Previous NALPS studies include MIS 2 in Luetscher et al. (2015; though this was not branded as “NALPS”), MIS 3 in Moseley et al. (2014), and MIS 4 to MIS 5 in Boch et al. (2011). The MIS 4/MIS 5 chronology (which is the part revised here), was constructed from seven speleothems from four cave sites including St. Beatus caves, Große Baschg cave (Baschg cave for short), Klaus Cramer cave and Schneckenloch (Boch et al., 2011). In this study, two additional samples from Baschg cave and one from Schneckenloch were analysed, plus one sample from Hölloch im Mahdtal (Hölloch cave for short), one from Grete–Ruth shaft, and six from Gassel Tropfsteinhöhle (Gassel cave for short). All cave sites are located on the northern rim of the European Alps (Fig. 1) and have small catchments of less than a few square kilometres. The distance between the most westerly and easterly caves is ca. 475 km. Details of the speleothems analysed in this study and their respective caves are given in Table 1, whereas images of the respective samples are given in Fig. S1 in the Supplement.

The 11 stalagmites were cut in half along their growth axis and polished by a professional stone mason. Pilot dating studies guided the sample size that was needed for high precision ages. Sub-samples of between 20 and 150 mg were hand-drilled using a handheld drill fitted with carbide burr-tipped drill bits of diameter 0.5 to 0.8 mm in a laminar-flow hood. The cleanest, densest growth layers were targeted for sampling.

Chemical procedures and aliquot measurements were undertaken in the Trace Metal Isotope Geochemistry Laboratory at the University of Minnesota. Separation and purification of U and Th aliquots from the sub-samples were undertaken using standard methods (Edwards et al., 1987) in a clean air environment. Samples were spiked with a dilute mixed 229Th-233U-236U tracer to allow for correction of instrumental fractionation and calculation of U and Th concentrations and ratios. Procedural chemistry blanks were on the order of 5–83 ag for 230Th, 2–523 fg for 232Th, 73 to 171 ag for 234U, and 0.2 to 1.6 pg for 238U. Aliquots of U and Th were analysed on a Thermo Fisher Neptune multi-collector inductively coupled plasma mass spectrometer (MC-ICPMS) in peak-jumping mode on the secondary electron multiplier (Shen et al., 2012).

Stable isotopes (δ18Ocalc and δ13Ccalc) were typically micro-milled at a spatial resolution of 250 µm (with the exception of GAS22 =200 µm and BA7 =500 µm) from the central axis of each sample (Fig. S2). In total 5000 new measurements were made for this study at the University of Innsbruck on a Thermo Fisher DeltaplusXL isotope ratio mass spectrometer linked to a GasBench II. Analytical precisions are 0.08 ‰ and 0.06 ‰ for δ18Ocalc and δ13Ccalc respectively (1σ; Spötl, 2011). All isotope results are reported relative to the Vienna PeeDee Belemnite standard. In addition to the main isotope track along the central axis, Hendy tests (Hendy, 1971) were also prepared for each sample as a first-order assessment of whether the respective stalagmite was deposited under conditions of isotopic equilibrium, though the preferred approach in recent years has been to reproduce the data in a second stalagmite (Dorale and Liu, 2009). Under the Hendy test criteria, δ18Ocalc values should remain constant along a single growth layer, and there should be no correlation between δ18Ocalc and δ13Ccalc that might otherwise indicate kinetic fractionation. Bayesian age models were constructed for all 11 samples using OxCal version 4.2 for Poisson-process depositional models (“p sequence”) and a variable “k parameter” of 0.001 to 10 mm a-1 (Bronk Ramsey, 2008; Bronk Ramsey and Lee, 2013).

