The application of the ARM in this paper is based on the work by . The ARM requires two meteorological archives: data to predict the spatial fields and a record of the spatial data from which the reconstructions are drawn. To predict the spatial fields we used daily station observations, while the RhiresD and TabsD datasets for 1961–2017 from MeteoSwiss serve as records of spatial data (see Sect. ). For a given day in the past, we screen the period for which we have spatial data (analogue pool) for the most similar day in terms of station data (best analogue). Precipitation and temperature fields from this day serve as an estimate for the day in the past. Predictands of the analogue method are thus the unknown spatial fields, while station data and weather types serve as predictors.

Given a sufficiently large analogue pool, the ARM generally has the advantage of reproducing the mean and variance of the target, as well as naturally occurring spatial patterns . With input from both more coarsely resolved data, e.g. reanalyses or weather types, and local information (station data), it can make use of more data sources than the simple downscaling or interpolation of station observations alone.

In order to maintain the physical consistency of the reconstructions, further conditions are established.

The day of interest and possible analogue days are required to be of the same WT to ensure similar synoptic-scale atmospheric conditions, e.g. wind fields . To account for the uncertainty in WT reconstructions, we did not restrict the analogue to the most probable WT but accepted additional WTs such that they cover the true WT with a combined probability of at least 95 % according to .

Following these conditions, the best analogue is defined as a day within the analogue pool, with the same weather type, and within the same time window that shows the most similar values of certain meteorological variables from a defined set of stations to the day of interest.

Before the application of the ARM, station and gridded data are preprocessed. As observed variables have different scales, each measurement series is standardised. Furthermore, temperature data from weather stations and spatial fields are decomposed into a smoothed mean climatology and the respective anomalies. For each observation series and each cell of the gridded data, a smoothed mean seasonality curve is estimated by fitting the first two harmonics of temperature time series following Eq. () using linear regression. 1S=c0+c1sin⁡(2πdoyndoy)+c2cos⁡(2πdoyndoy)+c3sin⁡(4πdoyndoy)+c4cos⁡(4πdoyndoy) Here, doy denotes the day of year, ndoy is the number of days in a year, and c0, c1, c2, c3, and c4 are parameters to estimate. After the calculation of the analogue reconstructions using temperature anomalies, mean climatology is then again added to the reconstructed temperature deviation fields to get absolute temperature data. This procedure slightly alters the characteristics of the ARM, as adding climatology and resampled temperature deviations not only resamples known temperature fields, but also creates new ones. An elimination of the signal from climatic temperature changes over the last centuries did not improve the results and was not further pursued in this study.

With preprocessed data, the analogue method is applied. Following , the root mean squared error (RMSE) is used as a measure of similarity (Eq. ). This distance measure was chosen to reduce the levelling-out of station data resulting in decreased variance of reconstructions. As the RMSE is sensitive to large deviations, station data of the best analogues will closely follow observed ones, thus also reproducing local phenomena and extremes. To a certain extent, however, an underestimation of variance due to the levelling-out of station observations cannot be avoided, especially for days with less suitable analogues and for skewed distributions like precipitation. 2d(x,y)=∑i=1n(xi-yi)2 Here, x and y are vectors of observations from the day of interest and a day within the analogue pool, respectively; i denotes the different observations within this vector. Other measures of similarity like the Mahalanobis distance were not examined.

There are some limitations to the method, the most crucial one being the size of the analogue pool, which has to be large enough to provide reasonably matching analogues to a given atmospheric state . This is especially relevant for extreme events as they occur more rarely, and therefore less suitable analogues exist. Also, the coverage by observation series plays an important role, as stations might fail to capture local events like thunderstorms. Furthermore, the methodological setting described above, e.g. the coupling of precipitation and temperature reconstructions and the choice of the similarity measure, can have an impact on the capability of the ARM to correctly reproduce the variance and mean, as well as spatial distribution of the variables. As the coupling was established for reasons of physical consistency and the distance measure was chosen to maintain natural variability, these effects are not further assessed in Sect. 4. Finally, temporal consistency is not guaranteed.

