KM5c (also referred to as KM5.3) is an interglacial centred on a ∼100 kyr window of relatively depleted benthic 18O values, which immediately follows a pronounced δ18O peak during the glacial stage M2 (3.3 Ma; Fig. 1). Minor changes to orbital forcing during KM5 enables a wider target zone (3.205Ma±20kyr) for data collection because the potential for orbitally forced regional and time-transgressive climate signals is minimised (Haywood et al., 2013). A comparable approach has been adopted by the PRISM4 synthesis (3.190 to 3.220 Ma; Foley and Dowsett, 2019; see Fig. 1). Here, we focus on a narrow time slice from 3.195 to 3.215 Ma, to span approximately one precession cycle. The reconstructed atmospheric CO2 concentrations from boron isotopes in KM5c are 360±55 ppmv (for median boron-derived values (n=3), full range: 289–502 ppmv, Fig. 1; Foster et al., 2017). A wider range of atmospheric CO2 concentrations has been reconstructed for the whole mid-Piacenzian stage (356±65 ppmv for median values (n=36), full range: 185–592 ppmv; Foster et al., 2017).

The PAGES-PlioVAR working group agreed on a set of stratigraphic protocols to maximise confidence in the identification and analysis of orbital-scale variability within the mid-Piacenzian stage (McClymont et al., 2017). Sites were only included in the synthesis if they had either (i) ≤10 kyr resolution benthic δ18O data which could be (or had been) tied to the LR04 stack (Lisiecki and Raymo, 2005) or the HMM-Stack (Ahn et al., 2017) and/or (ii) the palaeomagnetic tie points for upper Mammoth (C2An.2n (b) at 3.22 Ma) and lower Mammoth (C2An.3n (t) at 3.33 Ma). At one site (ODP Site 1090) these conditions were not met (see Supplement), but tuning to LR04 had been made using a record of dust concentrations under the assumption that higher dust flux occurred during glacials as observed during the Pleistocene (Martinez-Garcia et al., 2011). At ODP Site 806, uncertainty over age control resulted from the absence of an agreed splice across the multiple holes drilled by ODP, and a new age model has been constructed (see Supplement). For some sites (see online summary at https://pliovar.github.io/km5c.html, last access: 30 June 2020), revisions to the published age model were made, for example if the original data had been published prior to the LR04 stack (Lisiecki and Raymo, 2005) or prior to revisions to the palaeomagnetic timescale (Gradstein et al., 2012). In total, data from 32 sites were compiled, extending from 46∘ S to 69∘ N (Fig. 2).

A multi-proxy approach was taken, to maximise the information available on changing climates and environments during the KM5c interval. Two SST proxies were analysed: the alkenone-derived U37K′ index (Prahl and Wakeham, 1987) and foraminifera calcite Mg/Ca (Delaney et al., 1985). Both proxies have several calibrations to modern SST: here we explore the impact of calibration choice on KM5c SST data, by comparing and contrasting outputs between proxies and between calibrations. Although the TEX86 proxy (Schouten et al., 2002) has also been used to generate mid-Piacenzian SSTs (e.g. O'Brien et al., 2014; Petrick et al., 2015; Rommerskirchen et al., 2011), these data are not included here because they could not be confidently assigned to the KM5c interval either due to low sampling resolution and/or because our age control protocol was not met.

The model outputs used here were generated from the 15 models that contribute to the Pliocene modelling intercomparison project, Phase 2 (PlioMIP2) (Haywood et al., 2020). The boundary conditions for the experiments and their large-scale results for Pliocene and pre-industrial climates are detailed elsewhere (Haywood et al., 2020, 2016b), so they are briefly outlined here.

The Pliocene simulations are intended to represent KM5c (∼3.205 Ma) and were forced with PRISM4 boundary conditions (Haywood et al., 2016b). Atmospheric CO2 concentration was set to 400 ppmv (Haywood et al., 2020), in line with the upper estimates of atmospheric CO2 from boron isotope data (Fig. 1; Foster et al., 2017). Lower estimates from the alkenone carbon isotope proxy (Fig. 1) are likely to reflect an insensitivity of this proxy to atmospheric CO2 in the Pliocene (Badger et al., 2019). All other trace gases, orbital parameters, and the solar constant were specified to be consistent with each model's pre-industrial experiment. The Greenland Ice Sheet was confined to high elevations in the eastern Greenland mountains, covering an area approximately 25 % of the present-day ice sheet. The Antarctic ice sheet has no ice over West Antarctica. The reconstructed PRISM4 ice sheets have a total volume of 20.1×106 km3, equating to a sea-level increase relative to the present day of less than ∼24 m (Dowsett et al., 2016).

