Bayesian Reconciliation of Overlapping Multi Step Forecasts: A Post Processing Framework for Cross Origin Coherence
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: "Bayesian Reconciliation of Overlapping Multi Step Forecasts".
Jane: Here's the gist of the paper. Multi-step forecasting systems often issue overlapping predictions that can disagree even from the same model, a phenomenon termed cross-origin incoherence The proposed Bayesian reconciliation framework treats these overlapping forecasts as noisy and potentially biased measurements of a common latent trajectory This method operates as a model-agnostic post-processing layer, requiring no access to the forecasting model’s architecture, parameters, or training data It jointly reconciles forecasts targeting the same future time while accounting for horizon-specific bias and uncertainty using a Bayesian state-space model.
Tom: Controlled simulations demonstrate substantial reductions in cross-origin incoherence and accurate latent trajectory recovery under the assumed data generating process.
Jane: Defining Cross-Origin Incoherence Cross-origin incoherence is defined as disagreement among overlapping forecasts that target the same future outcome but are issued at different forecast origins The cross-origin incoherence metric is formally defined as the average squared disagreement among all forecasts targeting the same future time t A perfectly coherent system has I = 0, meaning all forecasts for the same future date agree exactly.
Tom: The metric is used specifically as a post-processing diagnostic and as a measure of the degree of disagreement before and after Bayesian reconciliation.
Paper summary: Jane: We’ve been looking at how the authors of Bayesian Reconciliation of Overlapping Multi Step Forecasts: A Post Processing Framework for Cross Origin Coherence tackled this issue with a Bayesian state-space model to treat overlapping forecasts as noisy measurements of a common latent trajectory >
Tom: They're proposing this framework as something model agnostic, meaning it doesn't care if you used an AR one or an LSTM, its only job is to correct the structural failure mode where forecasts for the same future event shouldn't contradict each other depending on when they were made >
Lu: The core of it is that the inner sum in Equation one computes that average squared disagreement among all forecasts targeting the same future time t, and then they take an outer sum over all target times to get a single scalar measure of cross-origin incoherence >
Meng: So from what I’ve seen, if you're running a production system where you need multiple overlapping predictions for the same output, this is a tool you could use to enforce coherence after the fact >
Lalam: It’s essentially treating forecasts targeting the same timestamp as noisy measurements of a common latent process and using that structure to separate horizon specific bias from the overall level of that latent trajectory >
Tom: The title itself says it's about reconciliation, which is important because you can have multiple good predictions but they might fight each other when they overlap >
Jane: So what this implies for the broader area of forecasting is that we can have a system where different parts disagree on the exact future point, but we can use this Bayesian approach to make them all agree on a single coherent path >
Conclusion: Tom: So we've been looking at this paper about Bayesian Reconciliation of Overlapping Multi Step Forecasts, and what they're doing is taking all those overlapping predictions that might contradict each other and trying to make them agree on one path using a Bayesian approach >
Jane: Right, so the title itself says it’s about reconciliation, which is important because you can have multiple good predictions but they might fight each other when they overlap >
Lu: The authors are proposing this framework as something model agnostic, meaning it doesn't care if you used an AR one or an LSTM; its only job is to correct the structural failure mode where forecasts for the same future event shouldn't contradict each other depending on when they were made >
Meng: From my side, I’m wondering what that means for deployment; if we can use this after the fact, does it just fix a calibration issue or something deeper about how these models interact >
Lalam: It’s essentially treating forecasts targeting the same timestamp as noisy measurements of a common latent process and using that structure to separate horizon specific bias from the overall level of that latent trajectory >
Tom: So, what this implies for the broader area of forecasting is that we can have a system where different parts disagree on the exact future point, but we can use this Bayesian approach to make them all agree on a single coherent path >
Jane: It’s about taking these messy predictions and forcing them onto a shared reality, which is actually really cool because it gives us some concrete numbers on how much they were fighting each other >
University of Sao Paulo, Brazil
stat.ME
Submitted: 2026-10-04
Updated: 2026-10-31
Comments: 24 pages, 7 figures, 4 tables. This version corrects the empirical pipeline, fixes the LSTM forecasting bug, and adds an ETS benchmark
Code: https://github.com/Fajtos/forecast-reconciliation
License: http://creativecommons.org/licenses/by-nc-nd/4.0/
Importance score: 85/100
The gist: Multi-step forecasting systems often issue overlapping predictions that can disagree even from the same model…
Key concepts
- Cross-Origin Incoherence
- This is the disagreement among multiple forecasts that all aim to predict the same future outcome but come from different forecasting systems. A perfectly coherent system would have zero incoherence, meaning all overlapping predictions for a specific time point match exactly.
