This paper presents a generative approach to direction-of-change time series forecasting. Kernel methods are used to estimate densities for the distribution of positive and negative returns, and these distributions are then combined to produce probability estimates for return forecasts. An advantage of the technique is that it involves very few parameters compared to regression-based approaches, the only free parameters being those that control the shape of the windowing kernel. A special form is proposed for the kernel covariance matrix. This allows recent data more influence than less recent data in determining the densities, and is important in preventing overfitting. The technique is applied to predicting the direction of change on the Australian All Ordinaries Index over a 15 year out-of-sample period.