Isn it circular reasoning to develop the AI models on past data and then to back test the models on the same data, essentially making predictions about what the AI program already knows?
Yes, but we don’t do that (although we have seen other people fraudulently do this). When we developed the AI models, we were very careful to avoid back testing on data that the model had been trained on. In order to properly back test a model, one must only train the model on data that was available up to the point in time that a prediction is made. For instance, when back testing and making a prediction for July of 1992, the model must be trained on historical data up to, but no later than, June of 1992. Then, for the next month’s predictions for August of 1992, the model must only be trained on data up till July of 1992. This is very time consuming, but it is the right way to do it.
