Update Self-experimentation-calibration.md

This commit is contained in:
Nuño Sempere 2019-05-06 16:08:14 +02:00 committed by GitHub
parent 04f8571979
commit 106a5bb3a3
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23

View File

@ -231,4 +231,4 @@ If I were to redo this experiment, I'd:
## 4. Conclusion
In conclusion, I am surprised that the dumb model beats the others most of the time, though I think this might be explained by the combination of not having that much data and of having a lot of variables: the random errors in my regression are large. I see that I am in general well calibrated (in the particular domain analyzed here) but with room for improvement when giving 1:5, 1:6, and 1:15 odds.
In conclusion, I am surprised that the dumb model beats the others most of the time, though I think this might be explained by the combination of not having that much data and of having a lot of variables: the random errors in my regression are large. I see that I am in general well calibrated (in the particular domain under analysis) but with room for improvement when giving 1:5, 1:6, and 1:15 odds.