A high-resolution, two-millennium monthly downscaled climate dataset for North America Authors: P. R. Guaita, P. Crippa Published: 2026 ABSTRACT Long-term regional climate projections are critical for validating and implementing regional impact assessments and for capturing slow processes in ecological models. Earth System Model (ESM) simulations are the primary source of climate data, but their coarse resolution, systematic biases, and limited representation of uncertainty restrict their applicability. Here, we present DPastCliM-NA, a bias-corrected, high-resolution (0.2° × 0.2°) monthly reconstruction of temperature and precipitation for North America over the past 2000 years, with explicit uncertainty estimates. The dataset is derived by statistically downscaling MPI-ESM1.2-LR simulations against the GHCN-m observational dataset for 1875–2014. Downscaling is performed using Principal Component Regression, and incorporates a model selection strategy to prevent overfitting. Spatiotemporal dependencies in the residuals are captured through spatial error models and autoregressive–moving-average processes. The resulting reconstructions show substantial improvements over raw ESM output, while preserving the long-term climate trends and variability represented by the ESM. Background & Summary Temperature and precipitation are fundamental drivers of both human and natural systems. Their variability has been extensively documented through direct and remotely sensed measurements since the industrial era, providing crucial insights into modern climate change1,2. To understand long-term pre-industrial climate dynamics and provide a baseline for anthropogenic warming, paleoclimate proxies have been instrumental in reconstructing climate records over past centuries3–5. These reconstructions, notably those covering the last two millennia, have provided essential evidence for the rapid, human-driven global warming observed over the last 150 years6,7. While Earth System Models (ESMs) are crucial for projecting future climate, they face significant challenges. The models are often calibrated using relatively short modern datasets, limiting their ability to accurately simulate, for example, long-term ecosystem shifts that unfold over centuries8. In this sense, paleoclimate simulations are essential for testing model performance beyond the scientific measurements era, as coordinated by the Paleoclimate Modeling Intercomparison Project (PMIP)9. Despite these advances, the ESMs used in these experiments are typically characterized by coarse grid spacing (100–200 km horizontal resolution), often contain systematic biases10, and do not have uncertainty estimates. These features preclude their direct applicability in regional climate impact assessment studies and their integration into ecological models, as the relevant processes (e.g., droughts, ecotone shifts) happen at much finer scales and could be significantly affected by biases11,12. To overcome these limitations, downscaling techniques are employed to refine ESM output to finer spatial resolutions and correct systematic biases. Downscaling methods can be broadly categorized as dynamical or statistical approaches13. While dynamical downscaling provides physically based, high-resolution simulations, its computational expense limits its application to short time periods. In contrast, statistical downscaling is computationally efficient, establishing empirical relationships between large-scale variables and local-scale observations to refine climate model outputs. Most statistical downscaling methods define an empirical link (transfer function) between a set of predictors (modeled data to be downscaled) and a set of predictands (usually observed data), and they can be classified under different categories depending on the employed methodology14. Transfer function methods range from simple linear regression to more advanced machine-learning approaches15,16. These methods are often applied in the context of the Coupled Model Intercomparison Project Phase 5 and 6 (CMIP5 and CMIP6) ensembles to correct systematic biases and enhance spatial resolution for regional climate impact studies17,18. At the time of writing, two long-term, high-resolution global paleoclimate reconstructions are publicly available: CHELSA-TraCE21k, which uses a statistical downscaling method based on topography and ice cover19,20, and PaleoClim, which uses the delta method21,22. Both products offer multi-year climatological averages rather than continuous high-frequency time series. While these products offer very high spatial resolution, the resulting large data volumes limit their temporal resolution. Furthermore, these models do not directly provide uncertainty estimates due to their inherently deterministic methodology. Typically, other regional downscaled products available at the paleoclimatic scale exhibit similar limitations, providing outputs at decadal to millennial temporal resolutions and at a spatial resolution of 0.5° × 0.5°23–25. This study presents a set of statistically downscaled monthly mean reconstructions of 2m air temperature (tas) and precipitation (pr) for the past 2000 years over North America26, generated by combining ESM simulations and observational datasets. A key feature of our product is the explicit inclusion of uncertainty estimates. Reconstructions are provided at individual measurement stations, allowing end-users to regrid the data to a resolution of their choice. A gridded product is also available; to limit storage requirements for continuous monthly reconstructions, it is provided at a coarser spatial resolution (0.2° × 0.2°) than some existing products, but retains sufficient detail for a wide range of applications. Methods The Principal Component Regression (PCR)27 is a type of Model Output Statistics13 approach. This technique is not only a spatial downscaling method - i.e., it refines an initial product from a coarse spatial resolution to a finer target resolution - but also corrects systematic biases in the mean climatology and variability of ESM simulations by directly linking simulations to observational data. Once calibrated over present times, the PCR can then be applied to climate simulations spanning past centuries. PCR employs first a principal component analysis27 on the climate model output. Principal Component Analysis (PCA) decomposes the model output into eigenvectors (spatial patterns, or spatial modes) and scores (temporal coefficients), which are then used in a regression model against the observed data. However, not all eigenvectors and scores need to be included in the downscaling model. In this study, a model selection process is employed to retain only those modes that contribute to a reduced Root Mean Squared Error (RMSE). The PCR is calibrated locally, on a station-by-station basis, and separately for each month. Calibration is conducted at the station level because long-term, homogeneous gridded observational datasets at monthly resolution are often either regionally restricted or cover shorter time periods than station observations (1875–2014 in this study). Furthermore, using station data allows direct use of quality-controlled observations without introducing artifacts from prior spatial aggregation; interpolation to the target grid is performed only after downscaling. As a result, a given ESM is associated with (12 × number of stations) distinct downscaling models. These models are then applied to the selected time periods using ESM simulation data specific to each station location, producing downscaled ESM time series at each station. Finally, a gridded downscaled product over the entire domain is obtained by spatially interpolating these time series onto the target grid. Figure 1 illustrates the general workflow of the downscaling process.Fig. 1Workflow diagram for the downscaling procedure, from data preprocessing to gridding. Workflow diagram for the downscaling procedure, from data preprocessing to gridding. The following sections describe (1) the employed datasets (GHCN-m, MPI-ESM1.2-LR simulations), (2) the PCR downscaling, and (3) the metrics used for validation. These section constitute the main body of the methods, and are followed by an additional section reporting specific equations and methodology details. Datasets GHCN-m observations Surface air temperature and precipitation from the Global Historical Climatology Network-monthly (GHCN-m) dataset version 4 are used as predictors in the proposed downscaling approach28,29. These datasets contain thousands of land station timeseries worldwide, which are processed via automated pairwise homogenization and undergo extensive quality assurance29–31. For the purposes of this study, the dataset is filtered over North America (lon = [−175°, −50°], lat = [5°,72°]) and for the years between 1875 and 2014 (Fig. 2). Furthermore, timeseries shorter than 20 years for both temperature and precipitation are removed entirely. The filtered dataset is divided into calibration, model selection, and