In my tinkering outside of work, I've embarked (with a previous colleague/friend) on developing an automated ML based stock trading system that handles buying, selling, and retraining, while understanding that this is the least unique project to embark on, as everyone wants to make money and this is the stereotypical side project.
The first day we got it truly live, my friend pinged me saying that it ran the whole day without issue, but only made a single trade, because the predictions rarely exceeded our thresholds for buying, which were set to traditionally aggressive threshold when attempting to maximize precision.
The lack of strong predictions in either direction is an indicator that we were onto something. Stocks are notoriously hard to predict. If this wasn't the case, everyone would be making millions by running a basic program on his or her computer. The lack of strong predictions means that our system recognized the insufficiency in our inputs.
This indicated a launching point for us, the beginning of the research phase. The system ran, and it told us that we had work to do.