Forecast Reconciliation: Methods & Extensions

• Forecast Reconciliation is the process of adjusting incoherent forecasts to meet structural aggregation constraints in hierarchies and networks.
• It formulates the adjustment as an optimization problem, using methods such as MinT, FlowRec, and NLCR to project forecasts onto coherent subspaces.
• Recent advances improve scalability, robustness under covariance uncertainty, and accommodate nonlinear relationships for practical, large-scale applications.

Forecast reconciliation is the process of adjusting a set of structurally related forecasts—such as those organized in hierarchies, networks, or other systems with aggregation constraints—so that the final reconciled predictions obey those constraints exactly. Classical and recent methodologies frame forecast reconciliation as an optimization problem: given potentially incoherent base forecasts, one seeks coherent forecasts that are optimally close, under a chosen loss or distance metric. The field has advanced from tree-based hierarchical structures and linear constraints to general networks, nonlinear coherence manifolds, massive-scale settings, robust optimization under covariance uncertainty, and machine-learning-based direct mappings.

1. Formal Problem Definition and Mathematical Framework

Let $\mathcal{Y}$ denote the vector of all forecasts across multiple levels, nodes, edges, or paths in a structured system (e.g., regions in a hierarchy, flows in a network, or aggregates over both cross-sectional and temporal axes). The corresponding set of incoherent (base) forecasts is $y$. Reconciled forecasts $x$ must satisfy system-specific aggregation or conservation constraints (e.g., sums across levels, flow conservation, or nonlinear relationships such as rates). Formally, the general reconciliation problem is stated as:

$$\begin{aligned} & \min_{x} \quad L(x, y), \ & \text{subject to} \quad c(x) = 0, \end{aligned}$$

where $L(\cdot, \cdot)$ is a loss function (often squared Mahalanobis distance), and $c(x)$ encodes the system's constraints—linear in classical settings ($Sx=0$), nonlinear in extended frameworks ($g(x)=0$).

For hierarchical or network-structured systems, aggregation constraints are encoded by a binary summing or incidence matrix $S$, giving the coherent subspace $\mathcal{S} = \{\tilde{y} : S\tilde{y} = 0\}$. In the case of networks, node/edge/path coupling is handled by constructing block incidence matrices for vertices and edges ($y$0, $y$1, cf. (Sharma et al., 6 May 2025)).

The standard MinT (Minimum Trace) reconciliation for linear constraints has the closed-form solution:

$y$2

where $y$3 is the constraint matrix and $y$4 is an estimator for the base forecast error covariance matrix. The solution is the orthogonal projection of $y$5 onto the constraint subspace in the $y$6 inner product (Møller et al., 2024, Girolimetto et al., 2023).

In more general settings, reconciliation can require nonlinear optimization (NLCR for continuous constraints (Girolimetto et al., 24 Oct 2025)) or network-flow formulations (FlowRec (Sharma et al., 6 May 2025)), depending on the system's structure and loss.

2. Algorithmic Approaches: FlowRec, MinT, and Extensions

2.1 FlowRec: Network-Flow-Based Reconciliation

FlowRec introduces an optimization-based perspective for hierarchical forecast reconciliation on general networks, going beyond tree structures (Sharma et al., 6 May 2025). Let $y$7 be a directed network. Base forecasts are given for nodes, edges, and paths: $y$8. Reconciled forecasts $y$9 must satisfy

• For every edge $x$0:

$x$1

• For every node $x$2:

$x$3

These constraints are enforced simultaneously by formulating $x$4 for some bottom-level path forecasts $x$5. The flow reconciliation problem becomes a minimum-cost flow problem:

$x$6

where $x$7 or $x$8 give quadratic or linear programs with convexity ensuring global optimum. FlowRec enables convex, strictly differentiable losses and supports arbitrary graph structures, with $x$9 time for sparse graphs, dramatically improving over the $$\begin{aligned} & \min_{x} \quad L(x, y), \ & \text{subject to} \quad c(x) = 0, \end{aligned}$$0 cost of MinT in tree settings. It also supports efficient, error-bounded approximate reconciliation, local dynamic updates, and monotonicity under forecast improvements (Sharma et al., 6 May 2025).

2.2 Comparison: MinT vs. FlowRec

• MinT: Classical optimal-reconciliation method, limited to tree hierarchies and requiring $$\begin{aligned} & \min_{x} \quad L(x, y), \ & \text{subject to} \quad c(x) = 0, \end{aligned}$$1 estimation, with cubic computational complexity. It seeks a weighted projection in the base-forecast error metric.
• FlowRec: Generalizes MinT to non-tree aggregations and networks by removing covariance estimation in favor of direct network structure. Reduces computational cost to $$\begin{aligned} & \min_{x} \quad L(x, y), \ & \text{subject to} \quad c(x) = 0, \end{aligned}$$2 for sparse graphs, and admits fast approximate/dynamic update schemes for practical use at scale (Sharma et al., 6 May 2025).

Empirically, FlowRec matches or exceeds MinT’s accuracy while improving runtime and memory by factors of $$\begin{aligned} & \min_{x} \quad L(x, y), \ & \text{subject to} \quad c(x) = 0, \end{aligned}$$3–$$\begin{aligned} & \min_{x} \quad L(x, y), \ & \text{subject to} \quad c(x) = 0, \end{aligned}$$4 and $$\begin{aligned} & \min_{x} \quad L(x, y), \ & \text{subject to} \quad c(x) = 0, \end{aligned}$$5–$$\begin{aligned} & \min_{x} \quad L(x, y), \ & \text{subject to} \quad c(x) = 0, \end{aligned}$$6 respectively in simulation and benchmark datasets.

