We present a general framework for forecast model selection using a meta-learning approach. It involves computing a range of features of the time series which are then used to select the forecasting model. The model selection process is carried out using a classification algorithm - we use the time series features as inputs, and the best forecasting algorithm as the output. The classification algorithm can be built in advance of the forecasting exercise (so it is an "offline" procedure). Then, when we have a new time series to forecast, we can quickly compute its features, use the pre-trained classification algorithm to identify the best forecasting model, and produce the required forecasts. Thus, the "online" part of our algorithm requires only feature computation, and the application of a single forecasting model, with no need to estimate large numbers of models within a class, or to carry out a computationally-intensive cross-validation procedure. The proposed framework is compared against several benchmarks and other commonly used forecasting methods. The results of our proposed framework outperform the other commonly used automatic algorithms for forecasting.