The new records produced in this study (Fig. 2b) comprise 5000 δ18Ocalc measurements dated by 145 precise U–Th ages (Fig. S3, Table S1), which add to the NALPS chronology of Boch et al. (2011; Fig. 2a) that comprised 7141 δ18Ocalc measurements and 154 U–Th ages. Combined, the two chronologies cover the period 118.3 to 63.7 ka. Within this interval, the record is now 90 % complete, compared to 76 % in Boch et al. (2011). Where speleothems grew synchronously, major transitional events between stadials and interstadials (and vice versa) are all in agreement within uncertainty, which can be very clearly seen in Fig. S5. In the interest of completeness and transparency, we present all δ18Ocalc records here; however, some samples are cleaner than others as discussed in Sect. 4 (i.e. low in detrital Th as indicated by a higher 230Th/232Th ratio) and thus deemed to be more reliably dated. The NALPS19 chronology is therefore constructed from only the most reliably dated records from this study and Boch et al. (2011; Fig. 2c). Considering the construction of NALPS19 further and generally working from youngest to oldest, samples KC1 and HÖL19 are included on the basis that they are the only records available that cover the transitions into stadials 19 and 20. The transition into interstadial 20 is present in both SCH6 (this study) and BA1b (Boch et al., 2011). Both samples have comparable levels of detrital Th, and the dating precision of the transition in both samples is ca. 200 to 250 years. Given the comparative cleanliness and dating precision, as well as the reproducibility of the timing of the transition to within ca. 50 to 85 years (Fig. S5), both samples are included in NALPS19. Samples covering the transition into stadial 21 include GAS12 and 13 (this study) and BA1b (Boch et al., 2011). Samples from GAS12 and GAS13 are extremely clean with dating precisions of 250 to 300 years (Table 2), whereas those from BA1b are generally moderate to very clean. Critically though, GAS12 contains six ages and over 60 δ18Ocalc measurements within the transition, and GAS13, three ages and over 130 δ18Ocalc measurements (Fig. S2). On the other hand, BA1b has only three δ18Ocalc measurements in the transition, and one age which is quite dirty resulting in an age corrected to younger values by 760 years and a dating precision of 580 years (Boch et al., 2011). Based on the higher resolution and higher precision provided by GAS12 and GAS13, as well as the fact they are reproducible during the transition on subdecadal and decadal timescales, we therefore include GAS12 and GAS13 in NALPS19 and omit BA1b. EXC4 is then included for the interstadial 21 portion on the basis that it is clean. However, for this section it only contains the interstadial and no transitions, therefore it is excluded from the discussion on transition timing (Sect. 5.2). The transition into interstadial 21 is captured in BA1 (Boch et al., 2011), BA7 and GAS25 (this study). As discussed above, GAS25 is extremely clean, thus correction for detrital Th is negligible and the dating precision is on the order of 300 to 400 years (Table S2). In contrast, BA7 is the dirtiest of the samples with large corrections for detrital Th (Table S1), whereas BA1 is moderately dirty resulting in comparable shifts to younger ages (Boch et al., 2011). Ideally, the complete transition would be constrained only in GAS25 since this sample is the most reliable and best dated, but unfortunately this record is limited to growth mainly during and just after the transition. We therefore include GAS25 where it is applicable and omit BA1 and BA7, but then keep BA1 and BA7 for the parts of the record where there is no alternative available. The transition into stadial 22 is present in GAS25, BA5 (this study), and BA2 (Boch et al., 2011). The situation here is similar to the transition into interstadial 21, where GAS25 is the superior sample with higher dating quality. GAS25 therefore takes priority, whereas BA5 is included to complete the stadial part of the record. BA2 is completely omitted from NALPS19 on the basis that correcting for detrital Th causes shifts in ages in terms of centuries (Boch et al., 2011) as compared to decades in GAS25 (Table S1). The section of the record encompassing interstadial 23, stadial 24, and interstadial 24 is fully covered by GAS22, GAS27, GAS29, and HUN14, which are all extremely clean, well-dated records with typical dating precisions of 300 to 400 years (Table 2). Furthermore, the timing of the transition into interstadial 23 is reproducible to within 60 to 100 years between GAS27 and HUN14. The timing of the transition into stadial 24 is in agreement on the order of 40 to 60 years in GAS22, GAS29, and HUN14. Furthermore, the pattern of δ18Ocalc shifts across the whole interstadial 24 to 23 period is remarkably similar in the new speleothems analysed here to the pattern of events in NGRIP δ18Oice across the same period. This suggests the new speleothem samples are capturing a larger-scale climate signal, unlike EXC3 and EXC4 from St. Beatus cave (Boch et al., 2011), which show a distinctly different pattern in δ18Ocalc across this time period. The reason for the difference is unknown, and is likely due to some local influence or control at the cave site. We acknowledge that there is still value in the St. Beatus records, but they are not ideal for investigations into leads, lags, and synchronicity when more comparable records exist; thus they are not included in NALPS19. Finally, the new record from HUN14 is used to complete the gap that existed previously at stadial 25.

