Algorithms for multi-horizon forecasting of industrial production

The subject of the study is multi-horizon forecasting of the annual growth rate of the U.S. industrial production index over horizons from 1 to 12 months. This indicator is treated as a measure of current real economic activity and the phase of the business cycle. Forecasting is complicated by the limited length of monthly macroeconomic samples, strong correlation among macroeconomic and financial variables, and instability of relationships between predictors and output over time. After constructing lags, impulse variables, and rolling characteristics, the regressor space becomes high-dimensional, which requires stable estimation methods. The paper focuses on selecting a model that reduces forecast error relative to AR12 while preserving reproducibility and economic interpretability. The reproducibility of the computational procedure and results is also taken into account. The methodology is based on a direct multi-horizon scheme, a pseudo-real-time experiment with an expanding training window, time-series cross-validation, RidgeCV, LassoCV, ElasticNetCV, and a PCA-based factor model. The scientific novelty lies in testing a unified reproducible forecasting scheme in which AR12, RidgeCV, LassoCV, ElasticNetCV, and a PCA-based factor model are compared on the same sample, identical time-series folds, and a common feature panel. This design separates the effect of Ridge regularization from L1-based feature selection, mixed regularization, and factor-based dimensionality reduction. RidgeCV reduces RMSE relative to AR12 at the 1-month horizon from 1.77 to 1.28, or by about 27.7%. At the 2-month horizon, LassoCV performs best: 1.67 compared with 2.18 for AR12. At the 3-month horizon, RidgeCV and LassoCV produce the same RMSE of 2.07. At the 12-month horizon, RidgeCV underperforms AR12: 6.23 against 4.56, whereas LassoCV reaches 4.20. Thus, RidgeCV should be interpreted as an interpretable short-term specification, while LassoCV and ElasticNetCV are more robust alternatives at several medium and long horizons.