The best analogue may not perfectly fit all observations. To further improve the temperature reconstructions, we borrow from data assimilation techniques see, e.g.. The method used here is based on the ensemble Kalman filter , which is applied e.g. for data assimilation of ensemble forecasts from dynamical models. Here we use the best analogue in the same way as the forecast (termed background or first guess) and best n analogues as an ensemble. However, neither the analysis nor the covariance matrix (see below) is passed on to the next time step. This simplification is called ensemble Kalman fitting (EnKF) or offline data assimilation . The EnKF essentially minimises a least-squares problem . The state vector x that minimises the following cost function J is optimal in the case of Gaussian errors: 3J(x)=(x-xb)T(Pb)-1(x-xb)+(y-H[x])TR-1(y-H[x]), where xb is the first guess (background), in this case reconstructions from the best analogue. Pb is the background error covariance matrix that in this particular case is estimated from an “ensemble” of the best n analogues, i.e. temperature fields from the n most similar days to the day of interest. Vector y contains station observations, and operator H is used to extract the observations from the model space. R is the error covariance matrix of y-H(x). With station observations and the ensemble of the best n analogues, a new estimation of temperature xb that is the best estimate for the true atmospheric state x is calculated from Eq. (): 4xb=xb+K(y-Hxb), where xb denotes the updated state vector (analysis), xb and y are as described above, and K is the Kalman gain or innovation matrix calculated from the ensemble. In this and the following equations, H describes the Jacobian matrix of H(x) and extracts the values from the grid cell closest to the observation site of y.

We use an implementation in which each observation is assimilated sequentially (Eqs. , –). The fitting procedure is into two steps: an update of the ensemble mean x‾ (Eq. ) and an update of the anomalies x′ with respect to the ensemble mean (Eq. ). Equations () and () depict the calculation of the Kalman gain K for the ensemble mean and K̃ for the anomalies. 5ax‾a=x‾b+K(y‾-Hx‾b)5bx′a=x′b+K̃(y′-Hx′b)  with: y′=0 6aK=PbHT(HPbHT+R)-16bK̃=PbHT[(HPbHT+R)-1]T×(HPbHT+R+R)-1 Here, x‾a and x‾b denote the analysis and background of the ensemble mean and x′a and x′b the corresponding anomalies. Pb and R are the error covariance matrices as in Eq. (). The observation error R is roughly estimated to be 1 ∘C. The background error covariance matrix Pb is calculated from the best n analogues following Eq. (), where i and j denote grid boxes and k the ensemble members. 7Pi,jb=1nens-1∑k=1nens(xi,kb-x‾kb)(xj,kb-x‾kb) Following , instead of the full error covariance matrix, the conversion PbHT is calculated directly in order to save computational resources. Covariance matrices estimated from small samples may exhibit spurious covariances far away from the observation. Spatial localisation is often used to minimise these effects. In our case, the study areas is too small and the ensemble size sufficient such that localisation is not necessary (tests using a Gaussian weighting function did not show improvement).

For each day, the EnKF is applied to the analogue reconstructions using a selection of measurement series that exhibit an average monthly correlation with co-located data from TabsD above 0.975 (see Fig. ). This is to avoid measurement series subject to local influences, which are not resolved by spatial data and would thus lead to erroneous assimilations. To account for a bias between local measured temperature at a weather station and spatially aggregated temperature values of the corresponding grid cell, station data are adjusted by subtracting the mean bias between the measurement and grid cell value from the TabsD dataset over the period 1961–2017 for each month. This procedure prevents systematic biases in fitted temperature fields. The ensemble size is set to the 50 best analogues.