3. Extensions: Nonlinear Constraints and Robust Reconciliation

3.1 Nonlinear Forecast Reconciliation (NLCR)

In many applications, relationships among time series are governed by nonlinear constraints (e.g., ratios such as rates, proportions, or multiplicative identities) (Girolimetto et al., 24 Oct 2025). The NLCR method projects base forecasts $$\begin{aligned} & \min_{x} \quad L(x, y), \ & \text{subject to} \quad c(x) = 0, \end{aligned}$$7 onto the nonlinear manifold $$\begin{aligned} & \min_{x} \quad L(x, y), \ & \text{subject to} \quad c(x) = 0, \end{aligned}$$8 by solving:

$$\begin{aligned} & \min_{x} \quad L(x, y), \ & \text{subject to} \quad c(x) = 0, \end{aligned}$$9

with first-order conditions handled via the Lagrangian and solved with Sequential Quadratic Programming (SQP). While general distance reduction does not hold globally in nonlinear settings, the existence of an “improvement ball” or, under convex constraints, a hypograph-based global improvement, is formally established. Empirically, NLCR outperforms linear methods in domains involving ratios, yielding significant reductions in RMSE for demographic and economic series (Girolimetto et al., 24 Oct 2025).

3.2 Robust Reconciliation Under Covariance Uncertainty

MinT’s reliance on estimated forecast-error covariances, often noisy due to finite-sample effects, motivates robust reconciliation strategies (Aikawa et al., 23 Oct 2025). Robust reconciliation models the true covariance as lying within a box-type uncertainty set derived from bootstrap distributions:

$L(\cdot, \cdot)$0

and seeks $L(\cdot, \cdot)$1 minimizing the worst-case expected squared error over all $L(\cdot, \cdot)$2, formulated as a semidefinite program (SDP). This yields a reconciliation solution with formal worst-case guarantees under misspecified or unstable covariance estimates, with demonstrable empirical gains—especially at upper hierarchy levels and in settings with large residual covariance uncertainty (Aikawa et al., 23 Oct 2025).

4. Computational and Practical Aspects

Forecast reconciliation at massive scale (e.g., $L(\cdot, \cdot)$3 series) requires numerically efficient algorithms and careful model-constraint representation (Tianyu et al., 4 Feb 2026). For unconstrained or lightly constrained special cases (e.g., tree with disjoint row supports), share-based (top-down or bottom-up) solutions can be recovered as limiting cases of weighted least squares, and nonnegativity can be enforced automatically. For general overlapping structures, the optimization-based approach (via the KKT system or ADMM) allows reconciliation across multiple, intersecting hierarchies simultaneously (Tianyu et al., 4 Feb 2026).

High-level insights into computational complexity:

Method	General Structure	Complexity (sparse)	Memory
MinT (tree)	Linear (tree)	$L(\cdot, \cdot)$4	$L(\cdot, \cdot)$5
FlowRec	Arbitrary graph	$L(\cdot, \cdot)$6	$L(\cdot, \cdot)$7
Robust SDP	Linear (tree)	$L(\cdot, \cdot)$8 (general)	$L(\cdot, \cdot)$9 (typ.)
Billions-Scale	Overlapping	Iterative, sparse	$c(x)$0

For approximate reconciliation, projected subgradient or multiplicative-weight methods achieve geometric convergence and tight error bounds with substantially reduced computational cost (Sharma et al., 6 May 2025).

5. Empirical Performance and Applications

FlowRec was evaluated on simulated and real datasets (e.g., Australian Tourism, Labor, Traffic, Wiki) and compared to bottom-up and MinT:

• Accuracy: Improvements up to $c(x)$1 RMSE/MAE reduction over base; systematically outperforms both MinT and bottom-up at all hierarchy levels.
• Runtime/memory: $c(x)$2–$c(x)$3 runtime improvement, $c(x)$4–$c(x)$5 lower memory compared to MinT on simulated ARIMA experiments; for dense networks, gains are even greater.
• Real-world benchmarks: On tree hierarchies, FlowRec matches or slightly exceeds MinT, but is markedly faster and uses much less memory.
• Dynamic and approximate updates: Local adjustments and approximate reconciliation yield further computational benefits with provable error guarantees, supporting real-time and large-scale use (Sharma et al., 6 May 2025).

6. Theoretical Insights and Future Directions

• NP-Hardness: Reconciliation under the $c(x)$6 (sparsity-inducing) loss is NP-hard (by reduction from Exact 1-in-3-SAT), ruling out exact algorithms for zero-entry minimization in general (Sharma et al., 6 May 2025).
• Polynomial Solvability: For all convex, strictly differentiable losses ($c(x)$7 norms for $c(x)$8), reconciliation is solvable in polynomial time (strict convexity yields uniqueness).
• Lower Bounds: Any exact reconciliation algorithm under $c(x)$9 admits a runtime lower bound of $Sx=0$0 even for $Sx=0$1-norm losses, achieved by FlowRec for sparse graphs.

Important open directions include further extensions to probabilistic and non-Gaussian reconciliation in high-dimensions, distributed and online optimization strategies for streaming settings, scalable computation for overlapping and general constraint sets, and deeper integration with machine learning–based forecasting architectures.

---

References

• "Hierarchical Forecast Reconciliation on Networks: A Network Flow Optimization Formulation" (Sharma et al., 6 May 2025)
• "Forecast reconciliation with non-linear constraints" (Girolimetto et al., 24 Oct 2025)
• "Hierarchical Time Series Forecasting with Robust Reconciliation" (Aikawa et al., 23 Oct 2025)
• "Billions-Scale Forecast Reconciliation" (Tianyu et al., 4 Feb 2026)