CaSR consists of two complementary components, each generating precipitation fields distributed to users (Sect. ). The online component provides near-real-time regional surface reanalyses by coupling a Numerical Weather Prediction (NWP) model with a land-surface data assimilation system. In CaSR, the NWP system is the Regional Deterministic Reforecast System (RDRS), a reforecast-based adaptation of the operational RDPS . RDRS runs twice daily at 00:00 and 12:00 UTC, producing 48 h integrations. The land-surface assimilation is performed by the Canadian Land Data Assimilation System (CaLDAS; ), which uses a legacy 6-hourly configuration of the Canadian Precipitation Analysis (hereafter CaPA-6h) to improve the precipitation forcing applied to its land-surface model. It is important to note that CaPA-6h precipitation fields are not distributed directly to users, as they serve primarily as internal forcing for CaLDAS. Instead, users access the output of the coupled RDRS–CaLDAS system, which includes hourly precipitation and related surface variables.

The offline component generates an a posteriori daily precipitation reanalysis (hereafter CaPA-24h) using forecasts from the coupled system as background. In CaSR v3.2, the 6–12 and 12–18 h lead times are combined to form the 24 h background field (Fig. , Appendix ), consistent with earlier operational CaPA versions. A limitation of this approach is the reliance on different lead times depending on the hour of the day, which may introduce discontinuities and systematic biases (see Sect.  and Conclusions). For simplicity, the terms offline precipitation reanalysis, CaPA-24h, and CaSR precipitation reanalysis are used interchangeably hereafter.

Major innovations in CaSR v3.2 are the modernization of its underlying numerical weather prediction (NWP) model and the use of ERA5 instead of ERA-Interim atmospheric initial and boundary conditions. The system now employs an updated configuration of GEM (Global Environmental Multiscale; ), which has been operational since June 2024 . This updated configuration includes modernized physical parameterizations, higher vertical resolution, improved topographic representation, and several additional refinements . Together, these changes affect the structure and quality of the precipitation background fields and, consequently, the error characteristics of the precipitation analysis relative to earlier CaSR versions.

The offline precipitation reanalysis based on the CaPA system assimilates precipitation observations from surface stations, primarily sourced from networks operating in Canada and the United States. As shown in the right panel of Fig. , station density is higher in more populated regions, particularly in southern Canada and the eastern United States. Details on contributing networks, gauge types, and spatial coverage are provided in Table  in Appendix .

A key consideration is the change in data sources around the year 2000. Before 2000, observations were obtained from the Integrated Surface Database (ISD; ), whereas after 2000, data were sourced from ECCC operational archives, which do not extend as far back in time. In CaSR v3.2, the ISD dataset was fully updated and reprocessed compared to v2.1, taking advantage of improvements made by NCEI (National Centers for Environmental Information) and allowing for refinement of several processing aspects, including the incorporation of trace precipitation information and more accurate station location definitions. This shift of data source explains the sharp discontinuity in the number of assimilated stations (Figure , left panel), increasing from an average of about 2000 stations before 2000 per analysis to approximately 9000 afterward. The use of ISD data prior to 2000 also introduced additional challenges. The ISD format and metadata structure differ substantially from those of ECCC archives and from the standardized input expected by CaPA, requiring extensive preprocessing to harmonize station identifiers, timestamps, and precipitation codes, and to reconstruct missing metadata such as gauge type. In addition, the spatial coverage and data completeness of ISD records vary across networks and time periods, adding further uncertainty when matching ISD stations with their modern ECCC counterparts.

All precipitation observations undergo automated quality control (QC) procedures consistent with those applied in the operational CaPA system . These procedures, applied at each assimilation time step, include spatial and temporal consistency checks such as leave-one-out validation with neighboring stations. Additional filters account for seasonal effects, including intense summer storms and wintertime windy conditions that can affect measurement reliability. Seasonal variability in the number of assimilated 24 h precipitation observations is also evident in Fig. . The lower number of assimilated gauges during winter reflects the application of more restrictive QC filters under solid precipitation conditions. Although this reduces observational coverage, its impact on the analysis is partly mitigated by the generally higher skill of the background field during winter . Although designed for real-time operations, these QC methods have proven effective in the reanalysis context, but with some limitations. In particular, persistently low precipitation values from certain networks were not flagged by the automated QC. These issues were revealed during an intermediate CaPA-24h run, where monthly accumulations showed that some stations consistently reported unrealistically low totals over specific periods. Such stations were subsequently excluded from the assimilation. The same intermediate run also helped identify suspicious extreme events that escaped initial QC screening. Specifically, stations reporting monthly precipitation exceeding 300 mm while also exceeding five times the corresponding background-field total, were flagged and removed. This additional diagnostic step thus provided a complementary layer of quality assurance beyond the standard real-time QC procedures.

