Results

   - The winning probability was not always satisfactory, but it was not negligible.
   - I wondered whether the model provides meaningful predictions. I would have been satisfied if the model does a little better than the randomized blind draws.
   - So I compared it with the randomized draws. The result was interesting.
   - The winning probability triples the randomized draw in Lotto, and was 36% higher in Megamillion.
   - The model outperforms much better in higher winning (Even the Megamillion prediction shows the possibility of 5-digit matching), so it turned out to be MEANINGFUL!

LSTM VS Random (Lotto)
The green area indicates winning condition (Korean Lotto win starts from 3-digit match)

LSTM VS Random (Megamillion)
Megamillion wins on any case of matching.

   - Rarely, the probability had a big jump (We say the situation "Daebaak" in Korean)
   - Daebaak in Korean Lotto

   - Daebaak in Megamillion
   - The expectation was at least $280 on each prediction!!

Future work

   - I know LSTM is outdated model. I can try using one of the transformers or other approaches to deal with sequence of data.
   - I looking at more advanced ways to training and deployment to go beyond the structure of MLOps level 0. Leveraging one of awesome platforms, such as Azure, GCP, Sagemaker will be awesome to operate and monitor efficiently.