- Bayesian State-Space Framework
- The method models the entire forecast history as noisy observations of a hidden, smooth underlying process (a stationary AR(1) process). It uses Bayesian statistics to estimate this latent trajectory while accounting for different biases and uncertainties that occur at different forecast horizons.
- Model Agnosticism
- The framework is designed to work on any existing forecast, regardless of how it was generated. It does not require access to the forecasting model's internal structure or training data, making it a simple post-processing layer for improving forecast quality.
Terminology
Summary
Multi-step forecasting systems often issue overlapping predictions that can disagree even from the same model, a phenomenon termed cross-origin incoherence The proposed Bayesian reconciliation framework treats these overlapping forecasts as noisy and potentially biased measurements of a common latent trajectory This method operates as a model-agnostic post-processing layer, requiring no access to the forecasting model’s architecture, parameters, or training data It jointly reconciles forecasts targeting the same future time while accounting for horizon-specific bias and uncertainty using a Bayesian state-space model. p. 0 The method operates as a model-agnostic post-processing
layer, requiring no access to the forecasting model’s architecture, parameters, or training data. A Bayesian state-space model jointly reconciles forecasts targeting the same
future time while accounting for horizon-specific bias and uncertainty. A Bayesian state-space model jointly reconciles forecasts targeting the same
future time while accounting for horizon-specific bias and uncertainty. Controlled simulations demonstrate substantial reductions in cross-origin incoherence and accurate
latent trajectory recovery under the assumed data generating process.Where the paper says it
Where the paper says it
Defining Cross-Origin Incoherence
Cross-origin incoherence is defined as disagreement among overlapping forecasts that target the same future outcome but are issued at different forecast origins The cross-origin incoherence metric is formally defined as the average squared disagreement among all forecasts targeting the same future time t A perfectly coherent system has I = 0, meaning all forecasts for the same future date agree exactly. The metric is used specifically as a post-processing diagnostic and as a measure of the degree of disagreement before and after Bayesian reconciliation. p. 2 This formulation is closely related to the broader notion of forecast stability, recently
studied in multi-horizon forecasting as consistency in predicting the same future event as
the forecast origin changes (Ma et al. The present metric is used specifically as a
post-processing diagnostic and as a measure of the degree of disagreement before and after
Bayesian reconciliation.Where the paper says it
The Bayesian State-Space Framework
The framework reinterprets the forecaster’s entire forecast trajectory not as final predictions, but as noisy measurements of a latent coherent process The latent coherent trajectory is modeled as a stationary AR(1) process, µt = ϕµt−1 + ηt, where ϕ < 1 captures persistence in the underlying process. Each forecast is treated as a noisy measurement of the corresponding latent state, linked by the observation equation yˆAI t+ht = µt+h + δh + ϵt,h. p. 6 For the observation model, each forecast is treated as a noisy, possibly biased measurement of the corresponding latent state. The observation equation links the latent coherent
trajectory to the forecasts produced by the forecasting system. Here, δh is a horizon-specific additive bias (e. , systematic over- or under-prediction at
horizon h), and τ
2
h
is the horizon-specific forecast variance, quantifying the forecaster’s reliability at that horizon.Where the paper says it
Where the paper says it
Validation and Performance Metrics
The framework was validated through a comprehensive simulation study with 100 independent replications, achieving a mean incoherence reduction of 58.5% The method transforms raw forecasts with a mean incoherence of I = 0. p. 2 Alternative measures of disagreement could be based
on absolute pairwise deviations or forecast variance across common target dates. The squared
difference formulation adopted here is analogous to mean squared error and places greater
weight on severe contradictions among forecasts (Gneiting and Raftery, 2007; Timmermann,
2006). Such forecasts should ideally provide a consistent representation of that outcome, motivating the study of cross-origin incoherence among overlapping predictions, that is, when a
model’s multi-step forecast trajectory contains internal contradictions. The distinction is
therefore not between reconciliation and non-reconciliation, but between the structural constraints being reconciled: the present framework considers multiple forecasts of the same
target timestamp generated at different forecast origins, rather than consistency across temporal aggregation levels. Controlled simulations demonstrate substantial reductions in cross-origin incoherence and accurate
latent trajectory recovery under the assumed data generating process. An empirical
application to daily S&P 500 volatility forecasts reduces 90% prediction interval coverage from 94.Where the paper says it
Where the paper says it
