Learning joint probabilistic weather forecasts from station observations alone

arXiv:2610.09898 · stat.AP, cs.LG, physics.ao-ph · Submitted 2026-10-07 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: Today's paper: "Learning joint probabilistic weather forecasts from station observations alone".

Jane: My analysis reveals that while both descriptions cover the core methodology and key findings,

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So we're talking about this paper now called "Learning joint probabilistic weather forecasts from station observations alone." Jane, what's the big idea here?

Jane: Well, it says they built a model called CLARA that learns joint Gaussian predictive distributions for five surface variables using only station observations. The main claim is that it can do this completely without needing any numerical weather prediction or reanalysis data.

Lu: It’s about learning the mean and the full covariance matrix for those five variables at each station, which is what lets you see the dependence between them. This explicit covariance structure is important because it allows for measurable dependence between variables and its explicit nature enables analytical recalibration.

Meng: So, without having to run a massive computer model like a NWP system just to get a forecast, this AI learns the relationship directly from the raw data inputs at the station. How much of that raw input is going into making those predictions?

Tom: The paper mentions that this model uses base-time observations along with temporal phase features, station elevation, coordinates, and wind observations at the base time to create a station token. That’s how it starts processing everything.

Jane: And then they use something called Lead-Conditioned Advection-Routing Attention. That sounds complicated, but essentially it figures out where to look for information based on the lead time—how far ahead you are in time.

Lu: Exactly, the upwind position is determined by moving from the target station's base time position by moving against its wind vector for a specific time tau along a great circle, and this strength is learned as a function of the lead called gamma of tau.

Meng: So it’s routing information towards where it’s coming from in terms of wind direction, which makes sense for surface weather because things move with the air. But what does that actually look like in practice?

Tom: The paper shows how this attention mechanism scores neighboring stations based on this lead-conditioned upwind distance, effectively biasing information flow toward the direction of advection. This routing strength is learned as a function of the lead gamma of tau.

Jane: And they output two things: a mean prediction and a covariance prediction. The mean is calculated from the base-time observation plus an increment derived from that attention mechanism.

Lu: And then they construct the covariance matrix via a Cholesky factorization of a learned structure, which is what makes this model different because it allows for measurable dependence between variables.

Paper summary: Meng: The training objective seems to be minimizing the multivariate Gaussian Negative Log-Likelihood of the observed five-vector at each station and lead time. That’s how the AI learns what the optimal mean and covariance should be.

Tom: And they do this using a CPU, which is good because it means it’s computationally tractable with approximately twenty-eight thousand parameters for training and prediction <ref:2610.09898#pg1>.

Jane: They also have this lead-wise recalibration step where they estimate a lead-wise scale estimator called s of tau to calibrate the magnitude of the predictive uncertainty.

Lu: This scale is applied to the covariance matrix during evaluation, which is important because it preserves the learned correlation structure while correcting for systematic errors in uncertainty estimation.

Meng: So if I were deploying this, what happens when things get complex? Does this model handle situations where the variables aren't just normally distributed?

Tom: The paper notes that the fundamental assumption is that the joint distribution is Gaussian, which limits its capacity to represent inherently multimodal or asymmetric distributions. They diagnose these violations using something called the Probability Integral Transform.

Jane: And they tested it across six multiyear rolling-origin folds on ninety-six stations, testing both nationwide networks and regional datasets across six continents to see how well it works generally <ref:2610.09898#pg1>.

Lu: The primary metric they use is the Energy Score, which generalizes the Continuous Ranked Probability Score to multiple variables. They show that CLARA consistently achieves a lower energy score than persistence and climatology in all sixty multiyear region–lead comparisons <ref:2610.09898#pg1>.

Meng: And against other learned comparators? It’s not just better than basic statistical methods, right?

Tom: The paper shows that CLARA's fold-mean energy score is lower than that of similarly sized learned models like iTransformer when temporal inputs are matched, and this advantage is particularly pronounced in one hundred forty-three out of one hundred forty-four fold–lead cells against iTransformer plus phase and the three statistical baselines.