Finally, two Canadian datasets, Adjusted Daily Rainfall and Snowfall (AdjDlyRS) and Adjusted Hourly Rain and Snow (AdjHlyRS, newly included in CaSR v3.2), receive special treatment in the QC workflow. Unlike other data sources, these datasets are exempt from standard CaPA quality control filters because they have been pre-processed to correct known observational biases. AdjDlyRS addresses errors in manual gauge measurements, including undercatch, evaporation, and wetting losses , while AdjHlyRS provides bias-adjusted hourly precipitation from automated stations . Since these corrections already account for conditions that would normally trigger rejection (e.g., solid precipitation or high wind), applying the standard CaPA QC filters would be inappropriate and could discard valid bias-corrected observations. However, their inclusion introduces a potential for duplicate records, a known issue in reanalysis assimilation , since they originate from the same underlying SYNOP and ECCC sources already ingested separately. To prevent redundancy, the reanalysis retains only the adjusted record when an AdjDlyRS or AdjHlyRS station is located within 0.02° of an existing SYNOP or ECCC record, giving preference to the bias-corrected dataset. Although excluded from some rejection filters process, these adjusted observations are still subject to the spatial consistency checks implemented within the CaPA quality-control framework .

The integration of additional datasets, such as AdjHlyRS in CaSR v3.2, led to a substantial increase in the number of assimilated observations after 2000. This increase is clearly reflected in the differences between v3.2 and v2.1 (Fig. , subplot of the differences). While ISD provided valuable information prior to 2000, it is an external dataset over which there is limited control. In contrast, the enhanced data availability in v3.2 represents a tangible improvement in both the density and quality of assimilated observations in the offline precipitation reanalysis.

Three complementary references are used to evaluate the precipitation reanalysis: surface station observations in a leave-one-out framework for point-scale verification, and two gridded products, described below, for product intercomparison.

Precipitation estimates from the evaluated products were verified against daily accumulations from in situ gauge observations, using the station network described in Sect. . The verification process relied on the Frequency Bias Index (FBI) and the Equitable Threat Score (ETS), both derived from a 2×2 contingency table for binary precipitation events, and on the partial mean to characterize conditional precipitation intensity. The FBI measures the ratio between the frequency of predicted and observed events, while the ETS quantifies the fraction of correctly predicted events after accounting for hits expected by random chance. The partial mean provides complementary information by describing the average precipitation intensity conditional on event occurrence below a given threshold, thereby allowing frequency- and intensity-related errors to be disentangled. Definitions of metrics are provided in Appendix . Precipitation events were classified as “1” when daily accumulation exceeded a given threshold and “0” otherwise. These thresholds corresponded either to absolute values (ranging from 0.2 to 100 mm) when the entire domain was considered, or to percentiles (0, 20, 50, 70, 80, 95, and 99.9) computed from the non-zero observed precipitation distribution when regional analyses were performed. A minimum threshold of 0.2 mm was applied to define non-zero 24 h precipitation, both to remove drizzle-level background noise and to match the detection limits of standard precipitation gauges (typically 0.1–0.3 mm).

All metrics were computed for the CaSR offline reanalysis (versions v3.2 and v2.1), its background fields, and ERA5-Land precipitation. The reanalysis estimates used in the verification were derived using the leave-one-out (LOO) framework embedded within the CaPA algorithm (Sect. ). This approach mitigates the artificial skill inflation that would otherwise arise from using the same station observations in both assimilation and verification. Although this approach does not fully eliminate spatial autocorrelation between validation points , it provides an operationally feasible method for independent verification.

