Many forecasting problems come as a hierarchy. Item sales add up to store sales, and store sales add up to state sales. Electricity feeders add up to substations, which add up to a grid total. Forecasters predict every level and then reconcile them: they adjust the numbers so that children sum to parents, deciding how much to trust each level. The standard recipe, minimum-trace reconciliation (MinT), does this with one estimate of how forecast errors vary and co-vary across the hierarchy. That single estimate is used whether the forecast is for tomorrow or for six months out. The forecasting literature, however, shows that errors grow and change shape as the horizon lengthens: the future drifts away from the past, and multi-step errors compound. Our claim is that reconciliation has to follow that drift. Bottom-level detail is trustworthy at short lead times. Smoother aggregates should take over at long ones. Using one setting for every lead time sends the wrong level's errors down or up the hierarchy.
We propose horizon-conditioned reconciliation. The error covariance, and the robustness margin around it, become smooth functions of the lead time $h$, estimated jointly across all lead times. Pooling matters because long-lead residuals are scarce and overlapping. We test this on the hierarchical benchmarks used by prior reconciliation work, with results reported separately for each horizon. We also use the method to probe a puzzling null result: adding temporal aggregation levels to a hierarchy gave no clear benefit (Hierarchical Forecasting at Scale). We conjecture that the benefit is real at long horizons and harmful at short ones, so it cancels out when everything is averaged together.
Core Claim
MinT reconciles base forecasts $\hat y_h$ for all nodes at lead time $h$ into coherent forecasts
$$\tilde y_h = S\,(S^\top W_h^{-1} S)^{-1} S^\top W_h^{-1}\,\hat y_h ,$$
where $S$ is the fixed matrix that sums bottom-level series into every node, and $W_h$ is the covariance of base-forecast errors at lead $h$. In practice, people either reuse one-step residuals for every $h$ or estimate each $W_h$ separately from very few residuals. Hierarchical Time Series Forecasting with Robust Reconciliation showed that noise in the estimate of $W$ can make MinT worse than not reconciling at all. That paper fixes the problem by placing a bootstrap box around $W^{-1}$ and solving a min-max semidefinite program. Its uncertainty set is static across lead times, and it is solved by a generic solver that does not scale beyond a few hundred series.
Our hypothesis has three parts. First, the optimal reconciliation weights shift monotonically toward upper levels as $h$ grows. Second, the estimation noise in $W_h$ grows with $h$. Third, a low-dimensional, horizon-smooth parametrization captures both effects better than either a static estimate or independent per-horizon estimates. Concretely, we write
$$W_h = D_h^{1/2}\,\big[(1-\lambda_h) R_0 + \lambda_h R_\infty\big]\,D_h^{1/2},$$
where $D_h$ is a diagonal matrix of per-series error variances, each fit as a smooth increasing curve in $h$. $R_0$ and $R_\infty$ are short- and long-lead correlation matrices, shrunk toward a diagonal-plus-low-rank structure. $\lambda_h \in [0,1]$ is a monotone mixing curve. The robustness radius $\rho_h$ is likewise a monotone function of $h$, calibrated by a block bootstrap that respects the overlap of $h$-step residuals.
Research Program
The work has three parts. The first is diagnostic. On each benchmark, we fit "oracle" per-horizon reconciliation weights on held-out data. We then measure how the share of weight placed on upper-level forecasts changes with $h$, and how fast the error variance and correlation change at each level. This tests whether the problem exists. If the oracle weights are flat in $h$, the proposal is unnecessary and we will report that.
The second part is the estimator. We fit $D_h$, $\lambda_h$, $R_0$, $R_\infty$ and $\rho_h$ by penalized likelihood on rolling-origin residuals, pooling across all lead times. Pooling is what separates this from per-horizon MinT: long-lead covariances borrow strength from short-lead ones through the shared correlation structure and the smoothness constraints. The robust variant reuses the min-max formulation of Aikawa et al. with a horizon-indexed radius. For large hierarchies, we replace their semidefinite program with the closed-form MinT solution under the structured $W_h$, and we inflate the diagonal by $\rho_h$, which acts as an approximate worst case. The point of this replacement is to remove the scaling limitation that paper reports.
The third part extends the hierarchy along time. We add temporal aggregation levels, such as daily to weekly on M5, as extra rows of $S$. We then evaluate stratified by horizon to test whether the null result of Sprangers et al. hides a short-versus-long sign flip.
