In recent decades, much comparative testing has been conducted to determine which forecasting methods are more effective under given conditions. This evidence-based approach leads to conclusions that differ substantially from current practice. This paper summarizes the primary findings on what to do – and what not to do. When quantitative data are scarce, impose structure by using expert surveys, intentions surveys, judgmental bootstrapping, prediction markets, structured analogies, and simulated interaction. When quantitative data are abundant, use extrapolation, quantitative analogies, rule-based forecasting, and causal methods. Among causal methods, use econometrics when prior knowledge is strong, data are reliable, and few variables are important. When there are many important variables and extensive knowledge, use index models. Use structured methods to incorporate prior knowledge from experiments and experts’ domain knowledge as inputs to causal forecasts. Combine forecasts from different forecasters and methods. Avoid methods that are complex, that have not been validated, and that ignore domain knowledge; these include intuition, unstructured meetings, game theory, focus groups, neural networks, stepwise regression, and data mining.