Hierarchical Bayes meets hierarchical forecasting: A flexible framework for level-focused forecasts
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.
Tom: We're looking at the paper: "Hierarchical Bayes meets hierarchical forecasting".
Jane: Here's the gist of the paper. Decision-making in hierarchical systems requires probabilistic forecasts at all cross-sectional levels Our proposed fully Bayesian hierarchical forecasting framework shares information more effectively between and across levels than reconciliation alone.
Tom: Framework Overview The paper proposes a fully Bayesian hierarchical forecasting framework that shares information more effectively between and across levels than reconciliation alone.
Jane: This approach has the flexibility to softly penalise incoherence, subject to model specification, and to focus the global model and coherence update on hierarchical levels most relevant to decision outcomes.
Tom: The framework yields parameter estimates that are focused towards the forecasting goals and capture the requirement for coherency.
Paper summary: Tom: So we're looking at this paper, "Hierarchical Bayes meets hierarchical forecasting: A flexible framework for level-focused forecasts". Basically, it claims that decision making in systems with multiple levels needs probabilistic forecasts at every single level.
Jane: And current methods just make independent forecasts at each level and then try to fix them afterward to make sure they match up. But this new framework shares information better between and across these levels than those post hoc fixes do, and it has the flexibility to softly penalize incoherence depending on what you specify about the model.
Lu: What's really interesting is that this approach lets you focus the global model and that coherence update only on the hierarchical levels that actually matter for your decision outcomes. This means the parameter estimates end up being focused towards those goals, which captures that need for coherency in a way reconciliation alone doesn't.
Meng: So, how does it share information better? Is it some new kind of prior structure or something? What does this mean practically when we're trying to deploy these models at scale?
Tom: The paper sets up a hierarchical prior where you have a hyperprior over the parameters for each series in the hierarchy. This captures the idea that the parameters for each series are related through a population distribution, but they can still differ based on some variance specified by that hyperprior.
Jane: It's designed to mirror that cross-sectional hierarchy you are trying to forecast, so at every node, those child and parent series share a hierarchical hyperprior structure.
Lu: They introduce a coherency penalty where the likelihood is multiplied by p of the data given sigma squared cohIT. This penalizes incoherence between the upper-level series and lower-level series, but it's done distributionally, so incoherence is only softly penalized depending on how you specify your model.
Meng: So, if I'm running this on a real traffic stream, what is the practical impact of this soft penalty? Does it just make the forecast a little less accurate overall?
Tom: It allows for flexibility to balance that coherence against accuracy based on your model specification, which means you can tune how much weight you put on coherence versus raw predictive power.
Jane: Then they propose a level weighting factor, lambda i, to prioritize specific series or levels by multiplying the likelihood by that factor when updating parameters. This extra emphasis carries through to the parent or child nodes while still keeping that soft constraint on coherence with minimal adjustment needed for those weighted series.
Lu: The empirical validation on simulated data showed this hierarchical Bayes approach outperforms MinT reconciliation in terms of overall root mean squared error and error of sum squares when you account for the performance across the whole hierarchy.
Meng: And in real data, they say it clearly beats independent Bayesian forecasts, both with and without that MinT reconciliation applied. But what about those level-weighted approaches? Are they really the best performers at their respective levels?
Tom: The results showed that those level-weighted approaches are the clear best performers when you look at them on their own level. It suggests a strong focus on where the decisions actually happen in the system.
Jane: So, to sum up, this paper shows how to use a fully Bayesian hierarchical forecasting framework that shares info better than reconciliation by having flexibility to softly penalize incoherence and focus the model updates on the most relevant levels for your goals.
Lu: It helps estimate model parameters when they are tough to find from the data or when you already know those series share characteristics, which is a huge thing in complex systems. But they also flag that this method isn't great if your series don't share those characteristics or if you use different mathematical models for different series.
Meng: And what about the focusing technique? The authors admit that it can actually reduce performance if the data or model at an upper level is weak, so you can't just ignore that part of the hierarchy.