In summary, important updates in the NALPS19 chronology (Figs. 2 and S6) therefore include (1) the addition of the cooling into GS-20, (2) a revision of the GI-20c/GS-21.1/GI-21.1a period using multiple cleaner samples, (3) revision of the warming into GI-21.1e and cooling into GS-22, also using a cleaner sample, and (4) revision of the interval GI-23.1 to GI-25c, which includes the addition of the previously absent GI-25a and b and a more distinctive “shape” to GS-24 in line with NGRIP.

Figure 3 (split into 20 000-year time slices in Fig. S6) shows the NALPS19 δ18Ocalc record in comparison to other well-dated δ18O records from distant Northern Hemisphere regions over the interval 60 to 120 ka. Comparison of NGRIP and NALPS19 shows that the broad-scale pattern of shifts in δ18O was remarkably similar during this period, including down to centennial-scale events. Differences do, however, arise when considering the timing and duration of events, which we investigate further by applying the ramp-fitting function of Erhardt et al. (2019). The ramp-fitting function is similar to the one used by Mudelsee (2000), but instead uses probabilistic inference to define a transition via a linear ramp between two constant levels. Such an approach enables a consistent approach to chronological quantification of climate transitions (Mudelsee, 2000), unlike the more subjective approach of taking the first data point that deviates from the baseline of the previous climate state (e.g. Capron et al., 2010a; Moseley et al., 2014; Rasmussen et al., 2014). Adolphi et al. (2018) applied such a ramp-fitting method to the younger, late glacial portion of the NGRIP δ18Oice record (Adolphi et al., 2018), whereas Steffensen et al. (2008) applied another ramp-fitting method through the last deglacial. For this study, ramp-fitting was applied to (1) δ18Ocalc of the new NALPS19 record (this study); (2) δ18Ocalc of the Asian monsoon composite speleothem record (Cheng et al., 2016); (3) NGRIP δ18Oice on the GICC05modelext chronology, which is comprised of a composite layer-counted chronology to 60 ka (Svensson et al., 2008) followed by splicing of the ss09sea-modelled chronology (Johnsen et al., 2001) between 60 and 122 ka onto the younger annual-layer-counted chronology (Wolff et al., 2010); (4) NGRIP δ18Oice on the AICC2012 chronology, which is constructed using glaciological inputs, relative and absolute gas and ice stratigraphic markers, and Bayesian modelling (Veres et al., 2013); and (5) NGRIP δ18Oice on the AICC2012 chronology updated by aligning δ18O of the atmosphere as measured in EPICA Dome C with δ18Ocalc of Chinese speleothems (Extier et al., 2018).

