Reference Model Based Learning in Expectation Formation: Experimental Evidence

How do people form expectations about future prices in financial markets? One of the dominant learning rules that explains the forecasting behavior is the Adaptive Expectation Rule (ADA), which suggests that people adjust their predictions by adapting to the most recent prediction error at a constant weight. However, this rule also implies that they will continually learn and adapt until the prediction error is zero, which contradicts recent experimental evidence showing that people usually stop learning long before reaching zero prediction error. A more recent learning rule, Reference Model Based Learning (RMBL), extends and generalizes ADA, hypothesizing that: i) People apply ADA but dynamically adjust the adaptive coefficient with regards to the auto-correlation of the prediction error in the most recent two periods; ii) Meanwhile, they also utilize a satisficing rule so that people would only adjust their adaptive coefficient when the prediction error is higher than their anticipation. This paper utilizes a rich set of experimental data with observations of 41,490 predictions from 801 subjects from the Learning-to-Forecast Experiments (LtFEs), i.e., the experiment that has been used to study expectation formation. Our results concludes that RMBL fits better than ADA in all the experiments.