Quantifying the risk of global warming exceeding critical targets such as 2.0 ◦ C requires reliable projections of uncertainty as well as best estimates of Global Mean Surface Temperature (GMST). However, uncertainty bands on GMST projections are often calculated heuristically and have several potential shortcomings. In particular, the uncertainty bands shown in IPCC plume projections of GMST are based on the distribution of GMST anomalies from climate model runs and so are strongly determined by model characteristics with little influence from observations of the real-world. Physically motivated time-series approaches are proposed based on fitting energy balance models (EBMs) to climate model outputs and observations in order to constrain future projections. It is shown that EBMs fitted to one forcing scenario will not produce reliable projections when different forcing scenarios are applied. The errors in the EBM projections can be interpreted as arising due to a discrepancy in the effective forcing felt by the model. A simple time-series approach to correcting the projections is proposed based on learning the evolution of the forcing discrepancy so that it can be projected into the future. This approach gives reliable projections of GMST when tested in a perfect model setting. When applied to observations this leads to projected warming of 2.2 ◦ C (1.7 ◦ C to 2.9 ◦ C) in 2100 compared to pre-industrial conditions, 0.4 ◦ C lower than a comparable IPCC anomaly estimate. The probability of staying below the critical 2.0 ◦ C warming target in 2100 more than doubles to 0.28 compared to only 0.11 from a comparably IPCC estimate.