Tom: The coherency penalty itself acts as a diagnostic tool for when your hierarchical system might be misspecified, which is useful for understanding why your forecasts aren't performing as expected.
Jane: So, this work gives us a way to think about combining the structure of hierarchy with probabilistic forecasting in a way that prioritizes where the decisions are made while still keeping things coherent. That sets up the next step for how we model complex systems probabilistically.
Conclusion: Tom: So we're wrapping up this deep dive on "Hierarchical Bayes meets hierarchical forecasting: A flexible framework for level-focused forecasts".
Jane: Basically, this paper is about taking those complicated models that look at many levels, like different parts of a system, and using Bayesian methods to share information across all those levels much better than just fixing things after the fact.
Lu: It's a framework that lets you use probabilistic forecasts at every single level instead of just getting independent guesses for each part.
Meng: So, what does this actually mean when you're trying to predict something in a real system? What's the practical application here?
Tom: Well, the main point is that you can focus the model updates only on the levels that actually matter for your final decision outcomes. The parameter estimates end up being pulled toward those goals.
Jane: That flexibility comes from how they handle incoherence—which is when one part of your system doesn't match what another part expects to see. They use a penalty that lets you "softly" penalize that mismatch based on how you set up your model.
Lu: And they also have these level weighting factors, like lambda i, which let you prioritize certain levels by multiplying the likelihood by a number bigger than one. This emphasis stays with the parent or child nodes when updating their parameters.
Meng: But what are the limits? The authors themselves point out that this approach won't work if your different series don't actually share characteristics, or if you use completely different mathematical models for each level.
Tom: Exactly. And they also warn that if the data at a high level is bad, focusing on it can actually make the overall performance worse. It’s not a magic fix for everything.
Jane: Despite those caveats, the results on simulated and real data showed that this hierarchical structure outperforms other methods when you look at performance across the whole hierarchy.
Lu: It suggests that for complex systems, sharing information structurally across levels is a way to get better parameter estimates when they are hard to find from the raw data alone.
Tom: So, it's less about a single perfect forecast and more about building a smarter structure for forecasting across different scales of reality.
Jane: And that leads us nicely into how this structural sharing compares to those simpler reconciliation methods we talked about earlier.
Queensland University of Technology
stat.ME
Submitted: 2026-06-22
Updated: 2026-11-11
Comments: 44 pages, 5 figures. Updated following reviewer and editor feedback
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 86/100
The gist: Decision-making in hierarchical systems requires probabilistic forecasts at all cross-sectional levels Our proposed fully Bayesian hierarchical forecasting framework shares information more…
Key concepts
- Bayesian Hierarchical Forecasting
- This method uses Bayesian statistics to model data structured in a hierarchy. It allows information from lower-level series to inform the estimation of higher-level parameters, creating a shared probabilistic structure that captures how different parts of the system are related.
- Coherency Penalty
- This is a mechanism where incoherence between upper and lower levels is penalized by multiplying the likelihood. This allows the model to gently balance maintaining consistency across levels against achieving high accuracy in individual forecasts, rather than strictly enforcing perfect coherence.
- Level Weighting Factor ($\lambda_i$)
- This factor is used to prioritize certain series or levels during forecasting. By multiplying the likelihood by $\lambda_i > 1$, the model places greater emphasis on that specific level, influencing parameter updates in parent or child nodes while still respecting the soft coherence constraint.