Because the in situ network is highly heterogeneous in both space and time, a substantial thinning and quality-selection procedure was applied before verification to ensure spatial representativeness and observational reliability across the domain. Verification relied only on trusted gauges: synoptic stations during the warm season (JJA and SON), and manual synoptic stations during the cold and transition seasons (DJF and MAM), which are better suited for measuring solid precipitation. In addition, adjusted precipitation stations from the Canadian network (AdjDlyRS and AdjHlyRS) were included to enhance the reliability of the verification.

Spatial thinning was performed to reduce over-representation of densely instrumented regions, such as southern Canada and the northeastern United States, by retaining at most one station within each 0.1° × 0.1° grid cell. Within each cell, the station reporting the largest number of valid daily observations over the evaluation period was retained. Temporal thinning was also applied by requiring a minimum data availability of 8 % over the evaluation window. Although this threshold may appear low, it was selected after testing several values and found to be the minimum level at which verification scores stabilized, thereby maximizing the retained sample size while preserving score robustness.

Comparisons between CaSR versions 2.1 and 3.2 were performed only for stations whose coordinates and observed precipitation values matched exactly in both datasets. Despite extensive efforts to harmonize station identifiers and perform spatial matching between the two reanalysis versions, a small fraction of stations could not be consistently paired. Retaining an 8 % temporal completeness threshold thus helps preserve as many stations as possible for meaningful intercomparison while maintaining statistical robustness. The evaluation was conducted separately for each of the four seasons, over the full 1980–2018 period, as well as for two subperiods (1980–1999 and 2000–2018). This partitioning enables assessment of the potential influence of the reduced number of assimilated stations before 2000 on verification results (Sect. ). Maps illustrating the evaluation network after the spatio-temporal thinning are provided in the supplementary material.

Given the large spatial extent of the domain and the diversity of regional climates, the frequency of precipitation events – and consequently the skill scores – can vary substantially across space, complicating interpretation. To account for this variability, all metrics were computed both over the full domain and over predefined subregions. This regional breakdown is particularly important for scores such as the FBI and ETS, which are sensitive to event frequency and thus can be biased by climate regime . The subregions follow the classification of , illustrated in Appendix .

CaPA-24h precipitation reanalysis was compared with ERA5-Land and PRISM from three complementary perspectives: (i) representation of precipitation accumulation, (ii) representation of high-impact precipitation, and (iii) representation of the diurnal cycle of precipitation (the latter restricted to ERA5-Land and CaSR, as PRISM is not available at hourly resolution). Together, these evaluations provide a comprehensive view of how CaSR v3.2 reproduces both the mean and extreme characteristics of precipitation relative to established gridded datasets.

For the first aspect, seasonal precipitation totals were accumulated into 30-year time series (1980–2018) for each dataset and each grid cell, all interpolated to the CaSR v3.2 grid. Maps of mean seasonal accumulations were examined alongside relative differences between CaSR v3.2 and v2.1, CaSR v3.2 and ERA5-Land, CaSR v3.2 and PRISM, and ERA5-Land and PRISM. These comparisons highlight the regional distribution of differences between reanalysis-based datasets (CaSR v3.2, v2.1, ERA5-Land) and the observation-based PRISM dataset. Seasonal time series of accumulated precipitation were also averaged over Bukovsky regions to assess long-term temporal consistency and potential drifts. To better understand the origin of biases in seasonal precipitation totals, the total bias relative to PRISM was further decomposed into contributions from the frequency of wet days (precipitation >1 mm) and the mean precipitation intensity on wet days. This decomposition allows the identification of regions where small total biases may result from compensating errors between precipitation occurrence and intensity.

Beyond mean accumulation, the ability to reproduce extremes is critical for hydrological and climate applications. High-impact precipitation was therefore assessed using indices from the Expert Team on Climate Change Detection and Indices (ETCCDI; ). Four indices were considered: Rx1day (annual maximum 1 d precipitation), Rx5day (annual maximum 5 d precipitation), R95pTOT (annual precipitation from days exceeding the 95th percentile), and R99pTOT (same as R95pTOT but with the 99th percentile). Since results for Rx5day and R99pTOT closely resemble those of Rx1day and R95pTOT, respectively, only Rx1day and R95pTOT are presented here. Time series of indices were compared across seasons and for each grid cell using the Kling–Gupta Efficiency score (KGE; ) described in Appendix . A KGE of 1 indicates perfect agreement with PRISM, while its components highlight whether errors are dominated by bias, variability, or correlation. All indices were computed at the grid-cell level after interpolating each dataset to the CaSR v3.2 grid to ensure spatial consistency. To account for the short length of the time series, which can lead to substantial uncertainty in some KGE components and therefore in the KGE itself , a non-parametric bootstrap approach with 1000 replicates of the annual values, resampled in pairs was applied at each grid cell to estimate 90 % confidence intervals.