Modelling groups had some choices regarding the exact implementation of boundary conditions; however, 14 of the 15 models used the “enhanced” PRISM4 boundary conditions (Dowsett et al., 2016) which included all reconstructed changes to the land–sea mask and ocean bathymetry. Key ocean gateway changes relative to modern values are the closure of the Bering Strait and Canadian archipelago, and the exposure of the Sunda and Sahul shelves (Dowsett et al., 2016). The initialisation of the experiments varied between models (Haywood et al., 2020). Some models were initialised from a pre-industrial state while others were initialised from the end of a previous Pliocene simulation or another warm state. The simulations reached equilibrium towards the end of the runs as per PlioMIP2 protocol.

For all anomaly calculations we obtain pre-industrial SST from the NOAA-ERSST5 dataset for the years 1870–1899 CE (Huang et al., 2017), ensuring alignment between the KM5c proxy data and the KM5c model experiments (Haywood et al., 2020). This pre-industrial time window excludes the largest cooling linked to the Little Ice Age and predates the onset of 20th-century warming (Owens et al., 2017; PAGES2k Consortium, 2017). The global mean SST anomaly from the proxy data was obtained as follows: firstly, the SST anomaly between the proxy data and the NOAA-ERSST5 data was obtained for each location and the data collated into bins of 15∘ of latitude. It is assumed that the average of all the data in each bin represents the average SST anomaly for that latitude band. Next, the area of the ocean surface for each bin is obtained. The average SST anomaly is then the average of all the bins weighted by the ocean area in the relevant latitude band.

Meridional gradients were obtained in a similar way. A low-latitude SST anomaly was obtained as the weighted average of all the bins containing low-latitude SSTs (for example the 4∘×15∘ bins containing latitudes of 30∘ S–30∘ N). A high-latitude SST anomaly was obtained as the weighted average of all bins containing high-latitude SSTs (>60∘ N because there were no proxy data points >60∘ S). As only the Atlantic Ocean contained data points poleward of 65∘ N, the high-latitude region used in the gradient calculations for both proxies and models was focused on the longitudinal window from 70∘ W to 5∘ E. The meridional gradient SST anomaly is then the low-latitude SST anomaly minus the high-latitude SST anomaly, relative to the pre-industrial period.

There are some uncertainties in this calculation of the global mean SST, in particular, the fact that the proxy data are not evenly distributed throughout a latitude bin and also that some bins contain very few data points. There is a higher density of data in the Atlantic Ocean compared to the Indian Ocean and Pacific Ocean, and no high-latitude data are available to consider a Southern Ocean response (Fig. 2). Nevertheless, this method of calculating averages does attempt to account for unevenly distributed data and provides an SST anomaly (SSTA) that is comparable with model results. The impact of proxy choice was examined in the calculation of the global means and meridional SST gradients. As no Mg/Ca data were available >50∘ N or >30∘ S, we calculated global mean SST and the meridional SST gradients either including or excluding the Mg/Ca data; both results are outlined below and shown in Table 1.

KM5c is characterised by a surface ocean which is ∼2.3 ∘C (alkenones and Mg/Ca), ∼3.2 ∘C (alkenones-only, Müller98 calibration), or ∼3.4 ∘C (alkenones-only, BAYSPLINE calibration) warmer than pre-industrial values, with a ∼2.6 ∘C reduction in the meridional SST gradient. The global mean SST anomaly is higher than the 1.7 ∘C previously calculated for the wider mid-Piacenzian warm period (3.1–3.3 Ma), regardless of proxy choice (IPCC, 2014b). Previous analysis of a suite of models suggested that a climate state resembling the mid-Piacenzian was likely to develop and be sustained under RCP4.5 (Burke et al., 2018). The PlioMIP2 ensemble (Haywood et al., 2020) indicates that best estimates for mid-Piacenzian warming in surface air temperatures (1.7–5.2∘) are comparable to projections for the RCP4.5 to 8.5 scenarios by 2100 CE (RCP4.5=1.8±0.5∘, RCP8.5=3.7±0.7 ∘C; IPCC, 2013). Our proxy-based mean global SST anomaly is larger than in most PlioMIP2 models when we use only alkenones or alkenones and Mg/Ca combined (Fig. 3), and hence our results suggest that the global annual surface air temperature anomaly for KM5c would exceed the PlioMIP2 multi-model surface air temperature mean of 2.8 ∘C (Haywood et al., 2020). The higher global SST mean recorded in the KM5c proxy data, compared to the PlioMIP2 models, occurs despite the available atmospheric CO2 reconstructions indicating values below the ∼400 ppmv used in the PlioMIP2 models (Fig. 1). Our synthesis of SST data thus indicates that with atmospheric CO2 concentrations ≤400 ppmv (comparable to RCP4.5), the surface ocean warming response will likely be larger than indicated in models. Further work is required to increase the temporal resolution of the atmospheric CO2 reconstructions through KM5c, to improve our understanding of the reconstructed SST response to CO2 forcing, including whether (or by how much) the reconstructed atmospheric CO2 differs from model boundary conditions and whether other changes in the model boundary conditions also influence SST patterns (e.g. gateway changes outlined in Sect. 2.4).

Proxy choice, calibration choice, and site selection have all had an impact on the magnitude of the change in meridional SST gradient for KM5c compared to the pre-industrial period (Table 1). Focussing only on a Northern Hemisphere SST gradient leads to higher gradient anomalies than when all of the low latitudes are included (30∘ S–30∘ N) because it excludes the high SST anomalies of the Benguela upwelling sites (20–25∘ S; discussed below, Fig. 4). Smaller meridional SST gradient anomalies occur using BAYSPLINE (+0.03 to -1.66 ∘C) than the Müller98 calibration for U37K′ (-1.18 to -3.00 ∘C; Table 1), due to the increased low-latitude SST anomalies generated by BAYSPLINE (Fig. 4). Due to several (but not all) low-latitude sites recording negative SST anomalies for KM5c using foraminifera Mg/Ca, the inclusion of Mg/Ca data leads to a larger difference in the meridional SST gradient relative to the pre-industrial period (-2.19 to -4.08 ∘C). Further work is required to fully understand the negative KM5c SST anomalies in some of the low-latitude sites (discussed further below), given their impact on the meridional SST gradients. However, a robust pattern emerging from the data is that the KM5c proxy data detail smaller low-latitude SST anomalies than those of the mid- and high-latitude SST anomalies (Fig. 4), leading to a reduction in the meridional SST gradient relative to the pre-industrial period. Enhanced mid- and high-latitude warming has been observed in other warm intervals of the geological past, including the last interglacial and the Eocene (Evans et al., 2018; Fischer et al., 2018), and is a feature of future climate under elevated CO2 concentrations (IPCC, 2014a).

There is complexity in the amplitude of the KM5c SST anomaly by latitude and basin, which may reflect patterns of surface ocean circulation. In the Northern Hemisphere, relatively muted warming in the East Greenland Current (ODP Site 907, 69∘ N) may reflect the presence of at least seasonal sea-ice cover from ca. 4.5 Ma (Clotten et al., 2018). In contrast, relatively high SST anomalies at ODP Sites 642 (67∘ N) and 982 (58∘ N) track the northward flow of the North Atlantic Current, accounting for the enhanced warming relative to north-east Pacific IODP Site U1417 (57∘ N; Figs. 4 and 5). The large North Atlantic SST anomalies also contribute to an enhanced Northern Hemisphere meridional SST gradient of up to 4 ∘C (>60∘ N minus 0–30∘ N; Table 1). For the Southern Hemisphere, a signal of polar amplification is less clearly identified than for the Northern Hemisphere (Fig. 4), although we recognise that all sites are <46∘ S. Low KM5c SST anomalies (<2 ∘C) at DSDP sites 593 and 594 (41 and 46∘ S, respectively) might be accounted for by a similar positioning of the Subtropical Front close to New Zealand during KM5c as today (McClymont et al., 2016; Caballero-Gill et al., 2019). Antarctic Intermediate Water (AAIW) temperatures were also only ∼2.5 ∘C warmer than pre-industrial values during KM5c, suggesting a small warming in subantarctic waters where AAIW forms (McClymont et al., 2016). In contrast, large anomalies at ODP Sites 1125 (43∘ S, Pacific) and 1090 (43∘ S, Atlantic) reflect a greater sensitivity to expanded subtropical gyres during KM5c, contrasting with the Pleistocene equatorward displacement (and enhanced cooling) of subpolar water masses (e.g. Martinez-Garcia et al., 2010) which today places both of these sites poleward of the Subtropical Front.

Given that our proxy data meridional SST gradient calculations use only two sites to calculate the high-latitude SSTs (ODP Sites 907 and 642), which are also both from the Nordic Seas (Fig. 2), we explored the impact of expanding our high-latitude band into the mid-latitudes. We also explored narrowing the low-latitude band so that it does not include the Benguela upwelling sites, which have a significant data–model offset (Fig. 4) and may be influenced by localised circulation changes (see Sect. 4.2). Previous calculations of Pliocene meridional SST gradients have also considered differences between the mid-latitudes and low latitudes through time (Fedorov et al., 2015). Despite adding four more sites by expanding the high-latitude band to 45∘ N/S, the meridional SST gradients are reduced by <0.4 ∘C, from -1.18 to -1.56 ∘C using the original U37K′ data (Table 1). However, it is clear from the distribution of sites (Fig. 2) that our reconstructed KM5c SSTs (and thus the global mean and meridional gradients) have a strong signal from the Atlantic Ocean. There is a relative scarcity of sites from the Indian Ocean, Pacific Ocean, and Southern Ocean, but it is difficult to ascertain what impact this may have had on our global analysis. Further work is required to increase the spatial density of SST data for KM5c and the wider mid-Piacenzian stage, to better evaluate the magnitude of the warming and gradient changes outlined here.

For the mid-latitudes and high latitudes, we find broad proxy data–model agreements for most sites. In the North Atlantic Ocean, reconstructed SST KM5c anomalies from U37K′ fall within the ranges provided by the PlioMIP2 models (Fig. 4) for all but one site (IODP Site U1387, 37∘ N). The overall U37K′–model agreement for the North Atlantic Ocean suggests that, as proposed by Haywood et al. (2013), a focus on a specific interglacial within the mid-Piacenzian provides an improved comparison to the climate being simulated by the PlioMIP2 models. Thus, some of the data–model mismatch in previous mid-Piacenzian syntheses (e.g. Dowsett et al., 2012) may have been due to the averaging of warm peaks which may not have been synchronous in time between sites and/or with the interval being modelled. Disagreements occur between proxies (North Atlantic Ocean) and between proxies and models (Benguela upwelling, Gulf of Cadiz, Mediterranean Sea) (Fig. 4). Here, we explore the potential causes for these offsets in turn.

The largely U37K′-derived data from the North Atlantic Ocean tend to align with the warmest model outputs (Fig. 4), and the U37K′-SST anomalies also tend to be larger than those from Mg/Ca. A challenge for understanding the cause(s) of the U37K′-Mg/Ca differences is that only two sites have data from both proxies, and these do not show a consistent signal. There is good correspondence between U37K′ and the original published Mg/Ca SSTs for IODP Site U1313 (41∘ N), whereas at ODP Site 609 Mg/Ca SSTs (both calibrations) are between 4.6 and 6.1 ∘C cooler than U37K′. It has also been shown that the U37K′-Mg/Ca SST offset at Site 609 is not constant with time for the late Pliocene (Lawrence and Woodard, 2017). When using BAYMAG, warmer KM5c SSTs are reconstructed than the original published data at DSDP Site 603 and IODP Site U1313 (35 and 41∘ N; Fig. 4), but BAYMAG reconstructs SSTs only 0.8 ∘C warmer than the original published SSTs at Site 609. These mid-latitude North Atlantic Mg/Ca data are provided by G. bulloides, which may calcify at depth in the water column (e.g. Mortyn and Charles, 2003; Schiebel et al., 1997) and account for those sites where Mg/Ca reconstructions give lower reconstructed SSTs than from U37K′ (Bolton et al., 2018; De Schepper et al., 2013). Alternatively, an offset between alkenones and Mg/Ca might be accounted for if there is a seasonal bias to the U37K′ calibration (e.g. Conte et al., 2006; Schneider et al., 2010). Despite documented seasonality in alkenone production at high latitudes, it has been proposed that mean annual SSTs continue to be recorded by U37K′ in sediments (Rosell-Melé and Prahl, 2013), as indicated by the original U37K′ calibration (Müller et al., 1998). In contrast, BAYSPLINE explicitly assumes an autumn signal is recorded at Atlantic sites >45∘ N (Tierney and Tingley, 2018). Despite these differences in interpretation, BAYSPLINE values for KM5c are <0.7 ∘C cooler than the original published U37K′ data (Fig. S2). Although the North Atlantic U37K′ data align with a range of mean annual SST anomalies generated by the PlioMIP2 models (Fig. 4), three of the sites show alignment between U37K′ SSTs and the July–November values from the multi-model means (Fig. 5). In contrast, Site 907 aligns with cool spring temperatures in the models, perhaps reflecting production after sea ice melt.

