Kong Baptist University (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-07987.txt.

Introduction

Zhandos Abdikhadir and Terence Chong argue that many wobbly data series are better read as steady noise around a fixed trend in this 2025 paper. Mainstream models treat trends as random walks, which can blur sharp turning points. Their framework allows a trend with a few breaks, plus seasons, with calm variation around it.

They give methods to locate the breaks and count them, then fit the full model. Applied to American goods exports to mainland China from 2006 to 2025, the model pinned breaks to major economic events. In a big forecasting contest dataset it beat a popular rival and excelled on series with abrupt shifts, while trailing random-trend models overall. The payoff is a readable story of each series.

Catalog record

CREATORZhandos Abdikhadir Department of MathematicsHong Kong Baptist University
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE47 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-07987

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Modelling Nonstationary Time Series using Trend-Stationary Hypothesis Zhandos Abdikhadir Department of MathematicsHong Kong Baptist University Abstract This paper challenges the prevalence of unit root models by introducing the Linear Trend-Stationary Trigonometric ARMA (LTSTA), a novel framework for modelling nonstationary time series under the trend-stationary hypothesis. LTSTA decomposes series into three components: (1) a deterministic trend (modeled via continuous piecewise linear functions with structural breaks), (2) a Fourier-based deterministic seasonality component, and (3) a stochastic ARMA error term. We propose a heuristic approach to determine the optimal number of structural breaks, with parameter estimation performed through an iterative scheme that integrates a modified dynamic programming algorithm for break detection and a standard regression procedure with ARMA errors. The model’s performance is evaluated through a case study on US Real GDP (2002–2025), where it accurately identifies breaks corresponding to major economic events (e.g., the 2008 financial crisis and COVID-19 shocks). Additionally, LTSTA outperforms well-established univariate statistical models (SES, Theta, TBATS, ETS, ARIMA, and Prophet) on the CIF 2016 forecasting competition dataset across MAE, RMSE, sMAPE, and MASE metrics. The LTSTA model provides an interpretable alternative to unit root approaches, particularly suited for time series with predominant deterministic properties where structural break detection is critical. Keywords: Time series, trend-stationary hypothesis, segmented linear regression, Fourier series, ARMA, forecasting * * footnotetext: Email: [email protected] 1 Introduction Popularized by George E. P. Box and Gwilym M. Jenkins in their seminal work ( Box and Jenkins, 1970 ) , Autoregressive Integrated Moving Average (ARIMA) models are a foundational framework for analyzing and forecasting nonstationary time series. These models are based on the idea that nonstationarity (or a unit root) can be removed through finite-order differencing, thereby transforming unit root processes into a stationary form. Afterwards, the stationarized time series can be described using autoregressive (AR) and moving average (MA) terms. The AR terms capture the linear dependence on past observations, while the MA terms account for the linear dependence between current values and prior forecast errors. The influential findings of Nelson and Plosser, 1982 revealed strong evidence that most economic time series follow unit root processes, further supporting the relevance of ARIMA models in describing a wide range of time series movements and generating reliable short-term forecasts. While differencing nonstationary processes remains a cornerstone of time series analysis, Perron, 1989 challenged the traditional unit root hypothesis. Instead, he proposed that many economic time series are trend-stationary, meaning they can be described with stationary processes that fluctuate around a deterministic trend characterized by shocks or structural breaks. For example, Perron used the Great Crash of 1929 and the 1973 oil crisis as the primary sources of these shocks when evaluating the deterministic trend component in his analysis of macroeconomic time series. These structural changes typically arise from economic crises, policy shifts, technological disruptions, or pandemics. Such events may disrupt the data-generating process and challenge fixed-parameter models like classical ARIMA. Several methods that account for structural changes have been proposed to address this limitation. Among these approaches, Markov Switching Models (MSM, Hamilton, 1989 ) allow parameters to stochastically switch between discrete states like economic expansions and recessions. In contrast, the Self-Exciting Threshold Autoregressive (SETAR) and Threshold Autoregressive (TAR) models ( Tong and Lim, 1980 ) assume that regimes