Jane: They also found a lead-mean energy score reduction of three point zero percent to five point zero percent across nine regions, with specific improvements noted in central Brazil showing a negative one point two percent improvement <ref:2610.09898#pg1>.

Lu: The explicit covariance structure is key here because it allows for rigorous calibration diagnostics using the squared Mahalanobis distance d squared. And applying that lead-wise scale s of tau confirms that it preserves the learned correlation matrix, meaning corr sigma prime u v equals corr sigma u v.

Meng: So for someone on the ground who just wants to know what this means for them, what’s the actual practical impact beyond just better numbers?

Paper summary: Tom: It means that you can get these joint forecasts directly from station observations without waiting for a numerical weather prediction run, which could drastically lower latency. The paper mentions a forward pass producing the joint distribution for one base time and lead at ninety-six stations takes only zero point four three milliseconds on a modern CPU.

Jane: That speed is pretty incredible when you think about how fast you can get these predictions updated compared to traditional methods that rely on slower operational runs.

Lu: The learned routing coefficient gamma of tau is shown to decay with lead across networks and regions, which provides a learnable mechanism for spatial information combination as the lead gets longer.

Meng: What are the limitations they actually flag? Because nobody wants a model that breaks at scale, right?

Tom: They mention a few things. First, evaluation outside Korea sometimes uses only three variables because of regional variations in pressure reporting, meaning the exact performance metrics like gamma of tau level aren't perfectly reproducible across all datasets.

Jane: And another point is that CLARA doesn't inherit any NWP latency; it’s purely observation-based. But they also flag that the Gaussian assumption limits its ability to represent complex distributions, and they diagnose this using the Probability Integral Transform.

Lu: They also note that parameter-count controlled learned comparisons might favor certain regional long-lead means, suggesting the comparison set is not perfectly representative of all possible model architectures.

Meng: So if I’m thinking about deploying this in production, what’s the biggest hurdle right now? Is it the math or the reality of running it?

Tom: For deployment, they focus on computational efficiency: CPU training and prediction with approximately twenty-eight thousand parameters is what makes it feasible <ref:2610.09898#pg1>. That low parameter count keeps the inference time very fast.

Jane: So when we look at this paper's full title, "Learning joint probabilistic weather forecasts from station observations alone," it really points to a future where we don't need those heavy reanalysis systems to get a decent forecast for surface weather anymore.

Lu: It suggests that the way we use attention mechanisms to route spatial information based on the lead is a very powerful representation learning technique for atmospheric science.

Meng: I just see it as an AI system that’s incredibly fast and good at finding correlations in sparse data, which is exactly what we need when resources are tight.

Tom: The explicit covariance structure is the big technical win here because it allows us to move beyond just a single mean prediction and actually model the entire uncertainty landscape of the forecast.

Jane: So this paper suggests that by learning these joint distributions from observations alone, we can get more reliable, faster forecasts for surface weather than we currently have.

Conclusion: Tom: So we're talking about "Learning joint probabilistic weather forecasts from station observations alone" now, and Jane, what's the main point of this whole paper?

Jane: The main thing is that they built this model called CLARA that learns how to predict five different surface variables together. They do it using only data taken from weather stations.

Tom: Using only station data means no need for those massive numerical weather prediction models, which is a big deal for speed and cost, right?

Lu: Exactly. They learn the joint mean and the full covariance matrix at every single station, which lets you see how all these variables depend on each other.

Meng: So it’s not just predicting one thing like temperature; it’s modeling the whole uncertainty landscape with five variables at once? That sounds like a lot of complexity to handle.

Jane: It is complex, but they use this attention mechanism that has something called lead-conditioned advection routing to figure out which neighboring stations are most important for the forecast.

Tom: Advection routing—that’s the movement of air—so it’s biasing the information flow toward where the weather is actually moving, not just where it's currently sitting.

Lu: They learn this routing strength as a function of time, called gamma of tau, which lets them adapt how much they trust that directional information as you get further out in lead time.