testing subsets. Initially, of the 140 years in the record, 50 years are randomly assigned to model selection, 15 years to testing, and the remaining 75 years to calibration. However, to maintain reliable calibration statistics, if a station’s calibration period is shorter than 30 years, the entire time series of that station is assigned to testing. Figure 3 illustrates the dataset split for temperature and precipitation across the first 100 stations, as an example.Fig. 2GHCN-m stations for temperature (tas; a) and precipitation (pr; b), alongside their Kernel Density Estimation (KDE) as a function of latitude. Orange dots are calibration stations, green dots are testing-only stations. Starred stations indicate example stations shown in Fig. 5.Fig. 3GHCN-m dataset splitting into calibration, model selection and testing, for temperature (tas; a) and precipitation (pr; b), over the first 100 stations of the dataset. GHCN-m stations for temperature (tas; a) and precipitation (pr; b), alongside their Kernel Density Estimation (KDE) as a function of latitude. Orange dots are calibration stations, green dots are testing-only stations. Starred stations indicate example stations shown in Fig. 5. GHCN-m dataset splitting into calibration, model selection and testing, for temperature (tas; a) and precipitation (pr; b), over the first 100 stations of the dataset. ESM data: MPI-ESM1.2-LR The simulations considered in this study are from the experiments past2k32 and historical33 from the model MPI-ESM1.2-LR34,35, which is part of the PMIP49, one of the components of CMIP633. MPI-ESM1.2-LR includes coupled atmosphere, ocean, sea-ice, and land-surface processes. Its atmospheric component has a horizontal resolution of 1.875° × 1.8496° (~150–200km) and 47 vertical levels extending to the top of the atmosphere (~0.01hPa). The experiment past2k covers 0–1849 CE32, and is a transient simulation incorporating various forcings derived from reconstructed or observed datasets, including volcanic aerosols, greenhouse gas variability, land-use changes, and orbital variations, with a focus on key climate periods such as the Medieval Climate Anomaly, the Little Ice Age, and the Roman Warm Period. Other than the past2k experiment, the historical experiment (covering 1850–2014 CE) is included in this study in order to calibrate the downscaling model over the historical period. Temperature and precipitation fields for both experiments were obtained at monthly resolutions. MPI-ESM1.2-LR is selected because it provides a continuous, internally consistent monthly simulation spanning both the PMIP4 past2k (0–1849 CE) and historical (1850–2014 CE) experiments, allowing calibration and reconstruction within a unified modeling framework. At present, among PMIP4 models, only MPI-ESM1.2-LR provides monthly-resolved output covering the full two-millennium period required by our PCR-based downscaling framework. Downscaling with PCR In this work, we define a target downscaling grid over North America with a spacing of 0.20° × 0.20° (626 × 336 grid points, longitude × latitude). This horizontal resolution is compatible with the goal of a fine-scaled product with a monthly timestep, while retaining computational feasibility and manageable storage dimension. Henceforth we define Nt as the number of timesteps, Ns as the number of stations (6, 889 for tas and 11, 546 for pr), Ng as the number of nodes in the target grid (626 ⋅ 336 = 210, 336), and Ne as the number of selected eigenvectors/scores. The data value at a given timestep t and location x is denoted by M(x, t) or O(x, t) depending on whether it originates from the Earth System Model (ESM) or the observational dataset, respectively. The indexes are sometimes omitted for brevity: M(x) and O(x) refer to the full time series at location x. When both the time and location indices are omitted, M and O represent the entire dataset. The first step of the PCR (henceforth denoted as PCR-downscaling) consists in regridding the ESM maps from the historical experiment to the target downscaling resolution via spline interpolation to ensure consistency. More specifically, an Akima cubic Hermite interpolation is used, modified to reduce undulations by assigning greater weight to the side with the slope closer to zero in the local derivative estimates36,37. This spline-based interpolation is used at this stage to preserve large-scale gradients and ensure smooth transitions in the regridded ESM fields prior to PCA decomposition. From the regridded ESM maps, an anomaly matrix is defined by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\bf{M}}}^{{\prime} }={\bf{M}}-\overline{{\bf{M}}}\otimes {{\bf{1}}}_{{N}_{t}}^{T}$$\end{document}M′=M−M¯⊗1NtT (dimension Ng × Nt), where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{{\bf{M}}}$$\end{document}M¯ is the time average (vector of length Ng), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\bf{1}}}_{{N}_{t}}$$\end{document}1Nt is a vector of ones (length Nt). Then, PCA is conducted on the anomaly matrix, leading to its decomposition into E matrix of the eigenvectors (dimension Ng × Ne) and S matrix of the scores (dimension Nt × Ne): 1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\bf{M}}}^{{\rm{{\prime} }}}\approx {\bf{E}}{{\bf{S}}}^{T}$$\end{document}M′≈EST Once scores and eigenvectors are determined, a local regression model (the PCR-downscaling model) is defined, and calibrated independently for each month and each GHCN-m station, resulting in 12Ns PCR models of the form: 2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$g{({\bf{O}}({\bf{x}}))}^{{\prime} }={\bf{S}}\cdot {\boldsymbol{\beta }}({\bf{x}})+{\bf{u}}({\bf{x}})$$\end{document}g(O(x))′=S⋅β(x)+u(x)with β the regression coefficients vector (length Ne), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\bf{u}}({\bf{x}}) \sim {\mathcal{N}}\left(0,{\sigma }_{u}^{2}({\bf{x}})\right)$$\end{document}u(x)~N0,σu2(x) the error vector (length Nt, see below for treatment of spatiotemporal autocorrelation of the errors). For precipitation, the original observations are log-transformed to reduce skewness through the function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$g({\bf{O}}({\bf{x}}))=\log ({\bf{O}}({\bf{x}})+{{\rm{O}}}_{{\rm{t}}}({\bf{x}}))$$\end{document}g(O(x))=log(O(x)+Ot(x)). The translation term is determined over the calibration years I by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\rm{O}}}_{{\rm{t}}}({\bf{x}})=1+{\min }_{t\in I}\,{\rm{O}}({\bf{x}},t)$$\end{document}Ot(x)=1+mint∈IO(x,t). This is done to make the log transformation available for zero values, while keeping the offset proportional to the precipitation magnitude to regulate the transformation effect38. In the case of temperature, the link function g is the identity function, i.e. regression is performed directly on the raw temperature anomalies. By definition, each row of the anomaly matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\bf{M}}}^{{\prime} }$$\end{document}M′ sums to zero. Consequently, the maximum rank of this matrix is Nt − 1, and decomposition through PCA can lead to identifying up to Nt − 1 eigenvectors. Usually, only a subset of them is selected, and a common approach consists in choosing the leading modes in terms of explained variance. However, selecting the eigenvectors according only to their explained variance does not necessarily correspond to a general improvement in model performances27. For this reason, in this study, the eigenvectors/scores selection is guided by a criterion based on root mean square error (RMSE), rather than just on explained variance. Initially, all modes explaining at least 10% of the total variance are included. Iteratively, for each mode with explained variance above 0.001% that was not included, a new model with the new mode adjoint is calibrated again, and a mean RMSE is computed by averaging the RMSE from the evaluation of 50 generated realizations (new samples) against the model selection dataset. The number 50 is chosen to balance computational efficiency with estimate reliability. At the end of every iteration, each mode is thus associated with one RMSE value. The eigenvector/score associated with the lowest final RMSE is added to the set of the selected modes if it leads to a RMSE reduction by at least 1%. Otherwise, the iterative process stops. This methods allows up to a total of 16 modes to be selected. After calibration and model selection, the PCR models are applied to the ESM outputs from the past2k simulations to generate downscaled climate anomaly fields at each station location (Eq. (5)). These downscaled anomalies are then adjusted to reflect the long-term climatic trends simulated by the ESM. In particular, the mean values from the downscaled estimates are corrected using the difference (for temperature; Eq. (6)) or the ratio (for precipitation; Eq. (7)) between the ESM’s 30-year moving mean and its mean over the calibration period. To obtain individual realizations at each timestep, the term u(x) needs to be explicitly modeled to provide