Local linear forests and foreign data sharpen French wage forecasts
Local linear forests reduce French wage forecasting errors by nearly half when combined with cross-border data. European Central Bank researchers Michele Lenza and Claudia Marchini show that the technique outperforms standard benchmarks over one- to five-quarter horizons.
Adaptive weights and smooth signals
Michele Lenza and Claudia Marchini evaluate quarterly French negotiated wage growth from 1985Q1 to 2025Q1 across horizons of one to five quarters ahead.
The authors compare local linear forests with random forests, ridge regressions and a random walk benchmark.
The baseline local linear forest specification with adaptive weights across 500 regression trees achieves the lowest prediction errors.
At the five-quarter horizon, the model reaches a relative mean squared error of 0.5302 against the random walk, outperforming standard ridge regression at 0.5988 and plain random forests at 0.6769. The approach handles non-linearities and extrapolates beyond historical ranges by fitting weighted local linear models.
Signals beyond French borders
Adding macroeconomic variables from Germany and Italy significantly improves forecast accuracy across all horizons.
The full dataset incorporates cross-border measures of output gaps, productivity, capacity utilisation and wage indices.
While European labor markets remain institutionally segmented, real and nominal economic trends exhibit strong common dynamics.
The local linear framework proves particularly advantageous during the post-crisis era when wage growth dropped to historically low levels.
National borders mislead wage models
The paper demonstrates that local linear forests capture wage non-linearities without overfitting.
Disregarding cross-border data remains a blind spot for models focused solely on domestic indicators.
Central bank forecasters should adopt these flexible tools across the euro area.
Source: Forecasting wages with local linear forests
IN: