Quote:
Originally Posted by ToothSayer
We can certainly say something here. I'm just wondering what. I don't accept that a data point that is modeled to be 1/1,000,000,000,000 to hit, hitting, means we can say nothing about the veracity of the model because we have only one data point.
Again, you must define better your context. Things may change dramatically.
Consider this scenario. There are two hypotheses, H1 and H2. We know one of them must be true. H1 predicts event E to be 1:10^9, while H2 predicts E to be 1:10^13. Event E actually happens. Just use Bayes' rule and see that our belief on H1 actually increases a lot thanks to E happening, even if it was so unlikely.*
In that scenario a single data point is hugely informative. On other scenarios it might be insignificant.
Some considerations.
- Consider always alternative hypotheses. An unlikely event by itself doesn't imply the hypothesis being unlikely
unless other hypotheses assign the event an higher likelihood.
- A model is built on a set of hypotheses. If predictions are off, you need to test single assumptions and see if you can improve the model.
*That scenario, even if much less extreme, is what indeed happened for election models. All the most serious models gave Clinton favourite. However, Nate Silver's model gave Trump the higher chances and Trump winning has cemented Silver's model as likely the best (even if people not understanding probs might consider the model a failure).