Uncertainty intervals are a secondary output. Transformer Conformal Prediction for Time Series reports that coverage of time-series conformal intervals falls below nominal beyond two steps. We will therefore report per-horizon empirical coverage of Gaussian intervals built from $\tilde W_h$, the reconciled covariance. We report it only as a diagnostic, not as a guarantee.
Baselines
The naive reference points are the unreconciled base forecasts and bottom-up aggregation. Every method shares identical base forecasters, so differences come only from reconciliation.
The first standard competitor is MinT with a static shrinkage covariance estimated from one-step residuals. This is the common practice. Aikawa et al. report a total-level Relative MAE of 0.966 for it on U.S. Walmart Sales.
The strongest existing small-hierarchy method is robust min-max reconciliation (Hierarchical Time Series Forecasting with Robust Reconciliation). It reports a total-level Relative MAE of 0.810 on Walmart, and the paper states only qualitatively that it beats the classical baselines on Australian Tourism, Births and Swiss Electricity, with no numbers in the file for those datasets. We rerun it with the authors' bootstrap settings on our rolling origins. We count our method as beating it if it achieves lower Relative MAE averaged over all levels and horizons, and in particular at the longest third of horizons, without losing at short horizons.
The strongest large-scale competitor is the sparse hierarchical loss of Hierarchical Forecasting at Scale. It builds coherence into training rather than adjusting forecasts afterwards, and it reports RMSE and MAE gains of up to 10% on M5. We train it with the same backbone and tuning budget, and we also test a horizon-weighted variant of it.
The most important comparison is per-horizon MinT: separate shrinkage $W_h$ for each $h$, with no pooling. It isolates whether our gains come from pooling and smoothness or merely from acknowledging that $W$ depends on $h$. Two further ablations remove one component each: a fixed $\lambda$, and a fixed $\rho$. All methods get the same rolling-origin windows and the same hyperparameter search budget, and report wall-clock time.
Why not just…?
Why not just estimate a separate MinT covariance at every horizon? This is the real threat. If residuals are plentiful, it may match our method, and the contribution would shrink to the diagnostic finding. Our bet is that it fails exactly where samples are scarce: quarterly tourism, monthly births, and long leads, where $h$-step residuals overlap heavily. The per-horizon MinT ablation decides this directly. We will report the result, including the sample sizes at which the two methods converge.
Why not just use robust reconciliation with a wider static margin? A single margin wide enough to protect long horizons is too conservative at short ones. Aikawa et al. themselves note over-conservatism that raises bottom-level error when the covariance estimate is already accurate. The fixed-$\rho$ ablation, evaluated per horizon, shows whether a horizon-indexed margin removes that trade-off.
Why not just build coherence into training, as Sprangers et al. do? Their loss weights all lead times equally and found temporal levels unhelpful. We run it both as published and with horizon weighting. If the horizon-weighted training loss closes the gap, that supports our core claim while favoring their mechanism, and we would say so.
Why not just fix the base forecaster's distribution shift, for example with Dish-TS? Dish-TS predicts the horizon's level and scale, and it reports more than 10% improvement over RevIN on long-term forecasting. It does improve the base forecasts, but it says nothing about how errors at different levels should be weighed against each other. We will use Dish-TS-normalized base forecasters as one configuration to check that our gains persist on top of it.
Proposed Experiments
We use the benchmarks from the file: U.S. Walmart Sales (14 series), Australian Tourism (35), Australian Births (10) and Swiss Electricity (31) as small hierarchies where the robust semidefinite program is feasible, and M5 (3,049 products, 12 levels) for scale. Metrics are Relative MAE and RMSE per level and per horizon bucket, plus per-horizon interval coverage and wall-clock time.
Four checks decide the outcome. First, the oracle upper-level weight share must increase with $h$ on at least three of the five datasets; otherwise the premise fails. Second, the pooled estimator must beat per-horizon MinT at long horizons on the short-history datasets; otherwise pooling is not the contribution. Third, on Walmart, total-level Relative MAE must improve on the 0.810 reported for robust reconciliation, with the improvement concentrated at long leads. Fourth, on M5 with temporal levels added, the horizon-stratified evaluation must show temporal aggregation helping at long horizons and hurting or neutral at short ones. Such a sign flip would explain the null result of Sprangers et al. rather than contradict it.