Where the paper says it
Model Agnosticism and Diagnostic Insights
The approach is model-agnostic and operates on already-generated forecasts as a postprocessing layer, requiring no access to the forecasting model’s internal architecture or training procedure The empirical results show that the framework works effectively regardless of the initial coherence level, serving both as a corrective tool for poorly calibrated systems (AR(1)) and as a rigorous coherence guarantee for well-behaved systems (LSTM). Furthermore, estimated horizon-specific parameters provide diagnostic information, such as showing that reliability parameters τh show a clear U-shaped pattern with best reliability at horizon 3. p. 4 The latent state here is not a reconciliation error but the coherent trajectory itself, and
the constraint is an identity constraint at each target time rather than a linear aggregation
constraint. The horizon-specific bias and reliability parameters estimated in our framework
have no counterpart in hierarchical reconciliation.Where the paper says it
Conclusion and Limitations
The empirical results provide evidence that the proposed Bayesian reconciliation framework can successfully enforce cross-origin coherence across overlapping multi-step forecasts The approach is model-agnostic and operates on already-generated forecasts as a postprocessing layer, providing diagnostic information without requiring access to model internals. Limitations include the assumption that horizonspecific reliability τ 2 h is constant over time and the need for extensions to handle heavy-tailed or skewed distributions The framework specifically addresses cross-origin disagreement by treating forecasts targeting the same future timestamp as noisy and potentially biased measurements of a common latent trajectory. p. 5 Recent work has explicitly introduced
stability-aware evaluation and optimization for forecasts targeting the same future events
as the forecast origin changes (Ma et al. In contrast, the present framework uses
Bayesian state-space inference to reconcile an already-generated forecast matrix by treating
overlapping forecasts as noisy and potentially biased measurements of a common latent trajectory. prior on δh, together with the sum-to-zero constraint, allows for horizon-specific systematic
biases while ensuring model identifiability. By Bayes theorem, the joint posterior distribution of latent states and parameters is then
given by,
p(µ, θ | {yˆ
AI
t+h|t}) ∝ p({yˆ
AI
t+h|t} | µ, θ)
× p(µ | ϕ, σ2
η
) p(ϕ) p(σ
2
η
)
Y
H
h=1
p(δh) p(τ
2
h
),
(5)
where p(µ | ϕ, σ2
η
) denotes the AR(1) process density implied by Equation 2.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 Reconciliation Framework for Forecast Matrix Coherence: The improved system can take overlapping multi-step forecasts from different origins and treat them as
noisy and potentially biased measurements of a common latent trajectory.
This allows the system to enforcelogical coherence
by mapping the forecast matrix to a coherent latent state, effectively correctingcross-origin incoherence
without needing access to the forecasting model’s architecture. -
Model Agnostic Post-Processing Layer: The framework acts as a
model agnostic post-processing layer, requiring no access to the forecasting model’s architecture, parameters, or training data.
This enables its application to any AI system, including modern neural architectures like LSTMs and AR(1) models, by operating exclusively on theforecaster’s output matrix.
-
Latent State Recovery and Diagnostic Capabilities: The system can estimate a coherent latent trajectory mu t through Bayesian inference. It also estimates horizon-specific bias (delta h) and reliability (tau h) parameters, which serve as
diagnostic of the forecaster’s behavior,
identifyingsystematic biases and reliability variations across forecast horizons.
-
Robustness to Forecast Generation Mechanisms: The framework can enforce coherence even when the underlying forecasting model departs from linear assumptions, such as using a threshold autoregressive (TAR) model with heavy-tailed Student-t errors, achieving
zero residual incoherence in the non-linear setting.
Abstract
Overlapping multi step forecasts often disagree, a form of vertical instability that we term cross origin incoherence. We propose a Bayesian reconciliation framework that treats overlapping forecasts as noisy, biased measurements of a common latent trajectory. Operating as a model agnostic post processing layer, it requires no access to model internals. Simulations demonstrate substantial incoherence reduction and robust latent trajectory recovery, remaining highly resilient (correlation 0.972) even under severe nonlinear stress tests. An empirical application to S&P 500 volatility evaluates AR(1), ETS, and LSTM models. The framework reduces incoherence by over 97% across all models while delivering fully calibrated probabilistic forecasts (93% to 94% coverage). Notably, it improves point accuracy for ETS and LSTM. Furthermore, the estimated horizon specific parameters act as a diagnostic X-ray, revealing hidden biases in neural networks without requiring access to their architecture
Sources
- Beyond Accuracy: A Stability-Aware Metric for Multi-Horizon Forecasting
- Robust Bayesian Filtering and Smoothing Using Student's t Distribution
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