Results of the ramp-fitting are shown in Table 3, Figs. 4 and 5, and S7. Unfortunately, results are not available for some transitions because their shape is incompatible with the transition model, which requires stable periods before and after the transitions. Where multiple NALPS19 speleothems grew synchronously, excellent agreement is found in the absolute timing of the transitions, which shows differences from as little as 10 years between GAS12 and GAS13 during the endpoint of the transition into GS-21.1, up to a maximum of only 163 years difference between GAS22 and HUN14 during the endpoint of the transition into GS-24.1 (i.e. within the 318 years uncertainty of GAS22 at this point; Table 3, Fig. 4). Similarly, for the NGRIP δ18Oice record on the GICC05modelext chronology, we find that the timing of the start of the respective transitions are in excellent agreement (2 to 119 years) between the ramp-fitting used in this study and the INTIMATE event stratigraphy scheme (Rasmussen et al., 2014; Table 3). Comparison between the timing of the ramp-fitted transitions in NALPS19 and the Asian monsoon speleothem records also show excellent agreement within centennial-scale uncertainties, with the exception of GS-20, which is older in NALPS19 by ca. 900 years (Table 3 and Figs. 4 and 5). The NALPS19 age for GS-20 is, however, in very good agreement on a multi-decadal scale with the GICC05modelext chronology (details below). It should be noted that a comprehensive comparison of the timing of transitions between NALPS19 and NGRIP on the three ice core chronologies is made difficult because of the large uncertainties associated with AICC2012 (ca. 3000–3200 years; Veres et al., 2013) and even the absence of uncertainties associated with GICC05modelext (Wolff et al., 2010). To deal with the absence of uncertainties in GICC05modelext, we take the approach of Abbott et al. (2012) and extrapolate the linear trend in ratio between age and uncertainty from the layer-counted 0–60 ka GICC05 chronology (Svensson et al., 2008), which yields an uncertainty of ca. 4.5 % by 120 ka (Table 3, Fig. 4). In reality, the uncertainty is likely to be considerably less since well-dated markers exist in some places (e.g. Guillou et al., 2019). Nevertheless, if only the central age is considered (where + indicates the respective chronology is earlier/older than NALPS19, and - is vice versa), excellent agreement in the absolute timing of the transition is displayed between NALPS19 and GICC05modelext for GS-20, which is offset by ca. -45 years, and GI-21.1, which is offset by ca. +20 to +80 years (Table 3, Figs. 4 and 5). Depending on the speleothem to which the comparison is made, the transition into GI-23.1 is offset by ca. +230 to +290 years (HUN14) or ca. +340 to +390 years (GAS27). The other transitions into GI-20 (+560 to +650 years), GS-21.1 (+490 to +570 years), GS-24.1 (-440 to -460 years), and GI-24.2 (ca. -550 years) display the largest of the offsets (Table 3 and Figs. 4 and 5). Comparison between NALPS19 and NGRIP on AICC2012 shows good agreement in the timing of GS-21.1 (ca. +8 to -40 years) and GI-20 (+50 to +140 years). The timing for GS-21.1 is further supported in this instance by the close agreement also of the Asian monsoon composite chronology (-70 to -150 years; Fig. 5). Elsewhere, the transitions in the NALPS19 chronology are consistently earlier than their counterparts in the AICC2012 chronology i.e. GS-20 (ca. -400 years), GI-21.1 (ca. -500 years), GI-23.1 (ca. -1900 years), GS-24.1 (ca. -2700 years), and GI-24.2 (ca. -2960 years), suggesting that some revision of the AICC2012 chronology may be needed. Extier et al. (2018) have also proposed such a revision for the period 100 to 120 ka, which is the interval in which there is the greatest discrepancy between AICC2012 and NALPS19. Application of the ramp-fitting to the AICC2012 chronology revised by Extier et al. (2018; AICC2012Extier) shows that there is much better agreement with NALPS19 during the 100 to 120 ka interval than existed for AICC2012 (Figs. 4 and 5). Specifically, the offset for GI-23.1 is ca. +350 years, GS-24.1 is ca. -400 years, and GI-24.1 is ca. -700 years.

The ramp-fitted transitions have also enabled an assessment of the duration of δ18O transitions in the respective chronologies (Table 3, Fig. 5). The quickest shift of 21 years is displayed for the AICC2012Extier transition into GI-20, whereas the longest shift of 489 years is present in NALPS19 for the transition into GS-19.2. Consistency in the duration of transitions between NALPS19 and the Greenland chronologies is displayed in particular for GS-20 (211 to 236 years), GS-21.1 (204 to 243 years), GS-24.1 (73 to 96 years), and GI-24.2 (32 to 47 years; Table 3, Fig. 5). The difference in durations for GI-21.1 (54 to 107 years) and GI-23.1 (68 to 138 years) is slightly larger but still comparable on multi-decadal timescales. The greatest difference between NALPS19 and the Greenland chronologies is displayed for GI-20, which varies between 21 and 114 years. Interestingly, with the exception of GI-24.2, the duration of transitions in the Asian monsoon are also comparable to the North Atlantic-sourced NALPS19 and Greenland chronologies (on multi-decadal and multi-centennial timescales; Table 3, Fig. 5).

