Concept drift of simple forecast models as a diagnostic of low-frequency, regime-dependent atmospheric reorganisation
Haokun Zhou · 2025
Read Concept drift of simple forecast models as a diagnostic of low-frequency, regime-dependent atmospheric reorganisation by Haokun Zhou (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-19638.txt.
Introduction
Haokun Zhou turns the failure of simple weather models into a measuring tool for climate change. Data-driven forecast models assume the statistics linking predictors to outcomes stay fixed, an assumption a warming climate breaks. Zhou trains plain linear models of sea-level pressure and near-surface temperature, place by place, on data from the nineteen fifties and from the two thousands, then tests them on recent years. The performance gap between the two vintages defines a local, readable measure of concept drift.
Splitting errors by frequency, weather regime, and region shows the drift is dominated by slow, low-frequency change and depends strongly on the prevailing circulation pattern. Over the North Atlantic and Europe, drift peaks in one phase of the regional oscillation. The diagnostic maps how the atmosphere is reorganising beneath the averages.
Catalog record
| CREATOR | Haokun Zhou |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 40 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-19638
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Concept drift of simple forecast models as a diagnostic of low-frequency, regime-dependent atmospheric reorganisation Haokun Zhou Abstract Data-driven weather prediction models implicitly assume that the statistical relationship between predictors and targets is stationary. Under anthropogenic climate change, this assumption is violated, yet the structure of the resulting concept drift remains poorly understood. Here we introduce concept drift of simple forecast models as a diagnostic of atmospheric reorganisation. Using ERA5 reanalysis, we quantify drift in spatially explicit linear models of daily mean sea-level pressure and 2 m temperature. Models are trained on the 1950s and 2000s and evaluated on 2020–2024; their performance difference defines a local, interpretable drift metric. By decomposing errors by frequency band, circulation regime and region, and by mapping drift globally, we show that drift is dominated by low-frequency variability and is strongly regime-dependent. Over the North Atlantic–European sector, low-frequency drift peaks in positive NAO despite a stable large-scale NAO pattern, while Western European summer temperature drift is tightly linked to changes in land–atmosphere coupling rather than mean warming alone. In winter, extreme high-pressure frequencies increase mainly in neutral and negative NAO, whereas structural drift is concentrated in positive NAO and Alpine hotspots. Benchmarking against variance-based diagnostics shows that drift aligns much more with changes in temporal persistence than with changes in volatility or extremes. These findings demonstrate that concept drift can serve as a physically meaningful diagnostic of evolving predictability, revealing aspects of atmospheric reorganisation that are invisible to standard deviation and storm-track metrics. a]Imperial College London keywords Atmospheric dynamics, Concept drift, Forecast models, Regime dependence † † authorinfo: E-mail: [email protected] 1 Introduction Climate change has rendered the atmosphere a fundamentally non-stationary system. Mean states, variability and the spatial structure of teleconnections are all evolving, and these changes inevitably modify the empirical relationships that underpin statistical and machine-learning (ML) prediction models. When such relationships change over time, models trained on past data degrade when applied to new conditions, a phenomenon known in the ML literature as concept drift [ 10 ] . Concept drift has been treated extensively as an algorithmic challenge, but much less as a physical signal that can be diagnosed and interpreted. Within the ML community, concept drift is typically defined as a change in the joint distribution of inputs and outputs or in the conditional mapping between them, and a large body of work focuses on detection and adaptation methods such as sliding windows, ensemble schemes and explicit drift detectors [ 10 , 14 ] . In Earth-system applications, similar concerns have begun to appear. Environmental and hydrological models have been shown to lose skill as land use, forcing and observation systems change [ 20 , 3 ] , and recent discussions of ML-based weather prediction highlight the vulnerability of purely data-driven models to changes in the climate regime encoded in reanalyses [ 23 , 19 ] . In these studies, however, concept drift remains primarily a problem to be mitigated rather than a lens through which to view atmospheric change. At the same time, the physical climate literature has developed a detailed picture of the dynamical and thermodynamical processes that organise low-frequency variability. Large-scale teleconnections such as the El Niño–Southern Oscillation (ENSO) and the North Atlantic Oscillation (NAO) strongly structure temperature and pressure variability over the Euro-Atlantic and North American sectors, and a substantial body of work has shown that their extratropical impacts are non-stationary on decadal timescales [ 11 , 1 , 16 , 12 , 4 ] . In parallel, studies of the eddy-driven jet, storm tracks and blocking have demonstrated that midlatitude extremes are sensitive to subtle changes in jet latitude, storm-track structure and blocking frequency, and that different diagnostics can yield different interpretations of circulation change [ 27 , 18 , 17 ] . Over land, and particularly over Europe, thermodynamic feedbacks further complicate the picture. European summer heatwaves have been shown to be amplified and prolonged by soil-moisture–temperature feedbacks, and regions of strong land–atmosphere coupling have shifted under anthropogenic warming [ 24 , 8 , 9 , 15 , 25 , 6 , 5 ] . These studies show that the same mean warming can have very different impacts on variability and extremes depending on soil-moisture regimes and the background circulation. Despite this rich physical context, there remains a conceptual gap between dynamical studies of regimes, teleconnections and land-surface feedbacks on the one hand, and ML-oriented discussions of concept drift on the other. Existing work tells us where and how the atmosphere has changed, but it does not directly tell us where the empirical “rules” learned by simple predictive models become obsolete. This paper seeks to close that gap by treating concept drift itself as a physically meaningful diagnostic of non-stationarity in atmospheric dynamics. Our central idea is that if we train simple, interpretable statistical models on different climate eras and evaluate them on the same recent period, then differences in their performance can be interpreted as evidence that the underlying input–output relationships have changed. By decomposing these performance differences by frequency band, circulation regime and region, we can ask where drift is dominated by slow background shifts, where it is associated with teleconnection reorganisation, and where it is anchored in land-surface feedbacks or circulation regimes. In this way, the concept-drift perspective becomes a bridge between ML robustness questions and physical climate diagnostics. In what follows, we develop this idea using ERA5 reanalysis, simple spatially aware linear models and a set of standard dynamical and thermodynamical diagnostics. Section 2 describes our data and methods. Section 3 presents four sets of results that together form a coherent picture: low-frequency, regime-dependent drift in midlatitude pressure fields; the relationship between ENSO reorganisation and simple-model drift; the role of land–atmosphere coupling in Western European summer temperature drift; and the decoupling of winter extremes from predictability shifts in the North Atlantic–European sector. In Section 4 we discuss the implications of these findings for both climate interpretation and ML-based weather prediction, before summarising our conclusions in Section 5 . 2 Data and methods We base our analysis on the ERA5 reanalysis [ 13 ] , focusing on daily mean sea-level pressure (MSLP) and 2 m air temperature over the Northern Hemisphere extratropics ( 20^{\circ} – 80^{\circ} N). To define distinct climate eras, we divide the record into three periods: 1950–1959, 2000–2009 and 2020–2024. The first two periods serve as training decades, while the most recent period serves as a common evaluation window. We restrict our main analyses to regions and seasons where ERA5 is considered reliable back to 1950 and interpret early-period results in light of documented differences in pre- and post-satellite-era reanalysis fidelity. At each grid point we train two classes of simple daily prediction model. The first class consists of pointwise persistence regressions, which predict next-day MSLP or temperature from the same variable at the same location on the current day. The second class consists of spatial gradient models, which predict central MSLP from a 3\times 3 neighbourhood: the central value and its local zonal and meridional gradients. All models are linear and are fitted independently at each grid point by ordinary least squares. For each model class, we estimate separate parameter sets for the 1950s and 2000s using only data