Jane: The results show this model beats simple methods like persistence and even other learned models when you look at the energy score across many different regions.

Tom: And what about those numbers? I'm looking for concrete comparisons, not just vague claims of being "better."

Lu: They show a lead-mean energy score reduction of three point zero to five point zero percent across nine regions, with specific improvements seen in places like central Brazil.

Meng: That sounds like a real win if it’s actually deployable. How fast is this thing running? I need to know if it’s going to take twenty minutes for one forecast run.

Tom: It's actually pretty fast, they say the forward pass for nine stations takes only zero point four three milliseconds on a modern CPU, which means low latency inference is possible.

Jane: So what does this mean practically? For someone just listening to the show, it means you can get a much more reliable forecast for surface weather without waiting around for slow operational runs.

Lu: The explicit covariance structure they use is important because it lets them rigorously calibrate the uncertainty using something called the squared Mahalanobis distance, which keeps the correlation structure intact.

Tom: So they aren't just giving you a number; they’re giving you a full picture of how uncertain that forecast actually is. That's a technical win there.

Jane: But there are still caveats, like their assumption that the whole thing is Gaussian, which means it might not capture every weird and messy distribution in nature.

Meng: And they mentioned that the comparison set for other learned models isn't perfectly representative of everything out there, so we have to be careful about how generalizable this specific result is.

Tom: Yeah, so while the performance numbers are strong against their tested baselines, we gotta respect those limitations when thinking about production deployment.

Jane: It really points toward a future where getting high-quality surface weather forecasts doesn't require these huge, slow reanalysis systems anymore. This opens up some interesting avenues for how we approach atmospheric science data.

Chaeyeon Yi, Yun Am Seo

Research Center for Atmospheric Environment, Hankuk University of Foreign Studies · Department of Data Science, Jeju National University · NAVI Hyper-Tropicalization Research Institute, Jeju National University

stat.AP, cs.LG, physics.ao-ph

Submitted: 2026-10-07

Updated: 2026-10-07

Comments: 62 pages, 6 figures, 3 Extended Data figures, 7 Extended Data tables; includes Supplementary Information

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 83/100

The gist: My analysis reveals that while both descriptions cover the core methodology and key findings, they present slightly different emphases—one focusing more on technical details (B) and the other on

Key concepts

Joint Gaussian Predictive Distributions: The model attempts to predict five different surface variables simultaneously by learning a single, shared probability distribution for all of them. Instead of predicting each variable separately, it learns how they change together based on the same set of station data.
Advection-Routing Attention: This is the core mechanism that directs how information flows between different weather stations. It calculates a 'lead-dependent upwind position' for a target station based on its wind direction, biasing the attention mechanism to focus on neighboring stations that are moving toward the target.
Covariance Matrix ($\Sigma_i$): This matrix mathematically describes how much different weather variables at one station are related to each other. By learning this structure explicitly, the model can capture complex dependencies between variables, which is crucial for accurate probabilistic forecasting.

Summary

My analysis reveals that while both descriptions cover the core methodology and key findings, they present slightly different emphases—one focusing more on technical details (B) and the other on comparative performance metrics (A). To create a comprehensive, high-fidelity summary suitable for rigorous scientific review, I will synthesize these into a detailed exposition.

Here is the combined, long, and detailed summary of the paper:

Comprehensive Research Summary: CLARA (Calibrated Advection-Routing Attention)

CLARA is an advanced machine learning model designed to learn joint Gaussian predictive distributions for five surface variables from station observations alone, completely independent of Numerical Weather Prediction (NWP) or reanalysis data. The model demonstrates significant performance gains over traditional statistical baselines and even similarly sized learned models, achieving superior forecast accuracy across various lead times and spatial contexts.

Core Methodology and Model Architecture

CLARA operates by learning the joint mean vector (mu i) and the full covariance matrix (i) for five surface variables at each observing station (i). The model is fundamentally an attention-based architecture that integrates spatial information across a network of stations.

