In this paper we show that Genetic Algorithms (GA) offer a simple explanation for human behavior in the Learning-to-Forecast experiment (Heemeijer et al., 2009). In this experiment two treatments are studied: under positive (negative) feedback realized price depends positively (negatively) upon individual forecasts. A common finding is that people converge quickly to the fundamental price under negative feedback, but the price oscillates under positive feedback. We propose to model agents as boundedly rational optimizers who learn how to maximize the forecasting precision of their prediction heuristics through GA. They use a simple linear (first-order) prediction rule which depends on past prices and predictions, the average price so far and the observed trend. The coefficients of the rule are not fixed; instead the agents are endowed with a small number of different coefficient sets and update their specifications with Genetic Algorithms optimization procedure. In each period they choose one set of coefficients based on its hypothetical performance from the previous period. As in the experiments, the agents are independent and cannot share information. In line with the experiments, we study simulations in which the initial predictions are sampled from the empirical distribution and hence the GA agents update their rules for 49 periods. Our results closely resemble the experimental data both at the individual and the aggregate level for both types of feedback. GA agents converge to the fundamental price under the negative feedback, but exhibit oscillating redictions in the case of positive feedback. Our model is also in line with previous analytical models and offers a unified explanation for their partial insights. Regardless of the feedback, heterogeneity in terms of forecasting rules persists. Moreover, asynchronous switching emerges endogenously. Agents learn to experiment with the rule choice only once per number of periods.