Abstract
Accurate forecasting of the U.K. gross value added (GVA) is fundamental for measuring the growth of the U.K. 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.
Introduction
Traditional time series forecasting methods, such as the Box-Jenkins ARIMA models, typically rely on the assumption of stationarity—the premise that statistical properties like mean and variance remain constant over time. However, real-world economic data frequently violates this assumption. The U.K. Gross Value Added (GVA), specifically the ABML series, exhibits significant nonstationarity characterized by long-term trends and evolving variance.
The ABML Dataset and Nonstationarity
The ABML series consists of quarterly observations of the U.K. GVA from 1955 to 2020. As shown in Figure 1, the series displays a clear upward trend with a dramatic collapse during the 2020 COVID-19 lockdowns.
Standard techniques often involve differencing the data to achieve stationarity. However, for the ABML series, even second differences fail to produce a stationary process. As illustrated in Figure 2, the variance of the second differences increases over time (a), and the magnitude of the 2020 shock (b) is unprecedented, creating a structural break that traditional models struggle to accommodate.
The FORLAP Methodology
To address these challenges, the researchers developed FORLAP (Local Partial Autocorrelation-based Forecasting). This method automates the Locally Stationary Wavelet (LSW) framework, which models a series as a collection of localized structures that evolve over time.
Key innovations in FORLAP include:
- Automated Window Selection: Using the local partial autocorrelation function (lpacf), FORLAP dynamically determines the optimal number of recent observations ($p$) relevant for the next forecast at each time step.
- Automatic Smoothing: It employs advanced spectral estimation to automatically select bandwidths for wavelet periodograms, eliminating the need for manual parameter tuning.
- Regularization: Optional covariance matrix regularization prevents forecast instability during periods of high volatility or sparse data.
Key Results
The performance of FORLAP was evaluated against standard Box-Jenkins models using both simulated data and the real-world ABML series.
- Simulation Benchmarks: FORLAP outperformed the Box-Jenkins baseline in two-thirds of nonstationary test cases and remained competitive even on stationary data.
- Economic Forecasting (ABML): During the volatile pandemic period, FORLAP achieved a 90% success rate in maintaining the true GVA value within its 95% prediction interval. In contrast, Box-Jenkins models only managed a 66-72% success rate, often failing to capture the scale of the impact.
- Interval Calibration: Using the Mean Interval Score (MIS), the authors demonstrated that FORLAP produces better-calibrated intervals that adapt intelligently to variance changes rather than simply widening them.
Conclusion
FORLAP represents a significant advancement in automated nonstationary forecasting. By replacing subjective manual tuning with statistically grounded estimators like lpacf, it provides a reproducible and robust tool for economic analysis. While future work will explore longer forecasting horizons and further theoretical consistency, FORLAP's ability to handle extreme "black swan" events like the COVID-19 pandemic makes it a valuable asset for modern econometrics.
Reference: Killick, R., et al. (2023). Automatic Locally Stationary Time Series Forecasting with application to predicting U.K. Gross Value Added Time Series under sudden shocks caused by the COVID pandemic. arXiv:2303.07772v1.
Summary updated