change deterministically when the series crosses a threshold value estimated from its own lagged values or exogenous variables, respectively. In this paper, we propose a model based on the unpopular Perron’s idea that univariate nonstationary time series can be described by trend-stationary processes. The proposed model is termed as the Linear Trend-Stationary Trigonometric ARMA (LTSTA). This paper presents the core concepts of the LTSTA and demonstrates its application through a case study of US Gross Domestic Product (GDP) data. To evaluate the forecasting performance of the proposed model, we compare it against several well-established methods using four evaluation metrics across 72 time series from a forecasting competition. The paper concludes with a discussion of the empirical findings and their implications. 2 Linear Trend-Stationary Trigonometric ARMA 2.1 Nonstationary Time Series Before introducing the proposed model, we review the foundational concepts and key characteristics of modelling nonstationary time series. Formally, a time series \{Y_{t}\} is strictly stationary if the joint probability distribution of any finite subset of observations is invariant under time shifts. This ensures that the statistical properties of the time series remain constant over time, enabling us to describe its behaviour. However, this notion of strict stationarity is too strong and rarely encountered in real-world applications. A more practical approach is to assume weak stationarity, where: • The mean \mathbb{E}[Y_{t}]=\mu is constant over time, • The variance \mathrm{Var}(Y_{t})=\sigma^{2} is constant over time, and • The autocovariance \mathrm{Cov}(Y_{t},Y_{t-k})=\gamma(k) depends only on the lag k . One of the reasons behind the assumption of weak stationarity is that if we want to construct the best linear predictor (in the mean squared error sense) for the strictly stationary time series, we only need to know the mean and autocovariance function without more detailed knowledge of the joint distribution ( Brockwell and Davis, 2016 ) . This principle is complemented by Wold’s Representation Theorem ( Wold, 1938 ) , which provides a complete characterization of weakly stationary processes through their infinite moving average representation. Let \{Y_{t}\} be a weakly stationary process with \mathbb{E}[Y_{t}]=0 and covariance function, \gamma(j)=\mathbb{E}[Y_{t}Y_{t-j}] , \forall j . Then, \{Y_{t}\} admits the decomposition: Y_{t}=\sum_{j=0}^{\infty}\psi_{j}\varepsilon_{t-j}+\eta_{t} (1) where: 1. \{\varepsilon_{t}\} is a white noise process satisfying \mathbb{E}[\varepsilon_{t}]=0 , \mathrm{Var}(\varepsilon_{t})=\sigma^{2}_{\varepsilon} , \mathrm{Cov}(\varepsilon_{t},\varepsilon_{s})=0 for t\neq s , 2. The coefficients \{\psi_{j}\} are square-summable \sum_{j=0}^{\infty}\psi_{j}^{2}<\infty with \psi_{0}=1 , 3. \eta_{t} is deterministic (perfectly predictable from its past), 4. \{\varepsilon_{t}\} and \{\eta_{t}\} are uncorrelated. Under suitable conditions, a weakly stationary process can be represented in the infinite autoregressive form, meaning that it can be written as a weighted sum of past observations plus white noise: Y_{t}=\sum_{j=0}^{\infty}\pi_{j}Y_{t-j}+\varepsilon_{t} (2) For convenience, the term stationarity refers specifically to weak stationarity unless otherwise stated. Although the representations ( 1 ) and ( 2 ) of stationary time series are elegant and describe the historical data well, they are often impractical as they require estimating an infinite number of parameters. Thus, a parsimonious representation of stationary time series using finite autoregressive and moving average components is preferred ( Box et al., 2015 ) . A process \{Y_{t}\} is said to be ARMA (p,q) with non-zero mean \mu if it can be defined as: Y_{t}-\mu=\sum_{i=1}^{p}\phi_{i}(Y_{t-i}-\mu)+\varepsilon_{t}+\sum_{j=1}^{q}\theta_{j}\varepsilon_{t-j} (3) or \phi(B)(Y_{t}-\mu)=\theta(B)\varepsilon_{t} where \{\phi_{1},\dots,\phi_{p}\} are the autoregressive coefficients with \phi_{p}\neq 0 ; \{\theta_{1},\dots,\theta_{q}\} are the moving average coefficients with \theta_{q}\neq 0 ; B is the backward shift operator such that B^{j}(X_{t})=X_{t-j} ; \phi(B)=1-\phi_{1}B-\dots-\phi_{p}B^{p} and \theta(B)=1-\theta_{1}B-\dots-\theta_{q}B^{q} are autoregressive and moving average operators respectively. An ARMA (p,q) process must satisfy causality and invertibility conditions to ensure it is well-defined and useful for inference. Causality means Y_{t} depends only on past/present white noise, while invertibility allows \varepsilon_{t} to be expressed via past observations. Mathematically, these require the roots of \phi(z) and \theta(z) to lie outside the unit circle |z|>1 , respectively. Additionally, if an ARMA process is causal, it is automatically stationary, as the conditions of stationarity and causality are identical. While ARMA processes are suited for stationary time series, they can be extended to describe the nonstationary behaviour observed in

Read the complete text