The NALPS19 chronology also enables new consideration of the duration of GS-22, which previously has been the subject of debate given the various different timescales presented in the literature (Boch et al., 2011; Vallelonga et al., 2012). Here, we use the same strategy as for the previous studies and define the duration of GS-22 as being from the midpoint of the δ18O transition into GS-22 until the start of the δ18O transition into GI-21.2 (Capron et al., 2010a; Vallelonga et al., 2012). The precursor event is defined as the start of the δ18O transition into GI-21.2 until the midpoint of the δ18O transition into GI-21.1e. All uncertainties are at the 95 % confidence interval. Based on multi-proxy annual layer-counting, Vallelonga et al. (2012) proposed the duration of GS-22 in the NGRIP ice to be 2894±198 years and the precursor event to be 350±19 years (together 3244±199 years, 2σ error; Table 4). The Vallelonga et al. (2012) layer-counted chronology thus indicated a longer duration for GS-22 than the GICC05modelext chronology (2620 years) and a shorter duration for the precursor event (300 years, together 2920 years; Table 4; Wolff et al., 2010). The ramp-fitting function was not able to constrain the transition into the precursor event (GI-21.2); thus we consider here the duration of the full period from the cooling into GS-22 to the warming into GI-21.1e, which in the NALPS19 chronology is 3993±155 years (Table 4). This finding is in agreement with the duration from the previous NALPS chronology of 3660±526 years (Table 4), but is ca. 1000 years longer than in GICC05modelext and 750 years longer than in the layer-counted chronology (395 years if the uncertainties are considered). In contrast, a relatively long duration of 4122±650 years has been proposed for NGRIP on the EPICA Dronning Maud Land (EDML) Antarctic ice core chronology (Capron et al., 2010b; Vallelonga et al., 2012), which is in agreement with the duration from NALPS19. Additionally, the duration of the same period as estimated from the Asian monsoon composite record is 4489±960 years. The speleothem δ18Ocalc records from both the Alps and China therefore support a longer duration for the period between the cooling into GS-22 to the warming into GI-21.1e, which is in line with the NGRIP-EDML chronology (Capron et al., 2010b; Vallelonga et al., 2012).

Speleothem deposits from the northern rim of the Alps now provide a near-complete record of δ18Ocalc variability during the last glacial period (Fig. 6; Boch et al., 2011; Moseley et al., 2014; Luetscher et al., 2015), which is remarkably similar to δ18O variability recorded in the NGRIP Greenland ice core during the same period. It is hypothesised that the moisture source for both regions during the last glacial period was the North Atlantic, with the primary control on the δ18O of precipitation in both Greenland and the Alps being temperature (Boch et al., 2011). Changes to the transport pathway have, however, been proposed for the northern Alpine speleothem record of the Last Glacial Maximum (LGM) between 26.5 and 23.5 ka induced by a southward shift in the North Atlantic storm track (Luetscher et al., 2015). The change to the transport pathway is, however, only considered to affect the LGM and not the remainder of the glacial (Luetscher et al., 2015).

We now consider the full glacial Alpine speleothem δ18Ocalc record in further detail. In addition to the NALPS records of Boch et al. (2011), Moseley et al. (2014) and NALPS19 (this study), we also consider an MIS 5 record from Siebenhengste (Fig. S9), a large cave system located on the northern rim of the Alps of Switzerland (Fig. 1), and a record from Kleegruben cave (Spötl et al., 2006), which is located in the Central Alps of Austria to the north of the main Alpine crest (Fig. 1). A thorough investigation of the controls on the δ18O of precipitation would require a sophisticated modelling approach, which is beyond the scope of this paper; thus here we appreciate that our investigation is a first consideration only. Furthermore, given the many different factors that can influence the δ18O of precipitation (Dansgaard, 1964; Rozanski et al., 1993; Clark and Fritz, 1997), it would be advantageous to have stable isotope information from fluid inclusions. Unfortunately, the speleothems presented here are largely devoid of fluid inclusions (Brandstätter, unpublished data).