Terminology
Summary
Decision-making in hierarchical systems requires probabilistic forecasts at all cross-sectional levels Our proposed fully Bayesian hierarchical forecasting framework shares information more effectively between and across levels than reconciliation alone. p. 0 Such post hoc corrections do not
incorporate hierarchical structure or decision goals into the underlying parameter estimation. We propose a fully Bayesian hierarchical forecasting framework
that shares information more effectively between and across levels than reconciliation alone.Where the paper says it
Framework Overview
The paper proposes a fully Bayesian hierarchical forecasting framework that shares information more effectively between and across levels than reconciliation alone. This approach has the flexibility to softly penalise incoherence, subject to model specification, and to focus the global model and coherence update on hierarchical levels most relevant to decision outcomes. p. 0 We propose a fully Bayesian hierarchical forecasting framework
that shares information more effectively between and across levels than reconciliation alone. Our approach has the flexibility to softly penalise incoherence,
subject to model specification, and to focus the global model and coherence
update on hierarchical levels most relevant to decision outcomes. Our approach has the flexibility to softly penalise incoherence,
subject to model specification, and to focus the global model and coherence
update on hierarchical levels most relevant to decision outcomes. This yields
parameter estimates that are focused towards the forecasting goals and capture the requirement for coherency, removing the need to estimate covariance
matrices for multi-step forecasting horizons.Where the paper says it
Where the paper says it
Hierarchical Structure and Prior Specification
The methodology leverages Bayesian hierarchical methods to share information both between series at the same level and across hierarchical levels within a probabilistic framework. A hierarchical prior is specified where a hyperprior p(φ) with hyperparameters φ for the parameters, θi for the i-th series within the hierarchy. p. 9 In particular, we choose a hierarchical prior, in which we specify a hyperprior p(φ) with hyperparameters φ for the parameters, θi for the i
th series
within the hierarchy, y1,i, . This captures the prior assumption that the
parameters for each series are related through the population distribution specified by p(φ), but may differ from each other by the given hyperprior variance. In particular, we choose a hierarchical prior, in which we specify a hyperprior p(φ) with hyperparameters φ for the parameters, θi for the i
th series
within the hierarchy, y1,i, . This captures the prior assumption that the
parameters for each series are related through the population distribution specified by p(φ), but may differ from each other by the given hyperprior variance. (2019)
employing quantile reconciliation in a temporal hierarchy and Ben Taieb et al. (2021) using copula models on sub-hierarchies of parent-child nodes.Where the paper says it
Where the paper says it
Where the paper says it
Coherence and Level Weighting Mechanisms
The paper introduces a coherency penalty where the likelihood is multiplied by p(dσ 2 cohIT), penalising incoherence between a set of upper-level series and lower-level series. This distributional constraint means that incoherence is only softly penalised, allowing the model to flexibly balance coherence against accuracy. p. 1 We
then apply a soft constraint to reduce incoherence across the forecast expected
values and full distributions, subject to the level of model misspecification. We
combine this with the flexibility to differentially weight levels on the basis of
their decision importance, optimising the coherent hierarchy towards accuracy
at a specific level of interest. We
then apply a soft constraint to reduce incoherence across the forecast expected
values and full distributions, subject to the level of model misspecification. We
combine this with the flexibility to differentially weight levels on the basis of
their decision importance, optimising the coherent hierarchy towards accuracy
at a specific level of interest. (2021) using copula models on sub-hierarchies of parent-child nodes. Both approaches improve on single-level or independent models, but are limited in their
applicability to a wider range of data and forecasting applications (Panagiotelis
et al.Where the paper says it
Where the paper says it
Where the paper says it
Empirical Validation and Results
The framework is applied to simulated data and real data forecasting Australian domestic tourism demand. In the simulation study, the hierarchical Bayes approach outperforms MinT reconciliation in terms of overall RMSSE and ES when accounting for performance over the full hierarchy. p. 2 This reconciliation-by-conditioning
approach yields the same posterior means asymptotically as MinT reconciliation. Reconciliation-by-conditioning also outperforms base bottom-up forecasts, but
is not compared to other probabilistic forecasting methods, and the calibration
of forecasts at upper levels is not explicitly evaluated. Hierarchical Bayes meets hierarchical forecasting: A
flexible framework for level-focused forecasts
Arwen Nugterena,
∗
, Mahdi Abolghasemi
a
, Kerrie Mengersen
a
, Christopher
Drovandi
a
aQueensland University of Technology, School of Mathematical Sciences, Centre for Data