spatiotemporal coherence in the downscaled series, as the residuals display spatial and temporal dependencies. First, their spatial correlation is modeled using a Spatial Error Model (SEM; Eq. (10)), which captures the spatial autoregressive structure that the residuals at nearby stations display. The resulting residuals are then further modeled in time using an Autoregressive Moving Average model (ARMA; Eq. (11)). This model is used to generate continuous time series of temporally correlated residuals extending across the full two-millennia reconstruction period. These temporal error series are then transformed back through the spatial error model, reintroducing the appropriate spatial correlation structure to produce timeseries with a spatiotemporal structure inferred from observations. After modeling the residuals, the downscaled values at individual stations are interpolated onto the 0.20° × 0.20° target grid. Natural neighbor interpolation is used to perform this step, as it is computationally efficient, locally adaptive, and well suited for interpolating irregularly spaced station data to a regular grid without imposing excessive smoothing. Because interpolation tends to smooth spatial variability, the resulting gridded fields generally exhibit lower variance than the original ESM fields. To correct this, the anomalies of the downscaled variables (relative to a 30-year moving mean) are rescaled to match the moving standard deviation of the ESM over the same time window (Eq. (12)). Finally, a land mask (rescaled to the target resolution39) is applied to the final downscaled products. Prediction intervals (PIs) are defined here in the statistical sense as interval estimates for a single realization of the PCR-downscaled variable. They are constructed via Monte Carlo simulations of the regression residuals, assuming Gaussian errors with variance equal to the model’s mean squared error. For each time step, random residuals are added to the fitted regression values. The resulting values are then transformed back to the physical variable scale, and the empirical 2.5th and 97.5th percentiles of the simulated distribution are taken as the bounds of the 95% PI. As with all transfer-function–based statistical downscaling approaches, the PCR framework assumes that the relationship between large-scale model modes and local-scale anomalies estimated during the calibration period (1875–2014 CE) remains valid when applied to earlier centuries. In this study, the regression is performed on anomaly fields rather than absolute values, and long-term changes in mean and variance simulated by the ESM are explicitly reintroduced through the moving mean adjustment (Eqs. (5)–(8)) and variance rescaling (Eq. (12)). Consequently, the assumption concerns the stability of the relationship between large-scale circulation modes and local anomalies, rather than stationarity of absolute climate statistics. Validation metrics Other than the PCR, a linear regression (LR) defined by using raw temperature and precipitation as predictors is also used for downscaling (LR-downscaling) for benchmarking. The LR-downscaling model is given by: 3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$g{({\bf{O}}({\bf{x}}))}^{{\prime} }={\bf{M}}({\bf{x}})\cdot {{\boldsymbol{\beta }}}_{LR}({\bf{x}})+{{\bf{u}}}_{LR}({\bf{x}})$$\end{document}g(O(x))′=M(x)⋅βLR(x)+uLR(x)with βLR the regression coefficient, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\bf{u}}}_{LR}({\bf{x}}) \sim {\mathcal{N}}(0,{\sigma }_{u,LR}^{2}({\bf{x}}))$$\end{document}uLR(x)~N(0,σu,LR2(x)) the error, whose space-time features are treated with the same methodology as for the PCR. Downscaled values at station locations are subsequently gridded to the target grid using the same procedure applied for PCR. The performance of the PCR-downscaled and LR-downscaled products are evaluated across multiple aggregation levels. Timestep-level metrics are computed by comparing the individual monthly timesteps of the downscaled products with observations from the GHCN-m testing dataset. The metrics falling within this category are: mean bias (MB; Eq. (13)), which measures whether the model underestimates or overestimates the observations; mean absolute error (MAE; Eq. (15)), which provides a measure of the overall model accuracy; Kling-Gupta efficiency (KGE; Eq. (17)), which summarizes correlation, bias, and variability into a single index and ranges between − ∞ and 1, with perfect prediction at 1. Station-level metrics are computed by first averaging both model outputs and GHCN-m observations at each station over the testing dataset, and then comparing these station-level means. The metrics in this category are: MB (Eq. (14)); MAE (Eq. (16)); Pearson correlation (r; Eq. (18)), which quantifies the spatial correlation between the station-averaged model outputs and observations. Both timestep- and station-level metrics are calculated for the full year and for individual seasons (MAM, JJA, SON, DJF). For comparison purposes, we also evaluate CHELSEA-TraCE21k19,20 using station-level metrics. This product provides only monthly climatologies averaged over centennial periods; therefore, we selected the most recent available climatology (1900–1990) and evaluated temperature and precipitation over the GHCN-m testing dataset. For each timestep-level and station-level metric, we test the null hypothesis of no difference between PCR and LR. Specifically, for the timestep-level metrics, we implemented a space–time block bootstrap procedure in order to account for both spatial and temporal dependence in the data. In this approach, observations are resampled in blocks defined by spatial clusters and temporal windows, and the difference in metrics between PCR and LR is recomputed for each bootstrap sample. Then, we estimate the two-sided bootstrap p-values for the differences in MB, MAE, and KGE between the two downscaling methods. For both temperature and precipitation, we use 1,000 bootstrap samples, 10 spatial blocks identified through k-means on the station coordinates, and temporal blocks of the size of 24 months. Instead, for the station-level metrics, statistical significance is assessed using parametric tests. For MB and MAE, we perform paired two-sided t-tests on the station-level errors (and absolute errors for MAE) obtained from PCR and LR, testing whether their mean difference is significantly different from zero. For the Pearson correlation coefficient, differences between PCR and LR are assessed using the Fisher r-to-z transformation. Finally, we calculate domain-level metrics, i.e., metrics that are computed either from gridded fields, or from spatially aggregated domain averages. Namely, we evaluate first the MB (Eq. (19)). Then, extremes in temperature and precipitation are evaluated by calculating the 10th and 90th percentiles for the modeled and the observed gridded products, and then the differences between the percentiles are computed (hereafter denoted Δ10th and Δ90th perc.). Finally, the structural similarity index metric (SSIM) is calculated between the mean modeled and observed gridded fields. SSIM is a widely used metric, originally developed for image quality assessment and more recently applied in climatic studies. For the mathematical formulation of this metric we refer to 40,41. SSIM evaluates the similarity between two images in terms of luminance, contrast and structural similarity, and ranges from 0 (no similarity) to 1 (perfect similarity). To summarize the comparative performance of the PCR- and LR-downscaling methods relative to the original ESM, we compute skill scores42. For each accuracy metric A, the skill score SSA expresses the fractional improvement (in percent) of the downscaled product relative to the ESM baseline, normalized by the difference between the ESM and a hypothetical perfect prediction. Positive values indicate improvement over the ESM, while negative values indicate degradation. Because the various performance metrics differ in their scales and units, they are combined into a Mean Skill Score (MSS; Eq. (21)) using a weighted average. Each metric is weighted according to the mean of its absolute values across all seasons and the annual evaluation, ensuring balanced contributions from all metrics. The MSS results in a percentage ranging between −∞ and 100%, which represents the percentage improvements (or degradations, when negative) that the PCR- and the LR-downscaled products achieved with respect to the total possible theoretical improvement between the original ESM and a perfect predictor. The MSS is calculated for timestep-level metrics first, and then jointly for station-level metrics and SSIM. Additional details and equations This section provides the formal mathematical expressions underlying the PCR application, long-term mean and variance adjustments, residual modeling, and validation metrics described above. Deterministic downscaling formulation After calibration and model selection, the PCR models are applied to downscale the ESM outputs from the past2k experiment. Since scores can also be obtained by projecting the anomaly matrix onto the eigenvectors (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\bf{S}}={({{\bf{M}}}^{{\rm{{\prime} }}})}^{T}\cdot {\bf{E}}$$\end{document}S=(M′)T⋅E), the estimated downscaled value \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\widehat{{\bf{O}}}({\bf{x}},t)$$\end{document}O^(x,t) would be obtained from Eq. (2) by inverting the definition of anomaly for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\widehat{{\bf{O}}}$$\end{document}O^, i.e. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$g(\widehat{{\bf{O}}}({\bf{x}}))=g(\widehat{{\bf{O}}}({\bf{x}})){\prime} +\overline{g({\bf{O}}({\bf{x}}))}$$\end{document}g(O^(x))=g(O^(x))′+g(O(x))¯, and then inverting the function g: 4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{{\bf{O}}}({\bf{x}})={g}^{-1}({({{\bf{M}}}^{{\rm{{\prime} }}})}^{T}\cdot {\bf{E}}\cdot {\boldsymbol{\beta }}({\bf{x}})+{\bf{u}}({\bf{x}})+\bar{g({\bf{O}}({\bf{x}}))})$$\end{document}O^(x)=g−1((M′)T⋅E⋅β(x)+u(x)+g(O(x))¯)However, to transfer the average climatic changes simulated by the ESMs to the downscaling model, the temporal mean of the observed variable \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{g\left(\widehat{{\bf{O}}}({\bf{x}})\right)}$$\end{document}gO^(x)¯ is adjusted by the local temporal mean of the ESM. In fact, this is the equation used to downscale the ESM simulations: 5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{{\bf{O}}}({\bf{x}})={g}^{-1}({({{\bf{M}}}^{{\rm{{\prime} }}})}^{T}\cdot {\bf{E}}\cdot {\boldsymbol{\beta }}({\bf{x}})+{\bf{u}}({\bf{x}})+\bar{g({\bf{O}}({\bf{x}}))}+{\bar{g({\bf{M}}({\bf{x}}))}}_{mov}-{\bar{g({\bf{M}}({\bf{x}}))}}_{cal})$$\end{document}O^(x)=g−1((M′)T⋅E⋅β(x)+u(x)+g(O(x))¯+g(M(x))¯mov−g(M(x))¯cal)with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\bar{\cdot }}_{mov}$$\end{document}⋅¯mov indicating the 30-year moving average, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\bar{\cdot }}_{cal}$$\end{document}⋅¯cal the average over the calibration years. For temperature, Eq. (5) becomes: 6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{{\bf{O}}}({\bf{x}})={({{\bf{M}}}^{{\rm{{\prime} }}})}^{T}\cdot {\bf{E}}\cdot {\boldsymbol{\beta }}({\bf{x}})+{\bf{u}}({\bf{x}})+\overline{{\bf{O}}({\bf{x}})}+{\overline{{\bf{M}}({\bf{x}})}}_{mov}-{\overline{{\bf{M}}({\bf{x}})}}_{cal}$$\end{document}O^(x)=(M′)T⋅E⋅β(x)+u(x)+O(x)¯+M(x)¯mov−M(x)¯cal and for precipitation: 7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{array}{ccc}\hat{{\bf{O}}}({\bf{x}}) & = & \exp ({({{\bf{M}}}^{{\rm{{\prime} }}})}^{T}\cdot {\bf{E}}\cdot {\boldsymbol{\beta }}({\bf{x}})+{\bf{u}}({\bf{x}})+\overline{\log ({\bf{O}}({\bf{x}})+{{\rm{O}}}_{{\rm{t}}}({\bf{x}}))}\\ & & +{\overline{\log ({\bf{M}}({\bf{x}})+{{\rm{M}}}_{{\rm{t}}}({\bf{x}}))}}_{mov}-{\overline{\log ({\bf{M}}({\bf{x}})+{{\rm{M}}}_{{\rm{t}}}({\bf{x}}))}}_{cal})-{{\rm{O}}}_{{\rm{t}}}({\bf{x}})\end{array}$$\end{document}O^(x)=exp((M′)T⋅E⋅β(x)+u(x)+log(O(x)+Ot(x))¯+log(M(x)+Mt(x))¯mov−log(M(x)+Mt(x))¯cal)−Ot(x) with Mt(x) defined analogously to Ot(x), i.e. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\rm{M}}}_{{\rm{t}}}({\bf{x}})=1+{\min }_{t\in I}{\rm{M}}({\bf{x}},t)$$\end{document}Mt(x)=1+mint∈IM(x,t). Equation (7) can also be expressed as: 8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{{\bf{O}}}({\bf{x}})=\frac{{\bar{{\bf{M}}({\bf{x}})+{M}_{t}({\bf{x}})}}_{geo,mov}}{{\bar{{\bf{M}}({\bf{x}})+{M}_{t}({\bf{x}})}}_{geo,cal}}\cdot {\bar{{\bf{O}}({\bf{x}})+{{\rm{O}}}_{{\rm{t}}}({\bf{x}})}}_{geo}\cdot \exp ({({{\bf{M}}}^{{\rm{{\prime} }}})}^{T}\cdot {\bf{E}}\cdot {\boldsymbol{\beta }}({\bf{x}})+{\bf{u}}({\bf{x}}))-{{\rm{O}}}_{{\rm{t}}}({\bf{x}})$$\end{document}O^(x)=M(x)+Mt(x)¯geo,movM(x)+Mt(x)¯geo,cal⋅O(x)+Ot(x)¯geo⋅exp((M′)T⋅E⋅β(x)+u(x))−Ot(x)with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\bar{\cdot }}_{geo}$$\end{document}⋅¯geo indicating the geometric mean (and the relative subscripts with the same meaning as above). In other words, the final downscaled variable \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\widehat{{\bf{O}}}({\bf{x}})$$\end{document}O^(x) accounts for and maintains the changes simulated by the ESM by adjusting the result of the downscaling model by the difference (in the case of temperature) or by the ratio (in the case of precipitation) between the moving 30-year mean and the mean over the calibration period (70 random years selected between 1875 CE and 2014 CE). Residual Modeling After defining the deterministic component of the downscaling model, we next describe the stochastic residual modeling framework used to generate spatiotemporally coherent realizations. In order to explicitly model the space-time dependency of the error term u(x), residuals are first sampled at the location x of every calibration station: 9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\bf{r}}({\bf{x}},t)=g{({\bf{O}}({\bf{x}},t))}^{{\rm{{\prime} }}}-{\mathbb{E}}(g{(\hat{{\bf{O}}}({\bf{x}},t))}^{{\rm{{\prime} }}})$$\end{document}r(x,t)=g(O(x,t))′−E(g(O^(x,t))′)Sampled residuals over calibration stations contain gaps in their timeseries. To fill these gaps, timeseries are gapfilled from other available calibration stations by thin-plate spline interpolation43. Thin-plate splines are employed here because they provide a smooth surface representation suitable for irregularly spaced station data while minimizing artificial local oscillations. As the observational network becomes sparser further back in time, complete timeseries were reconstructed using this method for the years 1915-2014, instead of the whole period 1875-2014. Then, a spatial error model (SEM) is employed to remove spatial dependency between nearby stations: 10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\bf{u}}({\bf{x}})=\lambda {\bf{W}}{\bf{u}}({\bf{x}})+{\boldsymbol{\varepsilon }}({\bf{x}})\Rightarrow {\boldsymbol{\varepsilon }}({\bf{x}})=({\bf{I}}-\lambda {\bf{W}}){\bf{u}}({\bf{x}})$$\end{document}u(x)=λWu(x)+ε(x)⇒ε(x)=(I−λW)u(x)where λ is the spatial autoregressive coefficient for the errors, W is the spatial weight matrix (neighborhood structure defined with Gaussian decay), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{\varepsilon }}({\bf{x}}) \sim {\mathcal{N}}(0,\sigma ({\bf{x}}))$$\end{document}ε(x)~N(0,σ(x)) are spatially independent errors with station-dependent variance (i.e., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{\varepsilon }} \sim {\mathcal{N}}(0,{\boldsymbol{\Sigma }})$$\end{document}ε~N(0,Σ), with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{\Sigma }}={\rm{diag}}({\sigma }_{1},...,{\sigma }_{{N}_{s}})$$\end{document}Σ=diag(σ1,...,σNs)). Ultimately, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\bf{u}} \sim {\mathcal{N}}(0,{(1-\lambda {\bf{W}})}^{-1}{\boldsymbol{\Sigma }}{(1-\lambda {\bf{W}})}^{-T})$$\end{document}u~N(0,(1−λW)−1Σ(1−λW)−T), and Σ and λ are determined via maximum likelihood estimation. The error timeseries ε(x) at each calibration station are modeled in time using an Autoregressive Moving Average model with p autoregressive and q moving average terms (ARMA(p, q)): 11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varepsilon ({\bf{x}},t)=\mathop{\sum }\limits_{i=1}^{p}{\varphi }_{i}\,\varepsilon ({\bf{x}},t-i)+\eta ({\bf{x}},t)+\mathop{\sum }\limits_{j=1}^{q}{\theta }_{j}\,\eta ({\bf{x}},t-j)$$\end{document}ε(x,t)=∑i=1pφiε(x,t−i)+η(x,t)+∑j=1qθjη(x,t−j)with φi and θj the autoregressive and moving average coefficients, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\eta ({\bf{x}},t) \sim {\mathcal{N}}\left(0,{\sigma }^{2}({\bf{x}})\right)$$\end{document}η(x,t)~N0,σ2(x) the i.i.d. Gaussian innovation at grid node x and timestep t. Predictions for each calibration station location for the past are initialized assuming ε(x, 0) = 0 and η(x, 0) = 0. Finally, the ε(x, t) values are transformed back through the SEM to obtain the modeled residuals across the two millennia for every calibration station. For temperature, p = 1 and q = 2, while for precipitation p = 1 and q = 1. These values were determined by identifying the lag at which the median partial autocorrelation and autocorrelation across timeseries ε(x) crossed the 0.05 threshold. Gridding and variance adjustment Once residuals are simulated and combined with the deterministic component, the resulting station-wise reconstructions are spatially interpolated and variance-adjusted as described below. Station-wise downscaled values are interpolated onto the 0.2° × 0.2° target grid using natural neighbor interpolation, maintaining computational feasibility over a large domain. The gridded maps thus produced display different (typically lower) variability than the original ESM. Therefore, the final gridded values (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{{\bf{O}}}}_{g}$$\end{document}O^g) are obtained by scaling the variable anomalies (relative to the 30-year moving average) by the 30-year moving standard deviation σmov at each node: 12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\widehat{{{\bf{O}}}_{g}}({\bf{x}})={\overline{\widehat{{\bf{O}}}({\bf{x}})}}_{mov}+(\widehat{{\bf{O}}}({\bf{x}})-{\overline{\widehat{{\bf{O}}}({\bf{x}})}}_{mov})\cdot \frac{{\sigma }_{mov}({\bf{M}}({\bf{x}})-{\overline{{\bf{M}}({\bf{x}})}}_{mov})}{{\sigma }_{mov}(\widehat{{{\bf{O}}}_{g}}({\bf{x}})-{\overline{\widehat{{\bf{O}}}({\bf{x}})}}_{mov})}$$\end{document}Og^(x)=O^(x)¯mov+(O^(x)−O^(x)¯mov)⋅σmov(M(x)−M(x)¯mov)σmov(Og^(x)−O^(x)¯mov) Validation metrics equations Finally, we formally define the validation metrics used to assess model performance. The timestep-level mean bias (MB) is defined as: 13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$MB=\frac{{\sum }_{i=1}^{\widetilde{N}}\left({\widehat{{\rm{O}}}}_{i}-{{\rm{O}}}_{i}\right)}{\widetilde{N}}$$\end{document}MB=∑i=1N~O^i−OiN~with Oi and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{{\rm{O}}}}_{i}$$\end{document}O^i indicating the individual observed and the estimated values, respectively, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\widetilde{N}$$\end{document}N~ the total number of samples in the testing dataset (for all considered stations and timesteps). The station-level MB formulation instead is the following: 14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$MB=\frac{{\sum }_{i=1}^{{\widetilde{N}}_{s}}\left(\bar{\widehat{{\bf{O}}}}({{\bf{x}}}_{{\bf{i}}})-\bar{{\bf{O}}}({{\bf{x}}}_{{\bf{i}}})\right)}{{\widetilde{N}}_{s}}$$\end{document}MB=∑i=1N~sO^¯(xi)−O¯(xi)N~swith \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bar{\cdot }$$\end{document}⋅¯ indicating the temporal average, xi the station position, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widetilde{N}}_{s}$$\end{document}N~s the number of stations considered. The timestep-level mean absolute error (MAE) is defined as: 15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$MAE=\frac{{\sum }_{i=1}^{\widetilde{N}}\left|{{\rm{O}}}_{i}-{\widehat{{\rm{O}}}}_{i}\right|}{\widetilde{N}}$$\end{document}MAE=∑i=1N~Oi−O^iN~with variables defined as in the MB formula. The station-level MAE is defined analogously to Eq. (14): 16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$MAE=\frac{{\sum }_{i=1}^{{\widetilde{N}}_{s}}\left|\bar{\widehat{{\bf{O}}}}({{\bf{x}}}_{{\bf{i}}})-\bar{{\bf{O}}}({{\bf{x}}}_{{\bf{i}}})\right|}{{\widetilde{N}}_{s}}$$\end{document}MAE=∑i=1N~sO^¯(xi)−O¯(xi)N~s The (timestep-level) Kling-Gupta efficiency (KGE) is a global indicator summarizing correlation, variability, and bias, and is given by: 17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$KGE=1-\sqrt{{(\rho -1)}^{2}+{(\alpha -1)}^{2}+{(\beta -1)}^{2}}$$\end{document}KGE=1−(ρ−1)2+(α−1)2+(β−1)2with ρ being the timestep-level Pearson correlation coefficient, α the ratio of the modeled to observed standard deviations, and β the ratio of modeled to observed means. The KGE ranges in ( − ∞, 1], with KGE=1 indicating perfect performance. The (station-level) Pearson correlation (r) is given by: 18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r=\frac{{\rm{Cov}}\left(\bar{\widehat{{\bf{O}}}}({\bf{x}}),\bar{{\bf{O}}}({\bf{x}})\right)}{\sqrt{{\rm{Var}}\left(\bar{\widehat{{\bf{O}}}}({\bf{x}})\right)\,{\rm{Var}}\left(\bar{{\bf{O}}}({\bf{x}})\right)}}$$\end{document}r=CovO^¯(x),O¯(x)VarO^¯(x)VarO¯(x)with Cov(⋅ , ⋅) and Var(⋅) indicating the spatial covariance and variance computed across all considered stations. For the evaluation of the gridded products, the GHCN-m dataset is gridded at every timestep in the same way as the downscaled product, over the 1875-2014 period and using only the testing dataset. If enough stations (more than 50) are available at a given timestep, the corresponding observed testing map is produced. Otherwise, no observed map is generated for that timestep. The MB in a grid node x is defined as: 19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$MB({\bf{x}})=\frac{{\sum }_{i\in {I}_{t}}\left({\widehat{{\bf{O}}}}_{g}({\bf{x}},{t}_{i})-{\bf{O}}({\bf{x}},{t}_{i})\right)}{{\widetilde{N}}_{t}}$$\end{document}MB(x)=∑i∈ItO^g(x,ti)−O(x,ti)N~twith \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${I}_{t}=1,...,{\widetilde{N}}_{t}$$\end{document}It=1,...,N~t the number of timesteps for which there are enough stations to grid observations effectively. Extremes in temperature and precipitation are assessed by computing the differences in the 10th and 90th percentiles. These percentiles are calculated separately for the modeled and observed gridded products, and then the differences between the percentiles are computed (hereafter denoted as Δ10th perc. and Δ90th perc.). Similarly, the gridded DPastCliM-NA temperature and precipitation fields are also evaluated against ERA5 monthly output44 for all available, overlapping years (1940–2014). To this end, ERA5 is regridded to the same resolution as DPastCliM-NA by natural neighbor interpolation. The MB is computed as in Eq. (19), but with O(x, ti) now referring to the ERA5 product. A skill score for a certain accuracy metric A42 can be defined as: 20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S{S}_{A}=100\cdot \frac{{A}_{t}-{A}_{r}}{{A}_{p}-{A}_{r}}$$\end{document}SSA=100⋅At−ArAp−Ar where At is the accuracy of the target prediction (PCR- or LR-downscaled product), Ar is the accuracy of the reference prediction (ESM), and Ap is the accuracy of a perfect prediction. As different metrics have different scales, their contributions to the Mean Skill Score (MSS) need to be weighted. The weight (wA) for each metric A is given by the mean of all absolute values of that metric across the full year and the four seasons. Then, the MSS for the PCR-downscaled product compared to the original ESM is defined as: 21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$MSS=\frac{{\sum }_{A\in M}{w}_{A}\cdot S{S}_{A}}{{\sum }_{A\in M}{w}_{A}}$$\end{document}MSS=∑A∈MwA⋅SSA∑A∈MwAwhere M are the considered metrics. The timestep-level MSS is calculated using M = {MB, MAE, KGE}, and the station-level MSS is calculated from M = {MB, MAE, r, SSIM}. The MSS for the LR-downscaled product is defined analogously. The MSS results in a percentage ranging between −∞ and 100%, representing the percentage improvements (or degradations, when negative) that the PCR- and LR-downscaled products achieved with respect to the total possible theoretical improvement between a perfect predictor and the original ESM. Data Records The DPastCliM-NA dataset26 is publicly available on Zenodo (https://doi.org/10.5281/zenodo.17245353) and contains station-wise and gridded climate reconstructions over North America. The following products are provided:_metadata_table_NA_020.csv: Station metadata table including the station ID, latitude (°N), longitude (°E), elevation (m), and flags indicating their use in calibration (flag_cal), model selection (flag_ms), or testing (flag_test)._MPI-ESM-1-2-LR_past2k_statreal_PCRdownscaled_NA_020.nc: NetCDF-4 file containing the DPastCliM-NA station-wise realizations. These files contain two main variables:Ods_hat_mat — reconstructed time series at each calibration station (matrix of size station × timestep, with rows in the same order as the metadata table). Stations that were not reconstructed (testing-only stations) have NaN rows.PI_mat— 95% prediction intervals associated with each reconstructed time series (matrix of size station × timestep × bound, with lower/upper bounds)._MPI-ESM-1-2-LR__PCRdownscaled_NA_020.zip: Archive containing multiple NetCDF-4 files with monthly gridded fields from the station-wise realizations. Each file covers a 100-year period, and file suffixes (e.g., 1-1200) indicate the corresponding range of timesteps included._MPI-ESM-1-2-LR__PI_PCRdownscaled_NA_020>.zip: Archive containing the corresponding gridded 95% prediction intervals for the monthly fields in the previous item. _metadata_table_NA_020.csv: Station metadata table including the station ID, latitude (°N), longitude (°E), elevation (m), and flags indicating their use in calibration (flag_cal), model selection (flag_ms), or testing (flag_test). _metadata_table_NA_020.csv: Station metadata table including the station ID, latitude (°N), longitude (°E), elevation (m), and flags indicating their use in calibration (flag_cal), model selection (flag_ms), or testing (flag_test). _MPI-ESM-1-2-LR_past2k_statreal_PCRdownscaled_NA_020.nc: NetCDF-4 file containing the DPastCliM-NA station-wise realizations. These files contain two main variables:Ods_hat_mat — reconstructed time series at each calibration station (matrix of size station × timestep, with rows in the same order as the metadata table). Stations that were not reconstructed (testing-only stations) have NaN rows.PI_mat— 95% prediction intervals associated with each reconstructed time series (matrix of size station × timestep × bound, with lower/upper bounds). _MPI-ESM-1-2-LR_past2k_statreal_PCRdownscaled_NA_020.nc: NetCDF-4 file containing the DPastCliM-NA station-wise realizations. These files contain two main variables:Ods_hat_mat — reconstructed time series at each calibration station (matrix of size station × timestep, with rows in the same order as the metadata table). Stations that were not reconstructed (testing-only stations) have NaN rows.PI_mat— 95% prediction intervals associated with each reconstructed time series (matrix of size station × timestep × bound, with lower/upper bounds). Ods_hat_mat — reconstructed time series at each calibration station (matrix of size station × timestep, with rows in the same order as the metadata table). Stations that were not reconstructed (testing-only stations) have NaN rows. Ods_hat_mat — reconstructed time series at each calibration station (matrix of size station × timestep, with rows in the same order as the metadata table). Stations that were not reconstructed (testing-only stations) have NaN rows. PI_mat— 95% prediction intervals associated with each reconstructed time series (matrix of size station × timestep × bound, with lower/upper bounds). PI_mat— 95% prediction intervals associated with each reconstructed time series (matrix of size station × timestep × bound, with lower/upper bounds). _MPI-ESM-1-2-LR__PCRdownscaled_NA_020.zip: Archive containing multiple NetCDF-4 files with monthly gridded fields from the station-wise realizations. Each file covers a 100-year period, and file suffixes (e.g., 1-1200) indicate the corresponding range of timesteps included. _MPI-ESM-1-2-LR__PCRdownscaled_NA_020.zip: Archive containing multiple NetCDF-4 files with monthly gridded fields from the station-wise realizations. Each file covers a 100-year period, and file suffixes (e.g., 1-1200) indicate the corresponding range of timesteps included. _MPI-ESM-1-2-LR__PI_PCRdownscaled_NA_020>.zip: Archive containing the corresponding gridded 95% prediction intervals for the monthly fields in the previous item. _MPI-ESM-1-2-LR__PI_PCRdownscaled_NA_020>.zip: Archive containing the corresponding gridded 95% prediction intervals for the monthly fields in the previous item. Data Overview Figure 4 shows the DPastCliM-NA mean temperature and precipitation fields across the two-millennia period, the mean 95% prediction interval (PI), and the corresponding anomalies over three different time windows (0-100, 950-1050, 1900-2000 CE). For temperature, uncertainty appears to follow a latitudinal gradient, with the PI being the largest at higher latitudes. For precipitation, the 95% PI (Fig. 4m) is proportional to the precipitation intensity, i.e. the associated uncertainties are represented in proportion to variability and magnitude of the climatological values. Figure 5 showcases monthly temperature and precipitation time series in four representative stations over three selected time windows, located near sites with paleoclimatic proxy records (Fig. 2). Two stations (tas-1 and pr-1) are close to Harvard Forest in Massachusetts (US), where multiproxy reconstruction provided evidence of variability in temperature and precipitation4. The other two stations (tas-2 and pr-2) are located in the Southwestern US, a well-studied region with tree-ring records documenting past megadroughts45.Fig. 4DPastCliM-NA absolute mean fields and mean 95% PI across the two-millennia period, and the corresponding anomalies over three different time windows (0-100, 950-1050, 1900-2000 CE) for temperature (tas; a–h) and precipitation (pr; i–p).Fig. 5Monthly mean values of temperature and precipitation in four example stations (starred locations in Fig. 2) over three time windows for the PCR-downscaled product (blue line), for the original ESM (orange line) and for the GHCN-m dataset (black line). The grey area is the 95% Prediction Interval for the individual monthly realizations. The dashed line in each panel indicates the 30-year mean centered in the time window, while the anomaly indicates its difference with respect to the 0-2014 CE mean (2k mean). DPastCliM-NA absolute mean fields and mean 95% PI across the two-millennia period, and the corresponding anomalies over three different time windows (0-100, 950-1050, 1900-2000 CE) for temperature (tas; a–h) and precipitation (pr; i–p). Monthly mean values of temperature and precipitation in four example stations (starred locations in Fig. 2) over three time windows for the PCR-downscaled product (blue line), for the original ESM (orange line) and for the GHCN-m dataset (black line). The grey area is the 95% Prediction Interval for the individual monthly realizations. The dashed line in each panel indicates the 30-year mean centered in the time window, while the anomaly indicates its difference with respect to the 0-2014 CE mean (2k mean). Technical Validation Evaluation Against GHCN-m station data Table 1 summarizes the timestep-level performance of the PCR-downscaled (DPastCliM-NA), LR-downscaled, and original ESM products compared with the GHCN-m testing dataset. Metrics were calculated on individual observations using the testing dataset and are presented for the entire year as well as for individual seasons.Table 1Timestep-level metrics for temperature (tas) and precipitation (pr) for Principal Component Regression (PCR), Linear Regression (LR), and the MPI-ESM1.2-LR Earth System Model (ESM).PeriodMetrictasprpPCRLRESMpPCRLRESMYear-roundMB0.5000.160.14− 1.090.6040.100.090.38MAE0.4202.202.212.930.0181.351.331.74KGE0.1120.950.950.880.5680.570.570.32MAMMB0.6120.380.36− 1.320.0220.060.030.63MAE0.6142.262.252.830.1841.271.261.76KGE0.5460.910.910.830.5860.540.540.27JJAMB0.0880.20.24− 0.820.6920.540.540.27MAE0.0141.831.862.460.9081.511.501.81KGE0.3580.830.830.770.3420.530.520.36SONMB0.6860.310.32− 0.480.8720.140.150.08MAE0.7302.072.062.680.1181.421.41.72KGE0.2340.920.920.860.5520.560.550.31DJFMB0.016− 0.27− 0.36− 1.740.7360.190.190.62MAE0.3902.642.663.780.0161.191.171.68KGE0.0560.660.56− 1.080.0320.600.630.3Metrics are: MB = Mean Bias, MAE = Mean Absolute Error, and KGE = Kling-Gupta Efficiency. Seasons are denoted as MAM (March–May), JJA (June–August), SON (September–November), and DJF (December–February). Reported p-values refer to tests assessing whether the metrics for PCR and LR differ significantly; values with p < 0.05 are shown in bold. Table 1Timestep-level metrics for temperature (tas) and precipitation (pr) for Principal Component Regression (PCR), Linear Regression (LR), and the MPI-ESM1.2-LR Earth System Model (ESM).PeriodMetrictasprpPCRLRESMpPCRLRESMYear-roundMB0.5000.160.14− 1.090.6040.100.090.38MAE0.4202.202.212.930.0181.351.331.74KGE0.1120.950.950.880.5680.570.570.32MAMMB0.6120.380.36− 1.320.0220.060.030.63MAE0.6142.262.252.830.1841.271.261.76KGE0.5460.910.910.830.5860.540.540.27JJAMB0.0880.20.24− 0.820.6920.540.540.27MAE0.0141.831.862.460.9081.511.501.81KGE0.3580.830.830.770.3420.530.520.36SONMB0.6860.310.32− 0.480.8720.140.150.08MAE0.7302.072.062.680.1181.421.41.72KGE0.2340.920.920.860.5520.560.550.31DJFMB0.016− 0.27− 0.36− 1.740.7360.190.190.62MAE0.3902.642.663.780.0161.191.171.68KGE0.0560.660.56− 1.080.0320.600.630.3Metrics are: MB = Mean Bias, MAE = Mean Absolute Error, and KGE = Kling-Gupta Efficiency. Seasons are denoted as MAM (March–May), JJA (June–August), SON (September–November), and DJF (December–February). Reported p-values refer to tests assessing whether the metrics for PCR and LR differ significantly; values with p < 0.05 are shown in bold. Timestep-level metrics for temperature (tas) and precipitation (pr) for Principal Component Regression (PCR), Linear Regression (LR), and the MPI-ESM1.2-LR Earth System Model (ESM). Metrics are: MB = Mean Bias, MAE = Mean Absolute Error, and KGE = Kling-Gupta Efficiency. Seasons are denoted as MAM (March–May), JJA (June–August), SON (September–November), and DJF (December–February). Reported p-values refer to tests assessing whether the metrics for PCR and LR differ significantly; values with p < 0.05 are shown in bold. The DPastCliM-NA product shows consistent improvements across all metrics and for both variables compared to the original ESM. For temperature, the largest improvements are observed in the MB. In particular, the MB reaches negligible values between − 0.27 and 0.38 °C, depending on the season, whereas the ESM shows MB values between −1.74 and −0.48 °C. This demonstrates the ability of the PCR method to correct the systematic cold bias of the ESM, leading to nearly unbiased temperatures on average. MAE and KGE are also improved (1.83–2.26 °C and 0.66–0.95), though to a lesser extent. For precipitation, improvements are also evident, although the original ESM already performs reasonably well in many aspects. KGE appears to benefit the most from PCR-downscaling, followed by MB and MAE. On the annual scale, KGE improves by 0.25, and MB is reduced by 0.28 mm day−1. Seasonally, PCR shows consistent performance across all four seasons, with no single season dominating. However, DPastCliM-NA temperature performs particularly well during DJF and MAM. In winter, KGE is substantially improved compared to the ESM, indicating that both mean values and variability align more closely with observations. Precipitation also improves consistently across