Introducing bounded rationality in a standard consumption-based asset pricing model with time separable preferences strongly improves empirical performance. Learning causes momentum and mean reversion of returns and thereby excess volatility, persistence of price-dividend ratios, long-horizon return predictability and a risk premium, as in the habit model of Campbell and Cochrane (1999), but for lower risk aversion. This is obtained, even though our learning scheme introduces just one free parameter and we only consider learning schemes that imply small deviations from full rationality. The findings are robust to the learning rule used and other model features. What is key is that agents forecast future stock prices using past information on prices.