Finally, the diurnal cycle of hourly precipitation was examined, as it is a widely used diagnostic of model performance and physical process representation . This evaluation is particularly relevant given that both offline and online precipitation fields for a given day are constructed using forecast segments of varying lead times (see Sect. ). It also provides insights on the hourly fields distributed to users (Sect. ), which are less frequently assessed than daily totals. The diurnal cycle was computed for each grid cell from hourly precipitation in the CaSR v3.2 offline reanalysis and ERA5-Land, averaged over 1980–2018 and stratified by season and region. For each region, the median cycle and interquartile range are presented. Although a direct observational benchmark is not included in this study, this analysis remains valuable for identifying structural differences in the representation of sub-daily precipitation variability.

Figure  presents the evaluation of daily precipitation performance over the 1980–2018 period across the domain, using three complementary metrics – FBI, ETS, and the partial mean – computed at multiple precipitation thresholds and seasons.

Overall, differences between CaPA-24h versions 2.1 and 3.2 are relatively small. This is expected, as both versions assimilate largely similar precipitation observations and are evaluated at station locations using a leave-one-out (LOO) framework, which naturally leads to comparable analysis statistics at station locations. This interpretation is further supported by the comparison with the corresponding background fields, for which more pronounced differences emerge: the CaSR v3.2 background exhibits reduced bias (FBI values closer to unity and partial means closer to observations) and higher skill (larger ETS values) than v2.1. Both CaPA-24h versions consistently show reduced bias and improved skill relative to their respective background fields, highlighting the added value of the assimilation procedure. Finally, because LOO analyses retain information from neighboring assimilated stations, verification scores in densely instrumented regions should be interpreted as an upper bound on the performance expected in truly ungauged areas. Since this residual spatial dependence is nearly identical for both CaPA-24h versions, however, it does not affect their relative comparison.

These results are broadly consistent with those reported by (see their Fig. 9), despite differences in both the evaluation period (2010–2014 versus 1980–2018 here) and the evaluation network. In both studies, the analyses exhibit lower frequency bias (FBI) and higher skill (ETS) than the background fields. A notable difference emerges in winter (DJF) frequency bias at high thresholds, where the longer 1980–2018 period enables improved sampling of rare, high-impact events compared to 2010–2014, leading to a distinct high-threshold signature. This feature is specific to winter conditions, as seasons dominated by liquid precipitation benefit from a much denser station network, whereas winter precipitation is evaluated using a substantially sparser observational network, requiring longer periods to adequately sample intense events. In particular, events exceeding 100 mm d⁻¹ occur much more frequently in the background fields (FBI close to 2) than in the analyses (FBI of approximately 1.0–1.1). However, the impact on the partial mean – defined here as the conditional mean precipitation intensity below a given threshold – remains limited, indicating that the excess detections in the background fields are primarily concentrated near the threshold rather than associated with strongly overestimated intensities. This characteristic is consistent across seasons: background fields generally exhibit larger partial means than observations at high thresholds (especially in JJA), while the analyses tend to slightly underestimate the partial mean for events above approximately 25 mm d⁻¹, with overall biases reduced in v3.2 relative to v2.1.

Including ERA5-Land in the comparison reveals a markedly different response. ERA5-Land tends to overestimate the frequency of light to moderate precipitation events while underestimating the occurrence of high-intensity events, resulting in a threshold-dependent performance of the frequency bias. The partial mean further shows that ERA5-Land systematically produces intensities higher than observed at low to moderate thresholds, with values generally comparable to the CaSR background fields up to about 25 mm d. Beyond this threshold, the partial mean reaches a plateau, indicating an underestimated contribution from higher-intensity events, consistent with the pronounced high-threshold FBI deficit–particularly in summer (JJA). ERA5-Land also exhibits substantially lower skill and larger biases than both CaSR analyses (v2.1 and v3.2) and their respective background fields, as reflected by the ETS metric, in agreement with previous evaluations reported by .

While domain-averaged scores provide a useful first-order assessment of system performance, they can mask substantial regional differences, particularly in regions characterized by sparse station coverage where the background field exerts a stronger influence on the analysis. Differences between CaPA-24h v3.2 and v2.1 are generally small across ETS, FBI, and the partial mean, indicating comparable skill over most regions, as documented in the Supplement. Noticeable discrepancies are mainly confined to the Desert and South regions, where the background field exhibits an excessive precipitation bias in v2.1 , which remains amplified in v3.2. Although the analysis step partially mitigates this degradation through observation assimilation, it does not fully compensate for the background bias, resulting in degraded FBI and partial-mean values, while event-detection skill remains largely unchanged.

More pronounced differences emerge when focusing specifically on the background field (Fig. ), which highlights improvements (degradations) in FBI or ETS scores in red (blue) across regions and seasons. These results indicate an overall reduction in frequency bias in v3.2, while changes in ETS are largely neutral, suggesting that the improvements primarily affect precipitation bias rather than event-detection skill. Winter performance remains more mixed, with a persistent degradation in the southern domain. Overall, these results for the background field are reassuring, as they indicate that in regions with sparse or no station coverage – where the background field dominates in CaSR–CaPA – v3.2 generally provides less biased and more reliable precipitation estimates than v2.1.

Figure  presents the FBI and ETS scores for CaPA-24h v3.2, its background field (BK), CaPA-24h v2.1, and ERA5-Land, evaluated over the CaSR domain for winter (DJF) and summer (JJA). Results are shown separately for two periods: 1980–1999, when observations were primarily sourced from the Integrated Surface Database (ISD), and 2000–2018, when ECCC operational archives were used.

Comparisons between the two evaluation periods reveal only limited changes in relative performance and a stable ranking of the datasets. ETS values are generally slightly higher during the 2000–2018 period than during 1980–1999 for most products in DJF at low to moderate thresholds, whereas no systematic differences are observed in JJA, suggesting a modest overall improvement in categorical precipitation skill. The skill gap between background fields and analyses remains comparable across periods, indicating that the relative impact of the assimilation step is maintained despite major changes in the observing system. As in the previous section, ERA5-Land consistently exhibits the weakest performance across all configurations, with lower ETS values and systematically low FBI values indicating an underestimation of precipitation occurrence, particularly at higher thresholds. This feature, also evident in the partial mean, is consistent across seasons and evaluation periods.

Differences between the two periods mainly emerge at moderately extreme thresholds and are primarily reflected in the FBI. These differences can be attributed to sampling effects arising from (i) differences in station density between the two periods, with fewer available stations prior to 2000 by construction, and (ii) a seasonally varying sampling effect, as winter precipitation is evaluated using a sparser station network than liquid-precipitation seasons. This interpretation is further supported by the ERA5-Land results, which also exhibit small but noticeable differences between the two periods despite being unaffected by the change in assimilated precipitation datasets after 2000, indicating that sampling effects may contribute to the observed variations, independently of changes in the observing system.

Figure  shows the seasonal spatial distribution of precipitation and the effects of updates in CaPA-24h v3.2 compared to v2.1. ERA5-Land is included for comparison due to extended use in the community and its matching horizontal resolution with CaSR, and PRISM serves as a high-resolution observational reference.

Comparisons between v3.2 and v2.1 (second row of Fig. ) show widespread precipitation increases across the southern part of the domain in all seasons, particularly over the Gulf of California and along the Gulf of Mexico, likely associated with changes in the background field and model physics affecting moisture transport and convective processes. These southern increases contribute to a degradation of the frequency bias in v3.2 relative to v2.1, especially for small to moderate intensity precipitation events, through an increased false alarm rate in arid regions (e.g., south of the Desert region following the Bukovsky classification). In the Rocky Mountains, precipitation increases especially at high elevations, primarily due to modifications in the CaPA background field linked to GEM model upgrades and a sharper representation of topography. In summer, precipitation over much of Canada is reduced in v3.2, while regions with dense station coverage (e.g., along the US–Canada border and in the eastern United States) remain largely unchanged, reflecting stronger observational constraints.The reduced summer precipitation accumulations in northern regions improve skill and reduce bias relative to v2.1, indicating that v2.1 was overall too wet in these areas.