Given the limitations of the ARM described in the previous section, precipitation reconstructions might also be affected by difficulties, e.g in reproducing extreme events due to the limited availability of matching analogues, and as a consequence have biases in the mean and issues regarding areas that are sparsely covered by weather stations. Therefore, analogue reconstructions of precipitation are post-processed. Although attempts have been made to assimilate precipitation with the application of a Kalman filter , in this paper, a much simpler approach is used: quantile mapping (QM). This method of model output statistics aims to transform the cumulative distribution functions (CDFs) of modelled precipitation to match the CDFs of observed precipitation by finding a statistical transfer function h ; i.e. it is mapping modelled to the observed distribution. As pointed out by , this procedure is asynchronous; that is, it does not consider any chronological aspects of precipitation. In its simplest application, QM corrects the model bias according to observed precipitation values and can be generally expressed by Eq. (): 8Po=h(Pm), where Po and Pm are observed and modelled precipitation, respectively, and h the transfer function . Based on the probability integral transform theorem , the transformation can be described as 9Po=Fo-1(Fm(Pm)), with Fm the CDF of modelled precipitation and Fo-1 the inverse CDF of the observed precipitation. To solve this equation, the distribution of the variable of interest has to be defined. In this paper, a parametric transformation using an exponential asymptotic function to estimate precipitation distribution was chosen following . This parametric transformation is described by Eq. (), where P^o denotes the best estimate of Po and parameters a, b, x, and τ are to be determined. 10P^o=(a+bPm)(1-e-(Pm-x)/τ) The best prediction of parameters a, b, x, and τ is estimated by minimising the residual sum of squares for wet days . To define wet days, a threshold for P>0.1 mm was set. Precipitation values beyond 0.1 mm were set to zero. Parametric transfer functions were calculated from all data within the calibration period 1 January 1961–31 December 2017 for each grid cell after , with Pm the values from the analogue reconstructions and Po the values from the RhiresD dataset. No discrimination between different seasons has been made. Based on the assumption that the transfer function derived from this period is robust, i.e. precipitation distribution is not subject to changes in time, these functions can then be extrapolated in time to transform the precipitation distributions of the reconstructed datasets back to 1864.

Note that the method as applied in this paper only corrects model bias. This simple application of QM was chosen to be able to extrapolate distribution correction in time, as more complex approaches would likely be less robust. To substantially improve e.g. dry–wet day discrimination or extreme values, other approaches have to be applied .

The validation of precipitation and temperature reconstructions follows common measures and strategies used in the validation of field forecasts . If not indicated otherwise, validation measures and skill scores are computed on absolute values.

We use the Pearson correlation coefficient for temperature, while for non-Gaussian-distributed precipitation the Spearman correlation is calculated. Note that for temperature, correlation is computed on anomalies from mean seasonality (compare Sect. ), so it reflects day-to-day variability rather than the seasonal cycle. Error magnitudes are indicated as root mean squared error (RMSE), as this measure is sensitive to larger errors. Furthermore, systematic biases between reconstructions and observations are evaluated.

Additionally, the mean squared error skill score (MSESS) or reduction of error statistic (RE value) is calculated for temperature reconstructions following Eq. (), allowing us to analyse the skill of reconstructions compared to mean climatology in terms of the mean squared error: 11MSESS=1-∑i=1n(xirec-xiref)2∑i=1n(xi0-xiref)2, with xrec the reconstruction, x0 a “no knowledge prediction” (in this case mean climatology), xref the reference data from TabsD, and i the time step (validation over time) or grid cell (validation over space). An MSESS value of 1 indicates a perfect reconstruction. With an MSESS of zero, the prediction skills of reconstruction and climatology are equal and values below zero denote a decline in skill compared to climatology . Note that this measure punishes variance; i.e. a reconstruction with the correct variance but zero correlation will have an MSESS of -1.

For precipitation reconstructions, another property of interest is the discrimination between wet and dry. For this purpose, the Brier score (BS) was calculated (Eq. ) that compares the predicted probability of an event to observations : 12BS=1n∑i=1n(yi-oi)2, where y and o denote the probability of rain in reconstructions and observations, respectively, and i is as above. As reconstructions do not provide probabilities, y and o are binary with 1= rain and 0= no rain with a wet–dry threshold of 0.1 mm. The BS describes the percentage of time steps (or grid cells) wrongly assigned as wet or dry.

