A1622
Title: Non-stationary challenges in partial least squares: Applications to forecasting
Authors: Shou-Yung Yin - National Taipei University (Taiwan) [presenting]
Abstract: The behavior of Partial Least Squares (PLS) is examined when both the target variable and predictors are non-stationary. It is shown theoretically that PLS may mechanically remove low-frequency stochastic trends in sample, so that the residual appears stationary even when regressors are structurally irrelevant. This filtering effect does not carry over out of sample: prediction errors continue to display non-stationary behavior. To clarify this contrast, an operator-based characterization of PLS is developed and connected to the Karhunen-Loeve expansion of non-stationary processes. Building on this distinction, a backward-greedy variable selection procedure is proposed, guided by a Partial-sum energy criterion, which measures the squared cumulative sum of out-of-sample residuals. Unlike standard unit-root test statistics, which exhibit a cliff-shaped landscape that defeats conventional search, this criterion provides a smooth signal for subset quality. Monte Carlo experiments confirm non-trivial power to recover the true cointegrating variables, and an empirical application to FRED-MD data shows that out-of-sample stationarity criteria yield predictor sets that differ materially from those selected by in-sample fit.
