Concept drift of simple forecast models as a diagnostic of low-frequency, regime-dependent atmospheric reorganisation

TL;DR

This study treats concept drift as a physically meaningful diagnostic of non-stationarity in atmospheric dynamics under climate change. By training simple, spatially aware linear predictors of daily $MSLP$ and 2 m temperature on the 1950s and 2000s and evaluating them on 2020–2024, the authors quantify drift as the relative $RMSE$ difference to map how empirical input–output rules reorganize over time. They find that drift is dominated by low-frequency variability and is strongly regime-dependent, with Western European summer drift largely driven by land–atmosphere coupling and winter drift decoupled from extremes; teleconnection reorganisation influences some regions but does not drive hemispheric drift. Importantly, drift is largely orthogonal to volatility or mean-change diagnostics, implying that concept drift provides a distinct and actionable view of evolving predictability that can inform regime-aware ML design for weather prediction in a warming world.

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 tp 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.

Figures (3)

• Figure 1: Large-scale circulation regimes modulate concept drift.a, Mean low-frequency MSLP concept drift over the North Atlantic--European sector stratified by NAO phase ($n=1{,}827$ days). Drift is significantly amplified during the positive NAO phase ($+40.9\%$ relative to neutral; $p < 10^{-6}$), indicating that the linear mappings governing zonal flow have degraded more than those governing blocking or meridional flow. Error bars represent the 95% confidence interval. b, Distribution of winter temperature drift in the North American Great Plains stratified by ENSO phase. Drift is significantly elevated during El Niño winters (mean $+0.167$ K) compared to La Niña or neutral conditions ($p < 10^{-12}$), reflecting a breakdown in teleconnection stationarity specific to the El Niño base state.
• Figure 2: Drivers of concept drift in Western European summer temperatures.a, Spatial relationship between mean summer warming (2000s minus 1950s) and concept drift (RMSE increase). While drift correlates with warming ($r=0.59$, $p<0.001$), warming magnitude explains only 34.6% of the spatial variance ($R^2=0.346$), indicating substantial residual non-stationarity. The solid red line indicates the linear fit; the dashed grey line ($y=x$) indicates the theoretical 1:1 scaling. b, Dependence of concept drift on changes in land--atmosphere coupling. Over land grid points (red circles), drift is strongly correlated with coupling changes ($r=0.63$, $p<10^{-15}$), whereas over ocean grid points (blue dots), the relationship is weak ($r=0.16$). This land--ocean contrast demonstrates that thermodynamic feedbacks, rather than atmospheric heating alone, govern the loss of predictability over continental Europe.
• Figure 3: Decoupling of concept drift from atmospheric volatility.a--c, Spatial maps of the change between the 1950s and 2000s in (a) concept drift (RMSE difference), (b) pressure volatility (variance change), and (c) temporal persistence (lag-1 autocorrelation change). Drift hotspots (e.g. Mediterranean, North Pacific) do not align with regions of increased volatility. d, Spatial scatter plot of drift vs. volatility change showing a negligible relationship ($r = 0.077$, $R^2 < 0.01$). e, Spatial scatter plot of drift vs. persistence change showing a moderate positive relationship ($r = 0.29$), indicating that drift is physically linked to changes in system memory rather than system noise.