Forecasts of climate rely on model projections, but derivation of sophisticated climate models from first principles is not currently feasible. Therefore, evaluating climate models with observations is essential. The development and improvement of global climate models is currently only based on comparison with and tuning to historical observations of climate (the instrumental record). Model simulations of the present climate are well-tuned and are in general agreement with each other. However, there is no clear relationship between model performance for present day and model behaviour for projections. Models show a range of sensitivities when predicting the future climate response to the emission of greenhouse gases. This indicates that the evaluation of models using observations of historical climate is insufficient. It is very difficult to reduce uncertainties on projections based on the instrumental period only and the use data from earlier periods is critical. A wide variety of different climate states are recorded in the geological record (spanning greenhouse to icehouse scenarios). The modelling of past climates, in combination with data from the geological record, provides a unique laboratory to evaluate the ability of models to forecast global change. While data is available from numerous intervals in Earth history, analysis is often constrained by the availability of material of the correct age and data collection is often very time consuming and expensive (e.g. for marine sediment cores). For this reasons, it is important that data on past climate and environments is utilised optimally and that challenges resulting from sparsity of the data as well as from temporal and spatial uncertainties are addressed in the best way possible. The earth system modelling and proxy reconstruction communities often have little contact with professional statisticians. Even in publications, ad-hoc methods are used instead of established statistical "best practice". If inappropriate statistical methods are used, inference about models and the earth system will be weakly supportable or plainly wrong. To avoid these problems and to realise the opportunity of improved earth system forecasting, sound statistical methods as advised by statisticians must be used. On the other hand, use of appropriate statistical methodology is often made difficult due to sparsity of data or lack of resources, and statisticians are not always aware of the resulting restrictions on the applicability of methods. Statisticians need to develop awareness of the restrictions and requirements caused by the sparsity of palaeoclimate data and the high complexity or climate models. The Past Earth Network will develop a shared, multi-disciplinary vision for addressing the challenges encompassed by the following four network themes. (1) Quantification of error and uncertainty of data: The uncertainties inherent in different forms of climate data must be well-understood. This is particularly challenging for palaeoclimate data, since uncertainties are often large and varied. (2) Quantification of uncertainty in complex models: The uncertainties in the output of the (complex and high-dimensional) models in use must be well-understood. (3) Methodologies which enable robust model-data comparison: Appropriate methods for model-data comparison must be used, taking into account the nature and sparsity of data. (4) Forecasting and future climate projections: This theme synthesizes the results from the first three themes in order to assess and ultimately improve the ability of climate models to forecast climate change. By addressing these four challenges, results produced by the Past Earth Network will help to better understand and reduce the uncertainties in climate forecasts and ultimately will contribute to the development of better climate forecasts.