In a first part, a leave-one-out validation was performed on daily gridded data within the period 1 January 1961–31 December 2017. For each day, the best analogue day is calculated, excluding data from 5 d before and after the day of interest, as spatial patterns from neighbouring days can be similar. Precipitation and temperature reconstructions are then validated against the RhiresD and TabsD dataset, respectively. To analyse the full time span of the dataset, reconstructions are compared to station observations in a second part. For this purpose, two independent station series from Schaffhausen and Grimsel Hospiz (see Sect. ) were used. Measurements were compared to reconstructions by extracting values from the corresponding grid cells without interpolation.

Figure  shows results from the validation over time for analogue precipitation reconstructions (a–d) and quantile-mapped data (e–h) against RhiresD data. Depicted are the rank correlation (a, e), RMSE (b, f), mean bias (c, g), and Brier score (d, h). The Spearman correlation coefficient for analogue reconstructions is 0.79 on average and attains values from 0.62 to 0.86, with maximum values in central Switzerland (a). Quantile mapping does not change the ranks of precipitation distribution, therefore the two correlation maps are identical. Regarding the RMSE (e, f), an average error magnitude of less than 5 mm in the Swiss Plateau, as well as the inner-alpine valleys and large parts of the canton of Grisons can be observed. Errors are larger in mountainous areas and in Ticino, reaching values of 6–15 mm. Post-processed data (f) reveal a negligible increase in these errors in the range of 0.1–0.6 mm. Analogue reconstructions show a negative bias between 0.2 and 0.5 mm in the Swiss Plateau (c). The underestimation is more pronounced in mountainous regions and in Ticino, with values of 0.5–1.6 mm. Using the quantile mapping approach described in Sect. , this bias could be eliminated for the given time span (g). The Brier score indicates relatively high error rates in the discrimination between wet and dry days at individual locations, with values between 0.13 and 0.23 (d). Post-processed data reveal slightly negative changes in terms of Brier scores (h).

While rank correlation values show satisfying results, the bias and RMSE patterns of ARM reconstructions could possibly be explained by an underestimation of extreme and convective precipitation, which occur in the Alpine region and in southern Switzerland, especially in summer. While quantile-mapped data correct the bias, error values still remain large. We will elaborate on this issue below by looking at seasonal patterns and extremes. Another problem of the reconstructions is indicated by the Brier score: on average, 17 % of days are wrongly assigned as wet or dry. This relatively high fraction is not improved with post-processing, as the quantile mapping approach used here is not designed to address this particular problem.

The validation of temperature reconstructions over time in Fig.  reveals a good correlation for unprocessed data, ranging between 0.76 and 0.95 with a mean of 0.91 (a). Correlation is slightly lower in Ticino and the southern valleys of Grisons. With ensemble Kalman fitting, the Pearson correlation could be increased to values between 0.83 and 0.99 with a mean of 0.96, showing similar spatial patterns (e). The error (RMSE) could also be reduced with post-processing from 1.52 to 0.96 ∘C on average (b, f). In the Swiss Plateau, the error attains values below 1 ∘C, while in the Alpine region, in the Jura Mountains, and in southern Switzerland RMSE values up to 2.7 ∘C can be observed. Unprocessed reconstructions show a systematic overestimation of temperature in the Swiss Plateau, in the Rhone valley in Valais, and in the northern valleys of Ticino with values up to 0.06 ∘C (c). On the other hand, temperatures at higher altitudes and in southern Ticino are underestimated by 0.05 to 0.15 ∘C. Post-processed data (g) show less bias and a more balanced spatial pattern, with values ranging between -0.08 and 0.03 ∘C and a mean of -0.01 ∘C. The MSESS compared to mean seasonality (d, h) is high all over Switzerland and could be increased from 0.83 to 0.93 on average using EnKF. The pattern follows correlation.