Science, 2 George St, Brisbane, 4000, Australia
Abstract
Decision-making in hierarchical systems requires probabilistic forecasts at all
cross-sectional levels. Current hierarchical forecasting methods typically generate independent forecasts at each level and reconcile them post hoc to ensure
coherence between upper and lower levels. We
combine this with the flexibility to differentially weight levels on the basis of
their decision importance, optimising the coherent hierarchy towards accuracy
at a specific level of interest. Ensuring coherence of forecasts began with single-level (bottom-up, topdown and middle-out) approaches that guaranteed coherence, but did not lever2Where the paper says it
Where the paper says it
Where the paper says it
Discussion and Limitations
The Bayesian hierarchical structure helps to estimate model parameters when they are challenging to identify from the given data or when it is known that series share characteristics. The approach is not appropriate if series do not share characteristics or if different parametric models are used for different series, though coherency penalty and level weightings still apply. p. 4 We leverage these properties to
share information both between series at the same level and across hierarchical
levels within a probabilistic framework. Bayesian hierarchical methods have been used within forecasting for applications such as sales transaction forecasting (Agosta and Inchiosa, 2023), building
electricity demand (Grillone et al. archical model can be focused on a level with better data quality and hence
improve forecasting performance at other levels. Our coherency penalisation
approach is incorporated into the training of model parameters, allowing simple
calculation of multi-horizon forecasts without the need to estimate multi-horizon
covariance matrices of forecasting errors. In this paper, we develop
a Bayesian hierarchical approach to generate hierarchical probabilistic forecasts
that share information between and across levels to improve forecast quality. We
then apply a soft constraint to reduce incoherence across the forecast expected
values and full distributions, subject to the level of model misspecification.Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Bayesian hierarchical forecasting allows for
partial information pooling informed by data
by specifying a hierarchical prior, which helpsshrink extreme or uncertain estimates of theta i towards the mean parameter estimate,
improving series with noisy or incomplete data. -
The proposed framework can be focused on specific levels within a global model to
maintain or improve performance while softly adhering to aggregation constraints,
meaning it canfocus on a level with better data quality and hence improve forecasting performance at other levels.
-
The coherency penalty allows the model to
flexibly balance coherence against accuracy, avoiding rigid constraints that may produce unrealistic predictions, particularly in the presence of model misspecification.
-
Level weighting allows prioritization by applying a factor lambda i > 1 to up-weight data from a selected series or level, which can
carry through to update the parameters of the parent or child nodes
while maintaining soft coherence. -
The system can outperform MinT reconciliation in point and probabilistic accuracy metrics overall because
a Bayesian hierarchical approach better captures the underlying structure through the sharing of information between and across hierarchical levels of the parameter hierarchy.
Abstract
Decision-making in hierarchical systems requires probabilistic forecasts at all cross-sectional levels. Current hierarchical forecasting methods typically generate independent forecasts at each level and reconcile them post hoc to ensure coherence between upper and lower levels. Such post hoc corrections do not incorporate hierarchical structure or decision goals into the underlying parameter estimation. We propose a fully Bayesian hierarchical forecasting framework that shares information more effectively between and across levels than reconciliation alone. Our approach has the flexibility to softly penalise incoherence, subject to model specification, and to focus the global model and coherence update on hierarchical levels most relevant to decision outcomes. This yields parameter estimates that are focused towards the forecasting goals and capture the requirement for coherency, removing the need to estimate covariance matrices for multi-step forecasting horizons. We demonstrate improvements in predictive accuracy metrics on both simulated data and Australian domestic tourism forecasting.
Sources
- Machine learning applications in time series hierarchical forecasting
- Hierarchical Bayesian Regression for Multi-Location Sales Transaction Forecasting
- Modeling the uncertainty on the covariance matrix for probabilistic forecast reconciliation
Related papers
- Learning Robust Treatment Rules for Censored Data
- Neural Estimation of Summary Statistics for Simulation-Based Inference with Non-Asymptotic Guarantees
- Spectral divide-and-conquer MCMC for long stationary time series
- Improving the Efficiency of Subgroup Analysis in Randomized Controlled Trials with TMLE
- Personalized Federated Vector Autoregression with Personalization Diversity
- Doubly Robust and Efficient Calibration of Prediction Sets for Right-Censored Time-to-Event Outcomes