seasons, with the exception of JJA, which exhibits a slightly wetter MB compared to MPI-ESM1.2-LR. Table 2 reports the station-level metrics for the downscaled and original ESM products, again evaluated against the GHCN-m testing dataset. Unlike the timestep-level metrics, these are computed by comparing station-averaged values rather than individual timesteps.Table 2Station-level metrics and Structural Similarity Index Metric (SSIM) for temperature (tas) and precipitation (pr) for Principal Component Regression (PCR), Linear Regression (LR), and the MPI-ESM1.2-LR Earth System Model (ESM).PeriodMetrictasprpPCRLRESMpPCRLRESMYear-roundr<0.0010.9980.9960.9410.1610.9820.9810.583MB0.2010.1490.137−1.0700.3340.0960.0930.361MAE<0.0010.3140.3451.6990.4850.1920.1940.804SSIM—0.9070.8470.786—0.9460.9450.88MAMr<0.0010.9930.9900.9310.8310.9580.9570.543MB0.2980.3640.350−1.290<0.0010.0590.0290.617MAE<0.0010.5930.6571.9440.9890.2920.2921.038SSIM—0.8990.8850.782—0.8750.8740.838JJAr<0.0010.990.9850.876<0.0010.9690.9620.660MB0.0020.1930.234−0.8100.6910.0060.0080.155MAE<0.0010.4510.5111.826<0.0010.2990.3280.878SSIM—0.8850.8740.875—0.9070.9040.858SONr<0.0010.9930.990.935<0.0010.9470.9380.601MB0.6620.2980.304−0.4800.2450.1370.1450.072MAE<0.0010.5320.5891.6330.0020.3810.3980.869SSIM—0.8880.7940.747—0.9240.9280.882DJFr<0.0010.9960.9950.9560.0450.9720.9730.655MB<0.001−0.270-0.340−1.710.2500.1830.1900.598MAE<0.0010.5810.6652.3340.1760.3130.3071.062SSIM—0.9210.9110.889—0.8320.8200.824Station-level metrics are: r=Pearson correlation, MB = Mean Bias, MAE = Mean Absolute Error. Seasons are denoted as MAM (March–May), JJA (June–August), SON (September–November), and DJF (December–February). Reported p-values refer to tests assessing whether the metrics for PCR and LR differ significantly; values with p < 0.05 are shown in bold. Table 2Station-level metrics and Structural Similarity Index Metric (SSIM) for temperature (tas) and precipitation (pr) for Principal Component Regression (PCR), Linear Regression (LR), and the MPI-ESM1.2-LR Earth System Model (ESM).PeriodMetrictasprpPCRLRESMpPCRLRESMYear-roundr<0.0010.9980.9960.9410.1610.9820.9810.583MB0.2010.1490.137−1.0700.3340.0960.0930.361MAE<0.0010.3140.3451.6990.4850.1920.1940.804SSIM—0.9070.8470.786—0.9460.9450.88MAMr<0.0010.9930.9900.9310.8310.9580.9570.543MB0.2980.3640.350−1.290<0.0010.0590.0290.617MAE<0.0010.5930.6571.9440.9890.2920.2921.038SSIM—0.8990.8850.782—0.8750.8740.838JJAr<0.0010.990.9850.876<0.0010.9690.9620.660MB0.0020.1930.234−0.8100.6910.0060.0080.155MAE<0.0010.4510.5111.826<0.0010.2990.3280.878SSIM—0.8850.8740.875—0.9070.9040.858SONr<0.0010.9930.990.935<0.0010.9470.9380.601MB0.6620.2980.304−0.4800.2450.1370.1450.072MAE<0.0010.5320.5891.6330.0020.3810.3980.869SSIM—0.8880.7940.747—0.9240.9280.882DJFr<0.0010.9960.9950.9560.0450.9720.9730.655MB<0.001−0.270-0.340−1.710.2500.1830.1900.598MAE<0.0010.5810.6652.3340.1760.3130.3071.062SSIM—0.9210.9110.889—0.8320.8200.824Station-level metrics are: r=Pearson correlation, MB = Mean Bias, MAE = Mean Absolute Error. Seasons are denoted as MAM (March–May), JJA (June–August), SON (September–November), and DJF (December–February). Reported p-values refer to tests assessing whether the metrics for PCR and LR differ significantly; values with p < 0.05 are shown in bold. Station-level metrics and Structural Similarity Index Metric (SSIM) for temperature (tas) and precipitation (pr) for Principal Component Regression (PCR), Linear Regression (LR), and the MPI-ESM1.2-LR Earth System Model (ESM). Station-level metrics are: r=Pearson correlation, MB = Mean Bias, MAE = Mean Absolute Error. Seasons are denoted as MAM (March–May), JJA (June–August), SON (September–November), and DJF (December–February). Reported p-values refer to tests assessing whether the metrics for PCR and LR differ significantly; values with p < 0.05 are shown in bold. At this aggregation level, the DPastCliM-NA product shows even larger improvements relative to the ESM. This is expected, as comparisons at the timestep level are more affected by temporal mismatches between model and observations due to internal climate variability, whereas station-level averages emphasize spatial patterns and long-term climatological agreement. For temperature, the PCR-downscaled product achieves r > 0.99 annually and across all seasons. Consistent with the timestep-level results, the station-level MB is reduced to near-zero values. In contrast to the timestep-level MAE, substantial improvements are also observed in station-level MAE, which decreases from 1.7–2.3 °C to a much narrower range of 0.3–0.6 °C. Similar behavior is found for precipitation, with marked improvements in spatial correlation, increasing from 0.54–0.66 to 0.95–0.98. Also for this variable, both MB and MAE are strongly reduced and become nearly negligible. Table 2 also reports the SSIM, which, rather than being a station-level metric, is more appropriately interpreted as a domain-level (gridded) metric. In this case as well, substantial improvements over the ESM are observed for both variables and across all seasons, with the exception of JJA for temperature and DJF for precipitation, where improvements are more limited. Furthermore, we conduct a comparative evaluation of station-level metrics for temperature and precipitation from the CHELSA-TraCE21k product19,20 over the period 1900–1990 (Table 3). For temperature, DPastCliM-NA shows improved MB and MAE across all seasons, while exhibiting an overall comparable correlation coefficient. Instead, for precipitation, CHELSA-TraCE21k shows a slightly better MB; however, DPastCliM-NA performs better in terms of correlation and MAE. Table 3Station-level metrics (MB, MAE and Pearson correlation) for temperature and precipitation from CHELSEA-TraCE21k, evaluated against the GHCN-m testing dataset over the period 1900-1990.SeasontasprMBMAErMBMAErAnnual− 0.9191.3870.990− 0.0150.6140.878MAM− 0.9121.3860.9810.0060.5570.859JJA− 0.7561.1940.9560.0180.680.865SON− 0.7671.2130.986− 0.1000.6590.88DJF− 1.241.7550.9810.0160.5600.893 Table 3Station-level metrics (MB, MAE and Pearson correlation) for temperature and precipitation from CHELSEA-TraCE21k, evaluated against the GHCN-m testing dataset over the period 1900-1990.SeasontasprMBMAErMBMAErAnnual− 0.9191.3870.990− 0.0150.6140.878MAM− 0.9121.3860.9810.0060.5570.859JJA− 0.7561.1940.9560.0180.680.865SON− 0.7671.2130.986− 0.1000.6590.88DJF− 1.241.7550.9810.0160.5600.893 Station-level metrics (MB, MAE and Pearson correlation) for temperature and precipitation from CHELSEA-TraCE21k, evaluated against the GHCN-m testing dataset over the period 1900-1990. Tables 4 and 5 report the mean skill score (MSS; Eq. (21)) corresponding to the timestep- and station-level metrics, respectively, for both the PCR- and LR-downscaled products. This score quantifies the fraction of the maximum theoretically achievable improvement from the original ESM to a perfect predictor.Table 4Mean skill score associated with the timestep-level metrics (MSS; Eq. (21)) for temperature (tas) and precipitation (pr) for PCR- and LR-downscaled products.PeriodtasprMSS(PCR)MSS(LR)MSS(PCR)MSS(LR)Year-round40.941.031.732.8MAM33.433.837.438.3JJA32.831.34.74.7SON29.029.010.49.3DJF49.847.637.038.7Values indicate the percentage of the maximum theoretically achievable improvement from the original MPI-ESM1.2-LR product to a perfect predictor. Table 4Mean skill score associated with the timestep-level metrics (MSS; Eq. (21)) for temperature (tas) and precipitation (pr) for PCR- and LR-downscaled products.PeriodtasprMSS(PCR)MSS(LR)MSS(PCR)MSS(LR)Year-round40.941.031.732.8MAM33.433.837.438.3JJA32.831.34.74.7SON29.029.010.49.3DJF49.847.637.038.7Values indicate the percentage of the maximum theoretically achievable improvement from the original MPI-ESM1.2-LR product to a perfect predictor. Mean skill score associated with the timestep-level metrics (MSS; Eq. (21)) for temperature (tas) and precipitation (pr) for PCR- and LR-downscaled products. Values indicate the percentage of the maximum theoretically achievable improvement from the original MPI-ESM1.2-LR product to a perfect predictor. Consistent with the improvements observed in the individual timestep-level metrics (Table 1), the MSS indicates that, at the timestep aggregation level, the PCR and LR approaches perform very similarly, yielding nearly identical improvements over the ESM. MSS values for both variables vary by season. For temperature, values range from approximately 29% (SON) to 50% (DJF). As expected, improvements for precipitation are smaller, ranging from 5% (JJA) to about 37–39% (MAM and DJF). Comparatively, values of the station-level MSS are much larger (Table 5). Here, the improvements of the DPastCliM-NA over the ESM become more apparent, with MSS values ranging from 48% (SON precipitation) to 80% (annual temperature). Similarly to the timestep-level MSS, improvements for temperature (63–80%) exceed those for precipitation (48–75%). However, in this case, the PCR approach yields systematically larger improvements than LR for