The relative differences between CaPA-24h v3.2 and ERA5-Land exhibit their strongest amplitudes over complex terrain, with consistent structures along the western cordillera in all seasons (Fig. ). In DJF and SON, a marked sign alternation is apparent across the coastal mountains and the Rocky Mountains: CaPA-24h tends to be drier than ERA5-Land along parts of the Pacific coastal ranges, while becoming wetter over portions of the interior high terrain, consistent with differing representations of orographic precipitation and the smoother fields in ERA5-Land. In MAM, negative differences dominate much of southern Canada and large parts of the continental United States, indicating generally wetter conditions in ERA5-Land during spring, whereas positive differences persist over high latitudes. In JJA, negative differences extend over most of Canada, suggesting systematically larger summer accumulations in ERA5-Land, while CaPA-24h remains wetter over Mexico and parts of Central America. Persistent positive anomalies over Mexico throughout the year point to limitations near the southern edge of the domain.

The bottom two rows of Fig.  compare CaPA-24h v3.2 and ERA5-Land against PRISM in terms of relative seasonal precipitation differences. CaPA-24h v3.2 generally shows closer agreement with PRISM than ERA5-Land, reflecting both the assimilation of similar station networks in CaPA-24h and PRISM and CaPA-24h enhanced ability to capture finer-scale spatial variability relative to the smoother ERA5-Land fields. However, both products exhibit pronounced discrepancies over the western United States, particularly along the Pacific coast and across major mountain ranges, reflecting the strong sensitivity of relative errors to complex topography. In these regions, CaPA-24h v3.2 generally shows positive differences relative to PRISM over windward slopes, while localized negative differences appear, especially in winter and spring. Large relative differences are also apparent over parts of the central United States, notably during winter and spring. These signals often coincide with regions of low seasonal precipitation totals (typically below 50 mm; first row in Fig. ), such that modest absolute differences can translate into large relative values. Consequently, relative differences in these areas should be interpreted with caution; complementary maps of absolute differences are provided in the Supplement (Fig. S2). Overall, CaPA-24h v3.2 exhibits finer spatial variability and more localized structures than ERA5-Land when compared to PRISM, particularly over complex terrain. While neither product consistently outperforms the other everywhere, CaPA-24h v3.2 tends to reproduce sharper gradients and regional contrasts that are more consistent with the high-resolution PRISM fields, whereas ERA5-Land shows smoother and more spatially homogeneous patterns across seasons. In summer (JJA), both CaPA-24h v3.2 and ERA5-Land show a systematic underestimation of seasonal precipitation accumulations relative to PRISM over large portions of the CONUS. This signal persists despite the higher density of precipitation observations assimilated in CaPA-24h during the warm season, indicating that it is unlikely to be primarily driven by observational coverage. Instead, it likely reflects limitations common to both reanalyses, including the representation of warm-season convective precipitation, scale mismatches between grid-cell averages and the high-resolution PRISM estimates, and smoothing inherent to the analysis systems. The underestimation is most apparent in regions where convective precipitation can dominate summer totals.

To further investigate the origin of these seasonal accumulation biases, the total precipitation bias relative to PRISM is decomposed into contributions from the frequency of wet days (precipitation >1 mm) and the mean intensity on wet days. This decomposition also reveals regions where apparently small total biases result from compensating errors between precipitation frequency and intensity.

Figure  presents the regional decomposition of the total seasonal precipitation bias, computed with respect to PRISM, into contributions from wet-day frequency and precipitation intensity on wet days for CaPA-24h v3.2 and ERA5-Land across selected Bukovsky regions. Across most regions and seasons, differences in total accumulation are primarily driven by biases in wet-day frequency rather than by systematic errors in intensity. In winter, ERA5-Land exhibits a pronounced positive frequency bias over the Mountain West, the Great Plains, and Desert regions, which largely explains its positive total bias in these areas, while intensity biases remain comparatively modest. In contrast, CaPA-24h v3.2 shows smaller and more balanced frequency biases across regions, with intensity biases generally close to zero, indicating a better balance between occurrence and intensity errors.