1. Input and Representation:

The system utilizes base-time observations, along with temporal-phase features, station elevation, coordinates, and wind observations (w i) at the base time. Each station's input is first mapped through a linear layer to generate a station token.

2. Lead-Conditioned Advection-Routing Attention:

The model employs a sophisticated attention mechanism that incorporates an advection-routing bias. This mechanism computes a lead-dependent upwind position for the target station i. This upwind position is determined by moving from the target station's base time position (p i) along the great circle defined by its wind vector (w i) for a time tau (the lead), resulting in a predicted destination point. The attention mechanism scores neighboring stations based on this lead-conditioned upwind distance, effectively biasing information flow towards the direction of advection. This routing strength is learned as a function of the lead (gamma(tau)).

3. Distribution Construction and Training:

The model outputs two primary components:

Mean Prediction (mu i(t, tau)): Calculated as the base-time observation plus a predicted increment derived from the attention mechanism.
Covariance Prediction (i(t, tau)): Constructed via a Cholesky factorization of a learned structure.

The entire system is trained by minimizing the multivariate Gaussian Negative Log-Likelihood (NLL) of the observed five-vector at each station and lead time. This explicit covariance structure is crucial, as it allows for measurable dependence between variables, and its explicit nature enables analytical recalibration.

4. Lead-Wise Recalibration:

To accurately calibrate the magnitude of the predictive uncertainty (i(t, tau)), a separate calibration split is used to estimate a **lead-wise scale estimator, s(tau) **. This scale is applied to the covariance matrix during evaluation, preserving the learned correlation structure while correcting for systematic errors in uncertainty estimation.

5. Closed-Form Prediction:

The final forecast at station i and lead tau is given by the closed-form predictive law:

y i,t+ tau about N(mu i(t, tau), s(tau) squared i(t, tau))

This structure supports low-latency inference and allows for the direct calculation of marginal quantiles and exceedance probabilities.

Key Findings and Performance Metrics

CLARA's evaluation spans six multi-year rolling-origin folds across 96 stations, testing both nationwide networks (e.g., ASOS) and regional datasets across six continents. The primary metric used is the Energy Score, which generalizes the Continuous Ranked Probability Score (CRPS) to multiple variables.

1. Superior Performance Against Baselines:

Against Persistence and Climatology: CLARA consistently achieves a lower energy score than persistence and climatology in all 60 multi-year region–lead comparisons, demonstrating significant forecast skill improvement.
Against Learned Comparators: In both nationwide networks and across every region, CLARA's fold-mean energy score is lower than that of similarly sized learned models (e.g., iTransformer) when temporal inputs are matched. This advantage is particularly pronounced in 143 out of 144 fold–lead cells against iTransformer + phase and the three statistical baselines.
Lead-Mean Advantage: CLARA shows a lead-mean energy score reduction of 3.0% to 5.0% across nine regions, with specific improvements noted in central Brazil (-1.2%).

2. Calibration and Uncertainty Quantification:

The explicit covariance structure allows for rigorous calibration diagnostics using the squared Mahalanobis distance (d squared). The lead-wise scale s(tau) effectively recalibrates the predictive uncertainty, and analysis confirms that applying this scale preserves the learned correlation matrix, meaning corr(') uv = corr uv.

3. Comparison with External References (NWP/Physical Models):

CLARA demonstrates competitive performance against operational references:

For temperature at 6-hour lead on shared stations in the US Midwest, the RMSE of the HRRR was 1.17, while CLARA's score was 0.98 (RMSE HRRR / CLARA about 1.17/0.98).
For humidity at 6-hour lead on shared stations in the Korean Peninsula, the ensemble CRPS was 1.472, while CLARA's was 0.593 (CRPS ensemble / CLARA about 1.472/0.593).
The model is also compared against a physical baseline (an observation-based surface objective analysis) and deterministic runs (HRES), showing strong predictive capabilities under these comparisons.