The large data–model discrepancy at 30∘ S reflects three sites which today sit beneath the Benguela upwelling system in the south-east Atlantic (20–26∘ S; Fig. 4). Part of the data–model discrepancy in the KM5c anomaly can be attributed to the models overestimating pre-industrial SSTs at the northern Benguela sites (NOAA-ERSST5 SSTs are 2–5 ∘C below the pre-industrial model range) and suggests that models are not fully capturing the local dynamics of the coastal upwelling today (Small et al., 2015). Realistic representations of the Benguela upwelling system today are proposed to require realistic wind stress curl and high-resolution atmosphere and ocean models (<1∘; Small et al., 2015). Most of the PlioMIP2 simulations use lower-resolution atmosphere and ocean models (Haywood et al., 2020). An increased density of proxy data reconstructing KM5c atmospheric circulation, as well as the application of high-resolution models, may help to understand the observed KM5c data–model discrepancy. Furthermore, there was a deep thermocline during the Pliocene (as reconstructed in the equatorial Pacific (Ford and Ravelo, 2019; Ford et al., 2015; Steph et al., 2006; Steph et al., 2010) and theorised globally (Philander and Fedorov, 2003)), so that warmer subsurface waters than today were upwelled, enhancing local warming. However, warming of ∼3.4 ∘C in subsurface waters (Ford and Ravelo, 2019) and ∼2.5 ∘C in intermediate waters (McClymont et al., 2016) for Pliocene interglacials suggests that Pliocene upwelling of warmer waters is unable to fully account for the 7–10 ∘C SST anomalies at Benguela sites for KM5c. Changes to the distribution of export productivity and SSTs indicate that an overall poleward displacement of the Benguela upwelling system occurred during the Pliocene, so that the main zone of upwelling likely sat close to ODP Site 1087 at 31∘ S (Etourneau et al., 2009; Petrick et al., 2018; Rosell-Melé et al., 2014). As the northern and southern Benguela regions are today marked by differences in the seasonality of the upwelling, a temporal shift in upwelling intensity may also account for some of the large SST anomaly (Haywood et al., 2020). Thus, the data–model disagreement may be accounted for by a combination of displaced upwelling and warmer upwelled waters, giving large SST anomalies in Benguela proxy data, alongside the challenges of modelling both the pre-industrial and KM5c upwelling system and its associated SSTs.

Data–model disagreement also occurs at two Northern Hemisphere sites where U37K′-SST anomalies exceed those given by the range of model predictions (Fig. 4). Punto Piccola (Sicily, 37∘ N) is located within the Mediterranean Sea, whereas IODP Site U1387 (37∘ N, Iberian margin) records the influence of the waters sourced from the Azores Current and the subtropical gyre. The data–model disagreement for KM5c reflects warmer SST estimates from the proxy data compared to the models, despite the good agreement for the pre-industrial period suggesting that locally complex ocean circulation in these near-shore and marginal marine settings may have been captured in the models. For Punto Piccola, the data–model offset is also likely to be a minimum because BAYSPLINE Mediterranean SSTs explicitly record November–May temperatures (Tierney and Tingley, 2018), and alkenone production below the sea surface has also been proposed (Ternois et al., 1997): both scenarios would act to raise mean annual SSTs further from those simulated in the PlioMIP2 models (Fig. 4). Further multi-proxy investigation is required to identify whether the data–model disagreements in the Benguela upwelling, Gulf of Cadiz, and Mediterranean Sea reflect challenges in modelling near-shore or complex oceanographic systems and/or biases in the temperature signal recorded by the proxy data.

The low-latitude U37K′-SST anomalies for KM5c align well with the PlioMIP2 models (Fig. 4). At ODP Sites 806 and 959, the Mg/Ca anomalies using the original calibrations are both +0.3 ∘C compared to the pre-industrial period (Fig. 4) and also align with the PlioMIP2 models. At Site 806 the BAYMAG KM5c anomaly (+1.7 ∘C) also aligns with the PlioMIP2 models. Only one low-latitude site has both U37K′ and Mg/Ca SST data: ODP Site 1143 (9∘ N) records KM5c anomalies of +0.8 to +2.5 ∘C (U37K′) or -0.7 to -1.3 ∘C (Mg/Ca). Although the U37K′ data align with the model outputs for Site 1143, the negative anomaly in Mg/Ca lies outside the model range for mean annual SST (Fig. 4).

