Portfolio risk is in an important way driven by 'abnormal' returns emanating from heavy tailed distributed asset returns. The theory of regular variation and extreme values provides a model for this feature of financial data. We first review this theory and subsequently study the problem of portfolio diversification in particular. We show that if the portfolio asset return distributions are regulary varying at infinity, then Feller's convolution theorem implies that the portfolio diversification is more effective than if the underlying distribution would be normal. This is illustrated by a simulation study and an application to S&P stock returns.