Nonlinear Probabilistic Forecast Reconciliation

1. Problem Statement

Forecast reconciliation is the process of adjusting independently generated forecasts (base forecasts) so that they satisfy known aggregation constraints, ensuring "coherence." While probabilistic reconciliation for linear constraints (e.g., sums) is well-established, many real-world forecasting problems involve nonlinear constraints.

Examples of such nonlinear relationships include:

• Ratios: Immigration rates (immigrants/population) or tourism rates (regional tourists/total tourists).
• Demographic Ratios: Mortality or unemployment rates.
• Geometric/Trigonometric relationships: e.g., $Z = X/Y$ or $Z = \sin(X) + \cos(Y)$.

Existing methods for nonlinear reconciliation are limited to point forecasts (single values). This paper addresses the gap by tackling probabilistic forecast reconciliation under nonlinear constraints, aiming to produce coherent probability distributions rather than just point estimates.

2. Methodology

The authors propose two distinct approaches to extend probabilistic reconciliation to nonlinear manifolds. They define the system where free time series $B_t$ map to constrained time series $U_t$ via a nonlinear function $f_u(B_t)$. The full vector $Y_t$ lies on a nonlinear coherent manifold $\mathcal{M}$.

A. Reconciliation via Projection

This approach generalizes the linear "push-forward" measure concept to nonlinear manifolds.

• Mechanism: It maps samples from the base forecast distribution (incoherent) to the nearest point on the nonlinear coherent manifold $\mathcal{M}$.
• Mathematical Formulation: Given a base sample $\hat{y}$, the reconciled sample $\tilde{y}$ is found by solving an optimization problem:
  $$\psi(\hat{y}) = \arg \min_{x \in \mathcal{M}} \| \hat{y} - x \|_W$$
  where $\| \cdot \|_W$ is a weighted norm defined by a positive definite matrix $W$.
• Implementation: Since $\mathcal{M}$ is nonlinear, the map $\psi$ has no closed-form solution. The authors use the Newton-Raphson algorithm to numerically solve the optimization for each sample.
• Weighting Schemes: Three variants are tested:
  1. OLS: Identity matrix (Euclidean distance).
  2. WLS: Diagonal matrix of error variances (accounts for scale).
  3. FULL: Full covariance matrix of errors (accounts for correlation and heteroscedasticity).
• Limitation: Computationally expensive as it requires iterative optimization for every sample.

B. Reconciliation via Conditioning (UKF-based)

This approach generalizes the "slicing" of the joint density onto the constraint manifold.

• Mechanism: It treats the constrained forecasts as noisy observations of the free forecasts and updates the distribution of the free series based on these constraints.
• Algorithm: The authors propose an algorithm inspired by the Unscented Kalman Filter (UKF).
  1. Assumption: The base forecast distribution of the free series is approximated as Gaussian.
  2. Unscented Transform: Instead of linearizing the function (as in Extended Kalman Filters), the method generates a set of sigma points representing the mean and covariance of the free series.
  3. Propagation: These sigma points are passed through the nonlinear function $f_u$ to estimate the mean and covariance of the constrained series.
  4. Update: A Kalman-style update is performed to condition the free series distribution on the observed (or base forecasted) constrained values, yielding a reconciled Gaussian distribution for the free series.
  5. Sampling: Samples are drawn from this reconciled Gaussian and mapped back through the full function $f$ to generate coherent samples for all series.
• Advantage: Avoids iterative optimization per sample; relies on deterministic sigma point propagation.

3. Key Contributions

1. First Probabilistic Nonlinear Framework: This is the first work to address probabilistic forecast reconciliation specifically for nonlinear constraints.
2. Methodological Extensions: Successfully extends both the Projection and Conditioning paradigms from linear subspaces to nonlinear manifolds.
3. UKF Adaptation: Introduces a novel application of the Unscented Kalman Filter not as a dynamical filter, but as a numerical tool for approximating conditional distributions under nonlinear constraints.
4. Comprehensive Evaluation: Validates methods on synthetic data (paraboloid, saddle, ripples surfaces) and two real-world case studies involving Swiss socio-demographic ratios and Australian tourism rates.

4. Experimental Results

The methods were evaluated using the Energy Score (for joint distributions) and CRPS (for marginal distributions) against a baseline (independent base forecasts) and a Probabilistic Bottom-Up (PBU) approach.

• Accuracy:
  • Both reconciliation methods generally outperform the base forecasts and PBU.
  • The UKF-based conditioning approach consistently achieved the best overall performance, slightly outperforming the projection methods in terms of accuracy (lower Energy Score and CRPS).
  • Among projection methods, WLS and FULL generally outperformed OLS, confirming that accounting for scale and correlation is crucial.
• Computational Efficiency:
  • The UKF approach is substantially faster than the projection approach.
  • In experiments, UKF was 3 to 4 orders of magnitude faster than the projection method (e.g., milliseconds vs. seconds for 1000 samples). This is because UKF avoids the heavy numerical optimization required for every sample in the projection method.
• Case Studies:
  • Swiss Demographics: UKF improved base forecasts by ~9.4% (Immigration) and ~3.7% (Citizenship).
  • Australian Tourism: UKF was the only method to improve upon the base forecast (2.6% improvement), while projection methods failed to beat the baseline in this specific ratio scenario.

5. Significance and Limitations

Significance:

• The paper provides a practical toolkit for handling complex, real-world forecasting hierarchies where simple summation constraints do not apply (e.g., rates, ratios, percentages).
• It demonstrates that conditioning via UKF offers a superior trade-off between accuracy and computational cost, making it viable for real-time or large-scale applications where projection methods would be too slow.

Limitations:

• Gaussian Assumption: The UKF method assumes the underlying free-level time series follow a Gaussian distribution. It may struggle with highly skewed, intermittent, or discrete data.
• Discrete/Mixed Data: The projection method requires ad-hoc modifications for discrete series, and the current UKF approach is unsuitable for them.
• Scoring Rules: The authors note that standard Euclidean-based scores (Energy Score, CRPS) might be misleading on nonlinear manifolds. They suggest future work on geodesic-distance-based scoring rules that are more appropriate for the geometry of the manifold.
• Optimality Guarantees: While linear reconciliation has theoretical optimality guarantees (MinT), the probabilistic guarantees for the nonlinear case remain empirical and are not yet fully derived.

Conclusion

The paper successfully bridges the gap between linear probabilistic reconciliation and nonlinear constraints. It establishes that reconciliation via conditioning using the Unscented Kalman Filter is the most effective approach, offering state-of-the-art accuracy with significantly reduced computational overhead compared to projection-based methods.