Overall, reconstructed temperature fields can be considered to very accurately reproduce the temporal evolution of the weather. Errors are relatively low, although in regions with sparse meteorological observations, larger errors are observed. Station coverage thus plays a crucial role for analogue reconstructions. The local field of larger errors in the western Jura near La Brévine might be explained by cold air pooling, which occurs frequently in this region during winter and is not captured by any of the measurement series used for reconstruction. Bias patterns suggest that ARM reconstructions have problems correctly reproducing vertical temperature gradients. A major issue here could also be inversions. In the vertical distribution of the temperature stations used, higher altitudes are only sparsely covered (see Fig. ), making the correct determination of vertical gradients and inversion heights difficult. Post-processed data seem to solve large parts of this problem, but this needs further investigation.

To find possible explanations behind the issues mentioned above and to gain more insight into the dataset, precipitation and temperature reconstructions are assessed in detail for differences between seasons and extremes.

Figure  depicts the rank correlation (a–d), RMSE (e–h), and averaged bias (i–l) over time of post-processed precipitation for each season. We see a relatively uniform correlation pattern over all seasons, with slightly higher values for summer (JJA) along the northern Prealps (c). Correlation values are lowest on the southern side of the Alps, especially in winter (a). Error values (e–h) show a similar spatial pattern throughout the year and are smallest from December to February (DJF) and slightly higher in spring (MAM) and autumn (SON); maximum values of the RMSE occur during the summer months and reach values of 8–15 mm in Ticino. The mean bias over time (i–l) still shows minor seasonal differences. In the Swiss Plateau and the Jura Mountains, the mean deviation is approximately zero. The largest differences occur in the Alpine region and in southern Switzerland, where autumn and winter precipitation show a tendency towards overestimation, while in spring and summer post-processed precipitation fields exhibit a mostly negative bias in western Valais and the Gotthard region.

The pattern of RMSE, with higher values during the warmer periods of the year and maxima in summer, supports the previous assumption that reconstructions have problems reproducing intensive or convective precipitation. The latter, which are local-scale phenomena, may not be detected by measurement stations, making station coverage an important issue in terms of obtaining reliable reconstructions.

Analysing the same for temperature (Fig. ), we see that Pearson correlation values (a–d) exhibit maximum correlation values in spring and summer, while in autumn and winter these values are slightly lower. The RMSE (e–h) is higher in winter than during the other seasons and reaches minima in the summer months. Maximum errors of up to 3 ∘C occur during winter in the Alpine region and the Jura Mountains (e). Overall, average bias (i–l) is only marginal for all seasons, with values between -0.2 and 0.1 ∘C. Generally, vertical temperature gradients seem to be corrected by Kalman fitting. However, we can see higher values and a distinct spatial pattern related to topography in winter, as can also be seen in the RMSE (e). This indicates that inversions, which occur more frequently during this season, also remain a problem in post-processed reconstructions.

The above-mentioned issue with cold air pools in the western Jura seems to be confirmed by the seasonal error patterns, although larger errors for this region persist throughout the year. Interestingly, the mean bias shows an underestimation of winter temperature for this region.

Post-processed temperature and precipitation data were further assessed by quantiles of Swiss mean temperatures and precipitation during 1961–2017 calculated from the MeteoSwiss TabsD and RhiresD datasets to analyse the accuracy of reconstructions in reproducing extremes (Fig. ). Note that, as quantiles were calculated for average values over Switzerland, local extremes do not necessarily correspond to the highest or lowest quantiles for the whole area. In the following, results from validation over space are shown to analyse the capability of reconstruction methods to reproduce spatial patterns. For comparison, validation results from the analogue method are indicated in grey.

The spatial correlation of precipitation (Fig. , top) is relatively low for low to moderate precipitation events and increases with precipitation quantiles. Looking at the RMSE, the mean and the spread of errors also increase with increasing precipitation. While the RMSE shows a median of less than 5 mm up to the 70 % quantile, for extreme precipitation events above the 95 % quantile errors attain values of 10–15 mm in the interquartile range. Compared to unprocessed data, a slight decrease in correlation and increase in the RMSE are visible. However, the bias is considerably improved. While, as argued before, analogue reconstructions indeed reveal a strong underestimation of extreme events, the median bias becomes approximately zero for all quantiles. Uncertainties, however, remain large. The Brier score reveals that for days with zero to low precipitation as well as for extreme events, the precipitation area is well represented in the reconstructions. The problems here lie in the correct reconstruction of precipitation areas for moderate events. Compared to unprocessed data, quantile mapping leads to a slightly better discrimination between wet and dry grid boxes for upper quantiles, while the BS becomes larger for lower quantiles.