In summer, it is noteworthy that in the South and Great Lakes regions the bias decomposition highlights a marked compensation between frequency and intensity errors in ERA5-Land. In these regions, ERA5-Land tends to overestimate the frequency of wet days while simultaneously underestimating precipitation intensity, resulting in near-zero total seasonal biases when only accumulated precipitation is considered. This compensation can give the impression of overall agreement with PRISM when examining total accumulations alone, despite substantial and physically meaningful discrepancies in the underlying precipitation characteristics. By contrast, CaPA-24h v3.2 exhibits smaller and more coherent biases in both frequency and intensity over the South and Great Lakes, leading to total biases that more directly reflect the underlying error structure. For the other regions, similarly to winter, CaPA-24h v3.2 shows a more balanced frequency–intensity bias compared to ERA5-Land, particularly over the Pacific region.

KGE maps and their components for the annual daily maximum precipitation (Rx1day) are presented in Fig.  for each season. The analysis focuses on the CONUS domain, where PRISM is available as reference, and on the 2002–2018 period. Given the limited number of annual samples, the corresponding bootstrap confidence intervals (Sect. S3 in the Supplement) are used throughout this section to identify robust signals.

A clear east–west contrast emerges in the KGE fields. East of the major mountainous regions, CaPA-24h v3.2 generally outperforms ERA5-Land in all seasons, with KGE values mostly ranging between 0.7 and 0.9. These high scores are mainly driven by bias and correlation components close to one, while the variability term is noisier, showing several blue patches indicative of higher variability in CaPA-24h relative to PRISM. Part of this noisiness reflects sampling uncertainty rather than actual disagreement, the variability ratio is the least constrained component of the KGE, with 90 % confidence intervals typically exceeding 0.5 in width (Sect. S3).In contrast, ERA5-Land systematically yields lower KGE values in this region, driven by larger biases (consistent with the accumulation results), weaker correlations, and reduced variability compared to PRISM. During summer, both reanalyses perform less well, with noisier KGE patterns, consistent with the challenges of capturing sub-grid convective storms . Nonetheless, CaPA-24h v3.2 maintains smaller biases and higher correlations than ERA5-Land, which even exhibits locally negative correlations. Variability remains particularly noisy in both datasets but tends to be systematically underestimated in ERA5-Land.

In the western mountainous regions, performance degrades for both reanalyses, particularly with respect to bias. Across all seasons, CaPA-24h v3.2 exhibits patches of very low KGE values (below -0.4), especially in winter and spring. Following the interpretation of , such values indicate that the reanalysis does not add skill compared to the climatological mean. In these regions, both bias and variability contribute to the degradation of KGE.

ERA5-Land fields appear markedly smoother than those of CaPA-24h v3.2, consistent with the coarser-scale nature of ERA5, from which ERA5-Land directly inherits its characteristics without additional downscaling or bias correction .

The same analysis applied to R95pTOT (Fig. ) complements Rx1day and other extreme precipitation metrics by assessing whether annual precipitation totals are dominated by very wet days or by light-to-moderate precipitation. KGE values for R95pTOT are generally lower than for Rx1day in both reanalyses. CaPA-24h v3.2 again outperforms ERA5-Land, particularly in MAM, JJA, and SON, while DJF remains the most challenging season to represent accurately.

Regional differences are evident. Both reanalyses perform poorly in the western mountainous regions, and ERA5-Land shows additional deficiencies over the eastern Appalachians, with KGE values below -0.4. In contrast, the southern part of the domain is relatively well represented, with KGE values around 0.7 or higher. Summer and autumn are the best-represented seasons for CaPA-24h v3.2, coinciding with more active precipitation regimes (summer convection and autumn cyclones), where the assimilation of surface data likely provides added value, especially in the eastern domain, where the surface observation network is denser.