temperature, with gains of 3.7–12.1%. For precipitation, the additional improvements provided by PCR relative to LR are negligible, reaching only 2.3%.Table 5Mean skill score associated with the station-level metrics (MSS; Eq. (21)) for temperature (tas) and precipitation (pr) for PCR- and LR-downscaled products.PeriodtasprMSS(PCR)MSS(LR)MSS(PCR)MSS(LR)Year-round80.271.775.074.7MAM71.768.062.062.1JJA63.057.865.563.2SON66.053.948.247.4DJF69.364.553.851.5Values indicate the percentage of the maximum theoretically achievable improvement from the original MPI-ESM1.2-LR product to a perfect predictor. Table 5Mean skill score associated with the station-level metrics (MSS; Eq. (21)) for temperature (tas) and precipitation (pr) for PCR- and LR-downscaled products.PeriodtasprMSS(PCR)MSS(LR)MSS(PCR)MSS(LR)Year-round80.271.775.074.7MAM71.768.062.062.1JJA63.057.865.563.2SON66.053.948.247.4DJF69.364.553.851.5Values indicate the percentage of the maximum theoretically achievable improvement from the original MPI-ESM1.2-LR product to a perfect predictor. Mean skill score associated with the station-level metrics (MSS; Eq. (21)) for temperature (tas) and precipitation (pr) for PCR- and LR-downscaled products. Values indicate the percentage of the maximum theoretically achievable improvement from the original MPI-ESM1.2-LR product to a perfect predictor. The differences between the two approaches are further assessed by testing whether the corresponding metrics from PCR and LR differ significantly (Table 1). At the timestep level, the two downscaling methods generally do not differ significantly; even when statistical significance is detected, the magnitude of the differences remains small and of limited physical relevance. In contrast, clearer differences emerge at the station level. For temperature, the PCR approach performs significantly better than LR annually and across all seasons in terms of r and MAE, while MB shows significant improvements in DJF and JJA. For precipitation, differences are less robust: although some metrics indicate statistically significant differences, their magnitude remains modest. Figures 6a,b and 7a,b present the mean bias (MB) for the PCR-downscaled product and the original MPI-ESM1.2-LR output over the testing dataset. The results show an improved MB for both temperature and precipitation across the continent (Figs. 6a, 7a), especially in data-rich areas such as the US and Canada, while improvements are less evident in the southern regions. The MB reaches uniformly low values for both temperature (within 1°C) and precipitation (within 0.5 mm day−1). The bias correction is particularly notable along the western coast, where the original ESM shows the largest biases. Here, temperature MB is reduced after PCR-downscaling, although some areas still exhibit a larger, mostly warm bias. Precipitation bias correction is extensive, with few areas displaying absolute bias larger than 0.5 mm day−1.Fig. 6Mean bias of temperature (a,b) and differences in the 10th (c,d) and 90th percentiles (e,f) over the testing dataset between PCR-downscaled or original ESM temperature fields and the GHCN-m dataset (Δ= Model - Obs). The testing period covers 1875–2014 (see Section 2.3). Numbers at the bottom of each panel indicate the domain spatial average ± standard deviation.Fig. 7Same as Fig. 6, but for precipitation. Mean bias of temperature (a,b) and differences in the 10th (c,d) and 90th percentiles (e,f) over the testing dataset between PCR-downscaled or original ESM temperature fields and the GHCN-m dataset (Δ= Model - Obs). The testing period covers 1875–2014 (see Section 2.3). Numbers at the bottom of each panel indicate the domain spatial average ± standard deviation. Same as Fig. 6, but for precipitation. Although the downscaling procedure did not explicitly target extremes, improvements in distributional percentiles are observed when comparing DPastCliM-NA and MPI-ESM1.2-LR with GHCN-m. For temperature, DPastCliM-NA fields inherit the percentile structure of the original ESM but with reduced discrepancies and more accurate estimates (Fig. 6c–e). Similar improvements occur for precipitation, with percentile differences generally ranging between − 0.5 and 0.5 mm day−1 (Fig. 7c–e). Despite these improvements, the North American Cordillera remains a region of systematic bias in the extremes for both variables. Particularly, the cold and warm biases inherited from the original ESM are predominantly found along this mountain range. For precipitation, dry extremes are represented correctly on average in the ESM, while DPastCliM-NA carries slightly drier estimates for low-precipitation months in this region. On the other hand, the substantial overestimation of the 90th percentile precipitation in the ESM is effectively corrected in DPastCliM-NA across the continent. A limitation inherent to this approach is the assumption that the statistical relationships calibrated over the instrumental period remain valid under different background climate states. While the anomaly-based formulation and explicit preservation of ESM-simulated changes in mean and variance reduce sensitivity to shifts in baseline climate, potential changes in large-scale circulation dynamics over the past two millennia could introduce non-stationarities that are not explicitly modeled. This assumption should be considered when interpreting long-term reconstructions. Evaluation Against ERA5 and Assessment of Spatial Structure To further assess the spatial completeness of DPastCliM in regions with sparse station coverage–particularly at high latitudes, in the southern part of the continent, and in topographically complex areas–we conduct an additional comparison against the ERA5 reanalysis over the modern reference period. ERA5 provides spatially continuous fields and therefore offers an independent benchmark for evaluating large-scale structure and regional biases. For near-surface temperature, DPastCliM-NA shows an overall positive bias relative to ERA5 (Fig. 8a), with an annual mean bias of about +0.5°C (Table 6). The warm bias exhibits a clear seasonal cycle, being strongest in SON (+0.7°C) and DJF (+0.6°C) and weakest in JJA (+0.2°C). The largest positive biases occur over topographically complex terrain in correspondence to the North America Cordillera, particularly in the northwestern part of the continent. Negative biases are more localized and occur mainly in parts of the northern Mexico, and some interior continental regions during the warm season.Fig. 8DPastCliM-NA annual mean bias for temperature (a) and for precipitation (b) against ERA5 fields (DPastCliM-NA minus ERA5) for the period 1940-2014. Numbers at the bottom of each panel indicate the domain spatial average ± standard deviation.Table 6DPastCliM-NA annual and seasonal MB for temperature and precipitation against ERA5 fields for the period 1940-2014.SeasontasprAnnual0.48 ± 1− 0.33 ± 1.04DJF0.61 ± 1.2− 0.12 ± 0.94MAM0.44 ± 1.13− 0.34 ± 1.09JJA0.22 ± 1.19− 0.44 ± 1.49SON0.67 ± 1.18− 0.43 ± 1.18 DPastCliM-NA annual mean bias for temperature (a) and for precipitation (b) against ERA5 fields (DPastCliM-NA minus ERA5) for the period 1940-2014. Numbers at the bottom of each panel indicate the domain spatial average ± standard deviation. Table 6DPastCliM-NA annual and seasonal MB for temperature and precipitation against ERA5 fields for the period 1940-2014.SeasontasprAnnual0.48 ± 1− 0.33 ± 1.04DJF0.61 ± 1.2− 0.12 ± 0.94MAM0.44 ± 1.13− 0.34 ± 1.09JJA0.22 ± 1.19− 0.44 ± 1.49SON0.67 ± 1.18− 0.43 ± 1.18 DPastCliM-NA annual and seasonal MB for temperature and precipitation against ERA5 fields for the period 1940-2014. For precipitation, DPastCliM-NA exhibits an overall modest dry bias across the study domain (Fig. 8b), with an annual mean bias of approximately −0.3 mm day−1. The dry tendency strengthens during the warm season. At the seasonal scale, the bias is weakest in DJF (about −0.1 mm day−1) and increases in JJA and SON (about −0.4 mm day−1). Spatially, negative precipitation biases are widespread over northwestern North America and, more locally, across the southern part of the continent. Positive biases are mostly present in the Gulf of Mexico area. For both temperature and precipitation, spatial variability substantially exceeds the magnitude of the domain-mean bias, indicating that regional variability is relatively large. Thus, DPastCliM-NA maintains realistic large-scale spatial patterns even in areas with limited station density, but systematic discrepancies are amplified in mountainous regions, where strong elevation gradients and localized precipitation processes are difficult to resolve. While the PCR framework implicitly incorporates elevation-dependent signals through station-based calibration and spatial predictors, no explicit lapse-rate correction or orographic adjustment is applied beyond what is learned statistically from the training data. Consequently, residual biases over complex terrain likely reflect both station sparsity and inherent limitations of large-scale predictors.