Automatic locally stationary time series forecasting with application to predicting UK gross value added time series

Rebecca Killick, Marina I. Knight, Guy P. Nason, Matthew A. Nunes, Idris A. EckleyJournal of the Royal Statistical Society Series C (Applied Statistics) 2024 · DOI

TL;DR

This paper introduces FORLAP, a novel automated forecasting method for locally stationary time series. FORLAP leverages a new local partial autocorrelation function (lpacf) to dynamically select the optimal window of past data for forecasting, outperforming existing methods, including Box-Jenkins, on both simulated and real-world economic data, even during periods of high volatility like the COVID-19 pandemic.

What Problem Does This Paper Address?

Forecasting economic time series, such as the UK's Gross Value Added (GVA), is crucial for understanding economic growth. However, these series often exhibit nonstationarity, meaning their statistical properties (like variance) change over time. Traditional forecasting methods, like Box-Jenkins, assume stationarity and can perform poorly when this assumption is violated. While methods for nonstationary series exist, they can be computationally intensive or require manual parameter tuning. This paper addresses the need for a practical, automated, and accurate forecasting method for nonstationary time series, specifically tackling the challenge of sudden economic shocks.

Methods

The core of this paper is the development of the FORLAP (Forecasting using Local Partial Autocorrelation) algorithm. FORLAP builds upon the Locally Stationary Wavelet (LSW) framework, which models time series whose statistical properties evolve over time.

Here's a breakdown of the key methodological contributions:

  • Local Partial Autocorrelation Function (lpacf): The authors introduce a new way to estimate the local partial autocorrelation function. This function helps identify how much recent past data is relevant for forecasting at a specific point in time, effectively determining a "forecasting window." This is a significant departure from previous LSW methods that relied on grid searches for this parameter.
  • Automated Parameter Selection: FORLAP automates the selection of the forecasting window size (denoted by p). By using the lpacf, it dynamically determines the optimal number of past observations to consider, eliminating the need for manual tuning or computationally expensive grid searches.
  • Leveraging LSW Framework: FORLAP operates within the LSW framework, which is well-suited for time series with time-varying statistical properties. It uses spectral estimation techniques to understand the local structure of the time series.
  • Modified Forecasting Procedure: The algorithm combines the automated p selection with existing LSW forecasting techniques, aiming to provide more accurate and reliable predictions.

Datasets

The primary dataset used for evaluation is the UK Gross Value Added (GVA) time series, specifically the ABML series, which is a key component of the UK's Gross Domestic Product (GDP). This series spans from quarter one 1955 to quarter four 2020 and exhibits significant volatility, particularly during the 2008 financial crisis and the COVID-19 pandemic.

In addition to the real-world GVA data, the authors also conducted extensive simulations using various stationary and nonstationary time series models. These simulated datasets were crucial for rigorously evaluating FORLAP's performance against other forecasting methods under controlled conditions.

Key Insights and Ideas

  • The Power of Local Structure: The paper highlights that understanding the local statistical structure of a time series is paramount for accurate forecasting, especially in the presence of nonstationarity.
  • Dynamic Windowing is Key: Forcing a fixed window of past data or relying on slow-adapting methods can lead to suboptimal forecasts. FORLAP's ability to dynamically adjust the forecasting window based on the local data structure is a major advantage.
  • Nonstationarity is the Norm, Not the Exception: The study reinforces that many real-world economic time series are inherently nonstationary, and methods that ignore this fact are likely to underperform.
  • COVID-19 as a Stress Test: The COVID-19 pandemic provided a real-world scenario of extreme volatility, serving as a rigorous test for forecasting methods. FORLAP's ability to maintain accuracy during this period is a strong testament to its robustness.
  • Beyond Stationarity: While Box-Jenkins is a benchmark for stationary series, FORLAP demonstrates competitive performance even on stationary data, suggesting its flexibility.

Key Results

  • Superior Performance on GVA Data: FORLAP significantly outperforms the traditional Box-Jenkins method for forecasting the UK GVA series, especially during periods of high volatility like the COVID-19 pandemic.
  • Improved Prediction Intervals: FORLAP consistently provides better-calibrated prediction intervals compared to other nonstationary methods, balancing interval width with coverage accuracy.
  • Simulation Study Success: In simulations, FORLAP outperformed Box-Jenkins on two-thirds of the tested scenarios and was competitive or superior to other nonstationary methods across the majority of models.
  • Robustness to Volatility: FORLAP showed strong performance even when faced with significant changes in variance and dynamics, as observed during the COVID-19 pandemic.
  • Automated Efficiency: The automated selection of the forecasting window (p) via lpacf makes FORLAP more practical and less computationally demanding than previous LSW forecasting approaches.

Limitations and Future Work

  • Bootstrap for Confidence Intervals: The authors suggest that a bootstrap test could be an alternative to asymptotic-based confidence intervals for selecting p.
  • Theoretical Understanding of p: Further theoretical work is needed to understand how well the estimated p reflects the true autoregressive order if the underlying process were indeed autoregressive.
  • Comparison with Other Nonstationary Methods: While FORLAP is compared to several methods, the authors note that code for some nonstationary forecasting techniques is not freely available, limiting direct comparisons.
  • Forecast Horizons: The current study primarily focuses on one-step-ahead forecasts (h=1). Future research could explore FORLAP's performance at longer forecast horizons.
  • Series Types and Availability: Further research is needed to understand FORLAP's performance on different types of nonstationary series and to ensure its availability for a wider range of users. The FORLAP forecasting method is available as an R package on CRAN.

Summary updated

Abstract

Abstract Accurate forecasting of the UK gross value added (GVA) is fundamental for measuring the growth of the UK economy. A common nonstationarity in GVA data, such as the ABML series, is its increase in variance over time due to inflation. Transformed or inflation-adjusted series can still be challenging for classical stationarity-assuming forecasters. We adopt a different approach that works directly with the GVA series by advancing recent forecasting methods for locally stationary time series. Our approach results in more accurate and reliable forecasts, and continues to work well even when the ABML series becomes highly variable during the COVID pandemic.
BibTeX
@article{killick2024automatic,
  author = {Rebecca Killick and Marina I. Knight and Guy P. Nason and Matthew A. Nunes and Idris A. Eckley},
  title = {{Automatic locally stationary time series forecasting with application to predicting UK gross value added time series}},
  journal = {Journal of the Royal Statistical Society Series C (Applied Statistics)},
  year = {2024},
  doi = {10.1093/jrsssc/qlae043},
}
Figure