Low KGE values are primarily driven by large biases, with both CaPA-24h v3.2 and ERA5-Land overestimating mean R95pTOT by at least a factor of two in several regions and seasons; the sign of these biases is robust to sampling variability, as the bootstrap confidence interval for β excludes 1 in 98 %–99 % of grid cells (Sect. S3). This result is somewhat counterintuitive, as one might expect mean annual R95pTOT values to be lower in the reanalyses given the finer native grid resolution of PRISM. However, because R95pTOT is defined relative to the 95th percentile, the higher threshold in PRISM leads to fewer exceedances and thus lower values. Conversely, the lower 95th-percentile threshold in CaPA-24h v3.2, and especially in the smoother ERA5-Land, produces more frequent exceedances. It should also be noted that percentiles are computed over the relatively short 2002–2018 period, which may limit their robustness. On the positive side, correlations for each season are close to one, indicating that the temporal variability of R95pTOT is well captured, particularly by CaPA-24h v3.2 and, to a lesser extent, ERA5-Land. This suggests that, with appropriate bias correction, the temporal evolution of R95pTOT could be accurately represented.

KGE and its components were also computed for Rx5day and R99pTOT (notshown; see Supplement). Regional and seasonal patterns for Rx5day are consistent with those obtained for Rx1day, with slightly reduced sampling uncertainty due to the temporal smoothing of the 5 d accumulation. For R99pTOT, spatial patterns resemble those of R95pTOT, but sampling uncertainty becomes very large (median β CI widths exceeding 3; Sect. S4), so that grid-cell scores for this index is difficult to interpret.

Analyzing the diurnal cycle of hourly precipitation provides insights beyond mean totals, as it reveals whether models capture the correct timing, amplitude, and regional structure of precipitation events. This diagnostic is particularly relevant in summer, when convection dominates , as many models tend to trigger rainfall too early in the day or misrepresent its intensity. In the following, the diurnal cycle is evaluated by comparing the hourly reanalysis from CaSR v3.2 (i.e. the disaggregated CaPA-24h, Sect. ) with ERA5-Land.

Figure  shows the median and interquartile range of the diurnal cycle of precipitation intensity for selected regions, displayed from 12:00 to 12:00 UTC over the 1980–2018 period, for winter (Fig. a) and summer (Fig. b). To ease the interpretation of UTC times, mean local solar noon is indicated in each panel, providing a fixed geographic reference that is independent of time-zone boundaries. In winter, as expected, the diurnal cycle exhibits weak variability across most regions, except in the Great Lakes region, where precipitation activity in both CaSR v3.2 and ERA5-Land exhibits a minimum near local solar noon and remains elevated from the evening through the early morning, and in the East region (not shown), where activity mainly occurs overnight and in the early morning, with a relative minimum around midday. Precipitation hourly rates are of similar magnitude in both datasets. These findings are consistent with , who reported weak wintertime diurnal variations over the contiguous United States. Small peaks appear in CaSR v3.2 at specific synoptic hours (06:00, 12:00, and 18:00 UTC), corresponding to the lead-time stitching inherent to the disaggregation process. These artefacts are more pronounced in wetter regions, such as the South and, to a lesser extent, the Desert.

During summer (Fig. b), the diurnal cycle is more pronounced, and the non-physical discontinuities in CaSR v3.2 are more visible. Ignoring these artificial peaks, both datasets reproduce the expected pattern: summer precipitation generally peaks in the late afternoon, around 3 to 6 h after local solar noon (i.e. 21:00–00:00 UTC depending on the region), over the EBoreal, MtWest, Desert, Great Lakes, South, and East (not shown) regions. Some areas, such as the Prairies, display a secondary maximum around midnight (06:00 UTC) in addition to a main late-afternoon peak (23:00 UTC). These patterns are consistent with the diurnal cycles documented by , highlighting the typical late-afternoon maximum of convective precipitation.

The artificial discontinuities in CaSR v3.2 at synoptic hours reach amplitudes of about 0.02 mm h⁻¹ in summer and are associated with decreases in precipitation intensity immediately following each peak. These features partly reflect the lead-time stitching inherent to the temporal disaggregation procedure, whereby precipitation amounts tend to be enhanced near forecast boundaries. In addition, part of the bias in the diurnal cycle may be related to model spin-up effects, as the atmospheric model typically requires 12 to 24 h to establish a dynamically consistent diurnal cycle of precipitation over the domain. During this adjustment period, precipitation characteristics have not yet fully stabilized and may differ from their representative diurnal cycle. The presence of these artefacts, and potential approaches to mitigate them, are discussed in the Conclusion as part of future improvements to the reanalysis. Overall, the temporal disaggregation procedure (Sect. ) does not appear to substantially alter the phase of the diurnal cycle itself, although event-based analyses could provide further insight.