We conduct a laboratory experiment in a fully-fledged macroeconomic model where participants receive information about future government spending shocks and are tasked with repeatedly forecasting output over a given horizon. By eliciting several-period-head predictions, we investigate forecast reaction to news and revision. The lab forecasts are consistent with stylized facts on reaction to news established in the survey literature. We find that subjects steadily learn the magnitude of the effect of the shocks on output, albeit not to full extent. We further find little support for fully backward-looking expectations. We rationalize the experimental data in the context of a Bayesian updating model, which provides a better description of the behaviors in longer-horizon environments and among more attentive and experienced subjects.