# 250th Place Saved by Public Single Model

Acknowledgments to the authors of original solutions
There is no technical specification about my solution, just a decision-making story, because I didn't spend much time on this task personally and joined lately here, mostly to investigate approaches that people have shared.
For the first final submission I slightly adjusted some public ensembles to enter the public medal range, hoping it might be sustained on private, because people guessed the shake-up would be low to none (for the top teams it did eventually). For the second sub, I would do something similar within a different set of public models based on overall public score, but two things about the data stopped me from that decision. First, it appeared to be small train/test datasets; another one is a significant amount of missing data in the test. Frankly, even one of these two points alone could raise a red flag.
Reading the best approaches, I assumed that the authors provided quite effective methodologies to handle missing values, but still it's a big enough uncertainty to fully rely on this. So it led me to go by a trustworthy path. As most of us well know, it's about choosing the best single model for the second final submission. In the past I've seen many times how good it can work, especially with similar data situations, but I've preferred to risk making ensembles of at least two models (and often failed).
Fortunately, here we have some prominent persons, who shared secure and solid approaches to using them as is. At first I wanted to take
[@hideyukizushi](https://www.kaggle.com/hideyukizushi)
's solution due to his astonishing contribution at many previous competitions and seeing how it can step into medal range on the private LB. However,
[@wasupandceacar](https://www.kaggle.com/wasupandceacar)
's notebook had a better score, definitely a single, 5-fold model, supported by many Kagglers in comments, yukiZ included. Taking all risks low, I consciously and confidently approved it as a backup variant. Happy to see it sustained during a moderate shake to get a bronze, although the first blend didn't survive. Notably, if I've added this model to/instead of other models in the blend, the public score surprisingly dropped, but the private rose, which, of course, I couldn't know a priori, so I drop off the risk and simply chose a single model, as I said.
In sum, I would love to give some respect to all who share such dependable approaches to lead us through the darkness of uncertainty.