Multi-output regression requires prediction of several continuous outcomes from the same predic-
tor set. Independent per-target models are simple to train, but they may leave residual association between
targets unused. We propose a residual-polishing extension of per-target Random Forest regression that
preserves the standard model-per-target workflow while adding a post-hoc correction based on out-of-bag
residuals, local residual averaging, and residual-covariance structure. The method is evaluated using con-
trolled synthetic scenarios and real multi-output datasets. The evaluation combines real-data benchmark
comparisons with repeated train/validation/test experiments, including confidence intervals, paired statistical
tests, and comparisons with several multi-output baselines. The results show that residual polishing can im-
prove the independent Random Forest baseline when recoverable local and cross-target residual structure
remains after marginal fitting

Multi-output regression requires prediction of several continuous outcomes from the same predic-
tor set. Independent per-target models are simple to train, but they may leave residual association between
targets unused. We propose a residual-polishing extension of per-target Random Forest regression that
preserves the standard model-per-target workflow while adding a post-hoc correction based on out-of-bag
residuals, local residual averaging, and residual-covariance structure. The method is evaluated using con-
trolled synthetic scenarios and real multi-output datasets. The evaluation combines real-data benchmark
comparisons with repeated train/validation/test experiments, including confidence intervals, paired statistical
tests, and comparisons with several multi-output baselines. The results show that residual polishing can im-
prove the independent Random Forest baseline when recoverable local and cross-target residual structure
remains after marginal fitting