Today, temperature has been shown to have the most dominant control on the δ18O of precipitation along the northern rim of the Austrian Alps (Kaiser et al., 2002; Hager and Foelsche, 2015), though other factors such as a changing moisture source, rain-out along the different transport pathways (continental effect), altitude (altitude effect), the North Atlantic Oscillation, and locally also the amount of rain (amount effect) all show some additional control (Ambach et al., 1968; Dray et al., 1998; Kaiser et al., 2002; Hager and Foelsche, 2015; Deininger et al., 2016). To consider the effects of these controls on the δ18O of precipitation during the last glacial period, we have first removed from the speleothem records the variability in mean ocean δ18O caused by fluctuations in ice volume (Fig. 6) using a rate of 0.012 ‰ m-1 (Rohling, 2013) and the sea-level curve of Grant et al. (2012).

Mean speleothem δ18Ocalc values for individual stadials and interstadials in the ice-volume-corrected record have been calculated for each sample (Figs. 7a, S8, Table S3). Excluding the samples associated with the LGM because of the different transport pathway (Luetscher et al., 2015), the δ18Ocalc range in mean interstadial values is 5.0 ‰ (Klaus Cramer (-7.9 ‰) and Siebenhengste (-7.9 ‰) to Kleegruben (-12.9 ‰), whilst the range in mean δ18Ocalc stadial values is slightly larger (but comparable) at 5.4 ‰ (Siebenhengste (-9.5 ‰) to Kleegruben (-14.9 ‰); Fig. 7a). We now consider the controls on δ18Ocalc during periods when more than one speleothem was deposited, specifically GI-23.1, GS-23.2, GI-24.1, and GS-24.1. Generally it is considered that the dominant control on the δ18O of precipitation in the northern and central Alps during the last glacial period was temperature, and the dominant moisture source was the North Atlantic (as both are today). The correlation between both temperature and distance from the North Atlantic as compared to mean δ18Ocalc was investigated to identify potential continental and rainout effects. In all instances, mean δ18Ocalc became increasingly lighter with increasing distance from the North Atlantic; a medium correlation is displayed for GI-23.1 (r2=0.64, n=4), GS-23.2 (r2=0.63, n=3), GS-24.1 (r2=0.57, n=4, two samples for Gassel), and a lower correlation during GI-24.1 (r2=0.16, n=3). This trend of lighter mean δ18Ocalc with increasing distance from the source would be expected with progressive rainout and is consistent with present-day observations.

Today, spatial variability of the δ18O of precipitation in the Austrian Alps is highly dependent on altitude (Hager and Foelsche, 2015). We find that medium to strong correlations between catchment elevation and mean δ18Ocalc existed during GI-23.1 (r2=0.49, n=4), GI-24.1 (r2=0.67, n=3), GS-23.2 (r2=0.79 n=3), and GS-24.1 (r2=0.74, n=4 (Gassel has two samples); (Fig. 7c)). For GI-24.1, the relationship shows that mean δ18Ocalc becomes increasingly lighter with increasing elevation (as would be expected for altitudinal controls on δ18O of precipitation). In contrast, the other examined time periods show an inverse relationship to what would be expected for altitudinal control, with mean δ18Ocalc becoming heavier with increasing elevation (Fig. 7c). Since GI-24.1 is the only event that does not include the high-elevation Siebenhengste site, the mean δ18Ocalc of 7H-12 was removed from the linear regression analysis for the three time periods showing an inverse relationship (Fig. 7d). This resulted in a switch to increasingly lighter mean δ18Ocalc with increasing elevation for GI-23.1, GS-23.2, and GS-24.1 (Fig. 7d; i.e. in line with an altitudinal control on δ18O of precipitation).