Six of the low-latitude sites have negative low-latitude SST anomalies in KM5c from foraminifera Mg/Ca; these occur regardless of whether the original or BAYMAG calibrations are applied and for both G. ruber and T. sacculifer-based reconstructions. The negative KM5c Mg/Ca-SST anomalies lie beyond those shown across the PlioMIP2 model range (Fig. 4), despite the absolute Mg/Ca SSTs reconstructed from these sites for KM5c falling within the model range for all but two of the sites (ODP Sites 999 (13∘ N) and 1241 (6∘ N); Fig. S5). However, the absolute SST values reconstructed for KM5c from Mg/Ca tend to align with the colder model outputs (Fig. S5).

Mg/Ca-SST calibration choice has no consistent impact on the KM5c anomalies (across all latitudes; Fig. 4). Therefore, the corrections for secular seawater Mg/Ca change and/or non-thermal influences over Mg/Ca, which are accounted for in BAYMAG (Tierney et al., 2019b), do not account for these cold tropical KM5c anomalies. For example, for ODP Site 806 in the western Pacific warm pool, BAYMAG SST estimates for KM5c are ∼1 ∘C warmer than the published Mg/Ca record (Wara et al., 2005). For Site 999 in the Caribbean Sea, BAYMAG SST estimates for KM5c are ∼0.5 ∘C cooler than the published Mg/Ca record (De Schepper et al., 2013). This also suggests that the impact of Mg/Caseawater change on SST is small on warm pool sites. The Mg/Caseawater correction used in BAYMAG is conservative, drawing on multiple lines of physical evidence (corals, fluid inclusions, calcite veins, etc.) (Tierney et al., 2019b). Given the variable directions of the offsets between published and BAYMAG SSTs shown here, the Mg/Caseawater correction is unable to account for the data–model offsets observed for the low latitudes.

CaCO3 dissolution in the water column and sediments could lead to a cool bias on the Mg/Ca SSTs (Dekens et al., 2002; Regenberg et al., 2006, 2009). However, the cool KM5c anomalies also occur if the forward-modelled core-top Mg/Ca SSTs from BAYMAG are used as the pre-industrial “reference” (Fig. S6). The cold low-latitude anomalies for KM5c could reflect an increase in the calcification depth of the foraminifera, since the surface-dwelling foraminifera analysed here calcify at a range of depths, particularly in the tropics where the thermocline is deep in comparison to mid-latitudes to high latitudes (Fairbanks et al., 1982; Curry et al., 1983). The negative anomalies are broadly smaller for G. ruber (-0.4 to -1.2 ∘C) than for T. sacculifer (-0.6 to -3.5 ∘C), consistent with a deeper depth habitat for the latter (Curry et al., 1983), although at Site 959 the G. ruber anomaly using BAYMAG is -3.8 ∘C. There is therefore a lack of consistency between sites, which is difficult to resolve when single species have been analysed for each of the sites through KM5c.

Where there are very large differences between BAYMAG and published Mg/Ca SST estimates, regardless of latitude (e.g. North Atlantic; Fig. 4), we suggest that some combination of calibration difference, Mg/Caseawater change, and/or other environmental factors including seasonality and calcification depth may offer an explanation. To fully investigate the cause(s) of offsets in Mg/Ca SST reconstructions requires future multi-species analysis for Mg/Ca for each site and multi-proxy analysis for each site. Such an approach would enable the exploration of a wider range of potential influences over both the Mg/Ca and U37K′-SST reconstructions and a reduction in the uncertainties of the reconstructed SSTs and their anomalies. Alongside foraminifera Mg/Ca and U37K′ analyses, additional proxies which are likely to add valuable information about water column structure and seasonality could include TEX86 (Schouten et al., 2002), long-chain diols (Rampen et al., 2012), and clumped isotopes (Tripati et al., 2010). Previous research has demonstrated that even within a single site there can be offsets between proxies which are not continuous through time (e.g. Lawrence and Woodard, 2017; Petrick et al., 2018), so that high-resolution and multi-proxy work is required to fully understand the offsets we have identified here. Resolving the causes of the different proxy–proxy and proxy–model offsets is important because they impact the calculation of the global mean SST anomaly relative to pre-industrial values; however, even with the inclusion of the overall cooler Mg/Ca data, the combined KM5c proxy data still indicate a global mean SST anomaly which is larger than most models from the PlioMIP2 experiments (Fig. 3).