Limited predictability is one of the remarkable features of deterministic chaos and this feature may be quantized in terms of Lyapunov exponents. Accordingly, Lyapunov-exponent estimates may be expected to follow in a natural way from forecast algorithms. Exploring this idea, we propose a method estimating the largest Lyapunov exponent from a time series which uses the behavior of so-called simplex forecasts. The method considers the estimation of properties of the distribution of local simplex expansion coefficients. These are also used for the definition of error bars for the Lyapunov-exponent estimates and allows for selective forecasts with improved prediction accuracy. We demonstrate these concepts on standard test examples and three realistic applications to time series concerning largest Lyapunov-exponent estimation of an experimentally obtained hyperchaotic NMR signal, brain state differentiation, and stock-market prediction.