Global Persistence, Local Residual Structure: Forecasting Heterogeneous Investment Panels

By Oleg Roshka

Rating

1505
Battle Count: 65

Relevance

6/10
The paper addresses cross-sectional investment forecasting on heterogeneous panels mixing macro, institutional, and firm-level data. While the prediction target is CapEx/Assets percentile ranks (not returns), the architecture provides improved rank forecasts that could inform sector rotation, factor timing, and cross-sectional stock selection. The tech/health block shows the largest gain (R²=0.808 vs 0.681 global), suggesting practical value for within-sector selection. However, portfolio-level evidence is preliminary and not statistically significant (IR=+0.47, p=0.14). The method is more directly relevant to macro/fundamental quantitative strategies than high-frequency trading. The cross-regime replication (US+UK/EU) and robustness to filing delays enhance practical applicability.

Implementation Complexity

6/10
The two-stage architecture is conceptually straightforward (pooled AR(1) + block-specific PCA+ridge), but practical implementation requires: (1) careful block assignment based on economic metadata, (2) quarterly re-estimation of all models, (3) exponentially-weighted demeaning of residuals, (4) ridge regularization tuning via cross-validation, (5) spectral radius clipping for DMD variants, (6) proper rolling-window evaluation protocol with expanding training sets, (7) multiple inference corrections (Holm-Bonferroni, DM-HAC, bootstrap). The Kalman filter variant adds state-space complexity. The method-equivalence finding (PCA≈DMD≈Ridge) simplifies engine choice. Main computational burden is the 1,000-permutation placebo test.

Reproducibility

5/5
Replication code, pseudonymised panel data, and committed hyperparameter predictions are available at https://anonymous.4open.science/r/harp-reproduction. Scripts reproduce all tables, figures, placebo permutations, UK/EU extension, and combined-panel analyses. Full reproduction takes approximately 15 minutes on a single CPU. All hyperparameters are inventoried in Table 17. Pre-registered analysis plan for UK/EU extension included in replication archive.

About this paper

Methodology: Heterogeneity-Aware Two-Stage Mixture Architecture. Problem types: Time Series Forecasting, Regression, Dimensionality Reduction, Panel Data Estimation, Cross-sectional Prediction.

The interactive Everscope explorer (charts, battles, favorites) loads below.