In this paper, I propose a flexible non-parametric method using Recurrent Neural Networks(RNN) to estimate the dynamic structure of most expectation formation models in macroeconomics. This approach does not rely on restrictive assumptions of functional forms and parametric methods but nests the standard approaches of empirical studies on expectation formation. Applying this approach to data on macroeconomic expectations from the Michigan Survey of Consumers(MSC) and a rich set of signals available to U.S. households, I find qualitatively new results: (1) agents' expectations about the future economic condition have asymmetric and non-linear responses to signals; (2) agents' attentions shift from signals about the current state to signals about future: they behave as if they were adaptive learners in ordinary periods and become forward-looking as the state of economy gets worse; (3) the content of signals on economic condition, rather than the volume of these signals, plays the most important role in creating the attention-shift. My method also allows me to apply the Double Machine Learning method to assess the statistical significance of these empirical findings. Finally, I show these stylized facts can be generated by a model with rational inattention, in which information endogenously becomes more valuable when economic status gets worse.