From this, we can conclude that reconstructions provide accurate precipitation fields for low to moderate precipitation events. For the benefit of unbiased reconstructions, a slight decrease in correlation and an increase in the RMSE and BS have to be accepted. Extreme events, however, are underestimated by ARM reconstructions and also show large errors for post-processed data. As extreme events by definition occur more rarely, the number of suitable analogues is limited. As argued in Sect. , more station data and a bigger analogue pool would also lead to more accurate results for extremes. Different post-processing methods might help to improve reconstructions, especially regarding wet–dry discrimination and extremes.

The validation of temperature by spatial mean temperature quantiles (Fig. , bottom) shows a considerable improvement for post-processed data compared to analogue reconstructions. Correlation values exhibit slightly better correlations for extreme temperatures, while reconstructed fields for medium temperatures are less correlated with TabsD data. RMSE values are higher for upper and lower extreme values. In general, errors could be significantly reduced with Kalman fitting. The average bias reveals that while analogue reconstructions tend to overestimate negative extreme values and underestimate extremely high temperatures, post-processed temperature data show a median of approximately zero for all quantiles. The bias pattern of the ARM can be explained as for precipitation by a limited number of suitable analogues for extreme events. Kalman fitting solves this problem. Furthermore, the spread of bias values is within ±1 ∘C for 4 times the interquartile range. Post-processed temperature reconstructions are thus also accurate and precise for extreme temperatures. MSESS values are better for upper and lower quantiles and show worse results around the median. As days around the median temperature are closer to average climatology, this pattern has to be expected. Nonetheless, the MSESS of post-processed data is still within the area of natural variability (see Sect. ).

In Fig. , reconstructed precipitation and temperature are compared to station observations from Schaffhausen (left) and Grimsel Hospiz (right) over the full length of the respective series. Plotting reconstructed values against observations, we see a large spread of values for analogue reconstructions (grey) and for post-processed data (red). While for Schaffhausen, quantile-mapped precipitation exhibits a distribution closer to station observations in the Q–Q plot (centre) compared to ARM reconstructions, we can see a tendency towards overestimation in post-processed precipitation reconstructions for Grimsel Hospiz. The seasonal pattern (bottom) reveals the lowest differences in autumn, whereas the uncertainty is highest during summer, which is again in line with more frequent convective activity during the latter season as discussed before. The systematic underestimation of precipitation by the ARM is adjusted by post-processing for all seasons. The larger uncertainty in the Alpine region discussed before is also visible in station data for which the spread of bias values over all seasons (bottom) is larger for Grimsel Hospiz. This is also the case for the RhiresD gridded dataset used for reconstruction (see Sect. ). The differences between reconstructions and station observations might at least partly be explained by the high spatial variability of precipitation; thus, spatially coarser gridded data can differ considerably from local measurements. For reasons of higher spatial variability, a less perfect fit has to be expected compared to temperature reconstructions. These are closer to observed values compared to precipitation and show a smaller spread of deviations. Kalman-fitted reconstructions are even more precise; not only is the spread of values reduced, but the tilt in distribution could also largely be corrected. Seasonal patterns of ARM reconstructions show larger deviations during spring and autumn. These seasonal differences are eliminated by EnKF. Reconstructed temperature fields thus accurately reproduce local temperature measurements, even for remote locations. With precipitation data, however, one has to be more careful when generating station series at individual locations.