Given that there is such limited availability of multiple speleothem δ18O records covering the same time periods, it is difficult to make firm conclusions on the controls of δ18Ocalc. Here though we offer some hypotheses based on the available data. We have shown that for a given time period δ18Ocalc trends towards lighter values with increasing distance from the North Atlantic (Fig. 7b). Despite this, there is some variability overprinted on top of this trend. For instance, even though Grete–Ruth is closer to the North Atlantic than Gassel cave, mean δ18Ocalc values for Grete–Ruth are consistently lighter than for Gassel (Fig. 7b). Since Grete–Ruth is located at a higher elevation than Gassel cave (Fig. 7c), the lighter mean δ18Ocalc values are likely a product of the altitude effect and associated cooler temperatures. In comparison, St. Beatus and Siebenhengste caves are located within 10 km of one another, and are the closest caves to the North Atlantic of all the caves investigated here. As expected, mean δ18Ocalc values are heavier for St. Beatus and Siebenhengste than for Grete–Ruth and Gassel caves (Fig. 7b). Closer investigation, however, shows that during GI-23.1, mean δ18Ocalc of low-elevation St. Beatus is lighter than the high-elevation Siebenhengste (Fig. 7b). Given the close proximity of the two caves, the condensation level (and therefore condensation temperature) would have been approximately the same, and thus one must consider the reason for the difference in mean δ18Ocalc for these two caves. Since the three caves at lower elevation (St. Beatus, Gassel, Grete–Ruth) follow the expected altitude-induced trend of lighter mean δ18Ocalc with increasing elevation (Fig. 7d), it seems the anomaly lies with the high-elevation 7H-12 stalagmite from Siebenhengste. One reason for the heavier-than-expected mean δ18Ocalc at Siebenhengste could be that the full annual signal is better preserved at high-elevation sites that are less exposed to evapotranspiration effects during the summer season than in more vegetated catchments. Alternatively, a summer bias towards isotopically heavier δ18Ocalc at the high-elevation site could for instance be caused by wind erosion resulting in relocation of the isotopically light winter snow, a process that has been well documented at various Alpine sites (Ambach et al., 1968; Bohleber et al., 2013; Hürkamp et al., 2019). Eventually, if firn developed above Siebenhengste during GI-23.1, then this would also limit the input of isotopically light precipitation causing a summer bias in the recorded signal. At present there is, however, no evidence to either support or reject the hypothesis of firn above Siebenhengste during MIS 5.

In summary, speleothems from the northern rim of the European Alps provide δ18Ocalc records for the majority of the last glacial period. As expected, the limited data set shows that mean δ18Ocalc for specific stadials and interstadials generally trends towards lighter values with increasing distance from the coast and with increasing altitude. An exception is the high-elevation 7H-12 stalagmite from Siebenhengste, which appears to record a stronger summer signal. Further investigation of the controls on δ18Ocalc in the northern Alps requires a more sophisticated modelling approach.

The recognition of centennial- to millennial-scale climate events, such as precursors to interstadials and “within-interstadial” depletions in δ18Oice (Capron et al., 2010a), led to the designation of the INTIMATE event stratigraphy for the Greenland ice cores over the last glacial period (Rasmussen et al., 2014). Typically, a precursor event is a feature of a stadial–interstadial transition; this includes GS-16.2, 17.2, 21.2, and 23.2 (Fig. 8; Capron et al., 2010a; Rasmussen et al., 2014). It is characterised in northern Alpine speleothems and Greenland ice cores by a rapid increase in δ18O from stadial to interstadial conditions. The event is short lived, lasting a maximum of a few centuries before δ18O returns to near-stadial conditions for another few decades to centuries, followed by the main transition into the interstadial. [Ca2+] in the Greenland ice cores varies almost simultaneously with these δ18Oice changes, where increases in [Ca2+] are associated with depletions in δ18Oice and vice versa. Changes in [Ca2+] are interpreted to reflect changes in dust concentration caused by changes in dust source conditions and transport pathways indicative of regional- to hemispheric-scale circulation changes (Ruth et al., 2007). In comparison, within-interstadial climate perturbations are characterised in general by smaller-amplitude depletions in δ18Oice that typically do not reach stadial values, and are often of shorter duration than the reversals at stadial–interstadial transitions. [Ca2+] also varies in tune with within-interstadial δ18Oice fluctuations, but similarly does not reach full stadial values. Such characteristics appear to be consistent in δ18O records from both Greenland ice cores and northern Alpine speleothems. The exception to such typical within-interstadial cold perturbations is the event at 107.5 ka in the NALPS19 chronology, which is designated GS-24.2 in the INTIMATE event stratigraphy scheme (Rasmussen et al., 2014). This drop in δ18Ocalc to stadial values occurred 978 years after the start of the interstadial, thus firmly making it a within-interstadial event rather than one associated with a stadial–interstadial transition. At present, the 107.5 ka event (GS-24.2) is the only centennial-scale δ18O event of such extreme amplitude occurring during an interstadial that is recognised in both Greenland and northern Alpine records. Because of this, it has been likened to the 8.2 ka cold event that occurred in the early Holocene (Alley et al., 1997; Capron et al., 2010a). Still, the δ18Oice excursion of the 8.2 ka event did not reach near-stadial values in NGRIP as GS-24.2 did, thus highlighting some differences between these two warm-interrupting cold reversals. In addition, Rasmussen et al. (2014) liken the within-interstadial GS-24.2 cold perturbation to stadial–interstadial transition events GS-16.2 and GS-17.2. Both the similarities between GS-24.2 and the 8.2 ka event and those of GS-16.2 and GS-17.2 suggest that such abrupt climate variability is not critically influenced by the size of the Greenland ice sheet (Capron et al., 2010a; Rasmussen et al., 2014).

