Data fitting is building a model that best explains the variability in your observed results. Prediction is using that model to estimate the outcome from a new observation. This is the essence behind leave one out (LOO) methods where a model is built using all but one of the data values, and then that model is used to predict the outcome for the data point that was left out. Failure of the model to correctly predict the outcome is an indication that the model is not very good. The method can be extended to leaving multiple observations out.
In econometrics, data -fitting is often called "estimation". Whereas a purely predictive model tends to start with an analytical equation first and then using "estimates" as parameter changes within that equation. Thereby allowing for a projection of likely states given a set of input parameters. Also data-fitting tends to place the emphasis on data, while prediction exercises tend to lead to a reformulation of the mathematical model.