The winter of 1887/88 was one of the most severe avalanche winters during the last 150 years, boosting efforts in avalanche prevention in Switzerland . Intensive snowfall in February and March 1888 brought large snow masses to Switzerland, leading to 1094 disastrous avalanches, damaging 850 buildings, destroying over 1300 ha of forest, and burying 49 people under the snow . Documentary data from provide a detailed description of this winter and a comprehensive survey of avalanche activity gathered by cantonal forestry offices. From a historical perspective, it has recently been assessed by . However, quantitative data on the weather of this avalanche winter are restricted to sparse station observations so far. The gridded weather reconstructions presented in this paper can help to analyse the 1887/88 winter weather quantitatively, thus helping to better understand the weather patterns leading to such an event. For demonstration purposes, we performed some simple calculations of monthly averages, mean snow precipitation, and the zero-degree level that are summarised in Fig. .

Shown are post-processed precipitation reconstructions aggregated over 1 month (a–d) for winter 1887/88 and the recently published monthly precipitation reconstructions by MeteoSwiss (e–h). Both datasets reveal similar patterns of monthly precipitation, although regional differences occur. More deviations between the datasets can be observed in the amount of reconstructed precipitation. From both datasets, large precipitation sums can be determined in December 1887 on the northern flank of the Alps and in the Jura Mountains. January shows little precipitation, with the highest values in the north-eastern Alps. February and March show extreme precipitation values in Ticino and the Gotthard region, as well as in March over the remaining part of Switzerland.

Daily reconstructions allow for going more into detail. For example, we can calculate the development of the zero-degree level from gridded temperatures by taking the intercept of a linear regression between temperature and altitude (bottom of Fig. ). Altitude data used here were aggregated from the Shuttle Radar Topography Mission (SRTM) 90 digital elevation dataset to fit the resolution of reconstructions. Another value of interest is the intensity of snowfall precipitation. In Fig.  (bottom), the average snow precipitation per snowfall area is shown, calculated from post-processed reconstructions and assuming an estimated 1 ∘C threshold , below which precipitation falls as snow. Grey shaded areas depict periods of high avalanche activity as reported by .

Analogue reconstructions (red) and assimilated temperature data (blue) show similar values. The extreme precipitation event in December 1887 coincides with a high altitude of the zero-degree level, thus leaving a snow-covered area restricted to higher altitudes. This event was followed by several cold episodes and low precipitation in January. The first avalanche period was dominated by low temperatures and intensive precipitation. During the second avalanche period, the reconstructed zero-degree level rises to approximately 600 m a.s.l. with almost no snow precipitation. March shows two periods of high temperatures and intensive precipitation in the middle of the month and during the third avalanche period.

Reconstructed precipitation patterns and the development of temperature are in line with the findings from documentary data that report strong snowfall during December, a dry January, and intensive precipitation again during February and March, especially in the Southern Alps. While the first two avalanche periods are determined by low temperatures, Coaz describes a sharp rise in the zero-degree level to about 2000 m a.s.l. preceding the third period that led to a high number of wet avalanches, which can clearly be seen in the reconstructions. However, during the second period, reconstructions show relatively low snow precipitation values, contesting the high avalanche activity. Nonetheless, avalanches are not only triggered by intensive precipitation. For example, the intensive snowfall period in December 1887 and the first two in March 1888 are not accompanied by more frequent avalanches. To analyse this, other factors like temperature and wind as well as the composition of different snow layers also play an important role and have to be assessed. A closer look at these periods would probably reveal more about the processes that triggered or prevented avalanches.

From precipitation and temperature reconstructions, new insight on the avalanche winter of 1887/88 can already be gained with simple methods. Using more sophisticated snow models that also take into account evaporation and snowmelt, high-resolution daily spatial data of the snow cover could be established that may be able to further explain avalanche activity. This is but one example of what the new daily reconstructions of temperature and precipitation could be used for. Analogue reconstructions have already been applied as input to numerical models, such as crop modelling e.g. and hydrological modelling e.g., but the list could be extended. Many other phenomena, e.g. heatwaves and droughts, can be analysed spatially, and making use of the long time span changes in climate and extreme events over time could be investigated.