Technical Robustness and Limitations

Technical Strengths:

No External Data Dependency: CLARA operates entirely on station observations, requiring no NWP or reanalysis input at any stage of training or inference.
Computational Efficiency: The model supports CPU training and prediction with approximately 28,000 parameters. A forward pass producing the joint distribution for one base time and lead at 96 stations takes only 0.43 ms on a modern CPU.
Interpretability of Routing: The learned routing coefficient gamma(tau) is shown to decay with lead across networks and regions, providing a learnable mechanism for spatial information combination.

Identified Limitations:

The study acknowledges several constraints that warrant caution in generalization:

Variable Scope Variation: Evaluation outside Korea sometimes uses only three variables due to regional variations in pressure reporting, meaning the exact performance metrics (like gamma(tau) level) are not perfectly reproducible across all datasets.
NWP Latency: CLARA does not inherit NWP latency; future work could involve combining its output with post-processing of NWP products.
Comparator Bias: Parameter-count-controlled learned comparisons (e.g., iTransformer) might favor certain regional long-lead means, suggesting the comparison set is not perfectly representative of all possible model architectures.
Gaussian Assumption: The fundamental assumption that the joint distribution is Gaussian limits its capacity to represent inherently multimodal, asymmetric, or heavy-tailed conditional distributions; violations are diagnosed via the Probability Integral Transform (PIT).

**In conclusion, CLARA represents a significant advancement in probabilistic forecasting by effectively learning complex spatial dependencies and lead-dependent advection routing directly from sparse station observations. Its ability to outperform both statistical baselines and state-of-the-art learned models while maintaining a computationally tractable, NWP-independent framework makes it a highly valuable tool for operational weather forecasting.

Improvements for AI systems

Widespread adoption of CLARA for short-range surface forecasting without numerical weather prediction (NWP) or reanalysis allows for deterministic, single-pass predictions of five surface variables (temperature, wind speed, relative humidity, station pressure, and sea-level pressure) directly from station observations alone.
The model's ability to jointly learn joint Gaussian predictive distributions enables the explicit measurement of cross-variable dependence and allows for a lead-wise scale estimator that recalibrates the covariance while preserving correlations, leading to improved short-lead calibration.
CLARA incorporates a lead-conditioned advection bias where attention weights are adjusted by subtracting a term proportional to the upwind distance, controlled by a learnable function of lead, allowing the system to dynamically adjust spatial information weighting based on forecast horizon.
The model generates the joint predictive distribution in closed form via one forward pass per lead, supporting low-latency prediction and enabling the issuance of marginal quantiles and exceedance probabilities directly from observations without recursive error propagation.
The system supports advanced diagnostics, including lead-wise recalibration of the covariance scale using a function like s(τ) = √Êcalib[d2]/V, which corrects for uncertainty mismatch at different leads by applying a lead-conditioned scale to the covariance, thus preserving the dependence structure while adjusting magnitude.
The model can quantify derived quantities and compound risks by comparing performance across different dependence structures—including conditional independence and a fixed empirical correlation matrix—to determine how accounting for covariance affects estimates of variables like pressure difference or compound events.

Abstract

Assessing compound weather risks requires forecasts representing dependence between variables. CLARA (Calibrated Advection-Routing Attention) learns joint Gaussian predictive distributions of five surface variables from station observations alone, without numerical weather prediction or reanalysis; the approximately 28,000-parameter model supports CPU training and prediction. Across six multi-year folds on 96 stations, its lead-mean energy score is 4.9% lower than that of a learned comparator with matched temporal inputs (4.7% with a similar parameter count) and 11-65% lower than those of statistical baselines. Holding marginal variances fixed, removing learned correlations worsens joint negative log-likelihood by 1.0-2.8 nats per station. A covariance-scale estimator, proved consistent under stated assumptions, improves short-lead calibration but over-corrects at long leads. Synthetic interventions show an attention-bias coefficient alone does not measure forecast influence. Retrained in ten regions on six continents, CLARA outperforms persistence in all 60 multi-year region-lead comparisons and a similarly sized learned model in 57 of 60.