During the deglacial and early Holocene, large-scale meltwater events are widely suggested as being responsible for causing some short-term climate reversals through the weakening of Atlantic meridional overturning circulation (AMOC; e.g. Broecker, 1994; Teller et al., 2002; Clark et al., 2001). Such cold reversals thought to be triggered by meltwater events include the Older Dryas at 14 ka (GI-1d, Rasmussen et al., 2014), the Preboreal Oscillation at 11.4 ka (e.g. Johnsen et al., 1992; Björck et al., 1996; Fischer et al., 2002), the 9.3 ka event (Fleitmann et al., 2008; Yu et al., 2010), and the 8.2 ka event (Alley et al., 1997). In contrast, however, not all freshwater injections led to cold events, and not all cold events were caused by freshwater injections (Stanford et al., 2006). For instance, both the Younger Dryas and Heinrich events occurred during times of already-colder sea surface temperatures and weakened AMOC, indicating that the input of freshwater from the iceberg armadas was not the initial cause of the AMOC slowdown (e.g. Hall et al., 2006; Henry et al., 2016).

In the case of the centennial-scale cold reversals of GS-16.2, GS-17.2, GS-21.2, GS-23.2, and GS-24.2 (Fig. 8), a possible mechanism for each of these events could be similar to the meltwater-triggered cold reversals of the deglacial. This hypothesis is supported when considering that events GS-17.2, GS-21.2, and GS-24.2 occurred shortly following episodes of rapid sea-level rise, which were in excess of 12 m kyr-1 in the high-resolution record of Grant et al. (2012; Fig. 8). Such rapid sea-level rise does not appear to have occurred prior to GS-23.2, though closer inspection of the sea-level curve shows that following the rise prior to GS-24.2, sea levels had remained elevated and underwent a series of rapid oscillations that are smoothed out in the rate-of-change curve (Fig. 8). Likewise, GS-16.2 did not occur coincident with an episode of sea-level rise, but did occur shortly after the rise associated with GS-17.2 (Fig. 8). Additionally, the rapid rises in sea level each began at times of increased ice-rafted debris (IRD) in the North Atlantic (McManus et al., 1999, on U–Th timescale), weakened AMOC, and increased ice volume as indicated by high benthic δ13C and planktic δ18O values, respectively, as well as pluvial periods in Brazil caused by a southward shift of the intertropical convergence zone (ITCZ; Wang et al., 2004; Fig. 8). In the late glacial, such episodes are associated with Heinrich events (Wang et al., 2004). Furthermore during glacial terminations, the sequence of events has been shown to include a Heinrich event, followed by short-lived warming, and then a millennial-scale return to cold conditions, and finally the transition to the interglacial (Cheng et al., 2009). Though on shorter timescales, the pattern of events during these specific stadial–interstadial transitions is similar to the pattern of events during glacial terminations. The oscillations of GS-16.2, GS-17.2, GS-21.2, and GS-23.2 at stadial–interstadial transitions can therefore be considered as being akin to the meltwater-triggered Preboreal Oscillation, which occurred shortly following the warming at the end of the Younger Dryas stadial during a time when considerable volumes of ice still existed, similar to the early glacial. These reversals at stadial–interstadial transitions during the early glacial period are therefore not so much warming events that punctuate cold periods (Capron et al., 2010a), but rather stadial–interstadial transitions that failed due to freshwater influx. On the other hand, GS-24.2, which occurred nearly 1000 years after warming occurred, is more similar to the Older Dryas in which a cold event punctuated a warm interval.