FIR model trained with JSNN on Y2 data (DIH in Y3).

Eric's submission using basic ann with 4 hn, variables gender (3 variables), age (10 variables), claims per year (2 vars), placeofsvc, and specialty.

unscaled data, JSNN model trained on valid set, 196 features.

OLS 196f DIH>5 neglected.

This entry represents the approximate best public score you can get using a constant value of 0.209179 for each member. This constant value was derived by analyzing some characteristics of the evaluation metric. Please search for "Optimized Constant Value" in the forums for more details on this approach. As you can see from its rank, a constant value (even a good constant) won't get you very far in this competition.

Regression

This is the score you get if you accidentally get your columns in an incorrect order and have your 3rd column be the value of the "ClaimsTruncated" column instead of the expected "DaysInHospital" column. This typically happens if you add a row number as the first column.

grouped by age

Entry where all "DaysInHospital" are zero. This represents predicting that nobody will go to the hospital.

This is the score you get if you submit the SampleEntry.csv file. The sample entry consists of random values between 0 and 15.

Linear model

This entry represents a worst-case scenario of predicting that every member will spend the maximum period of 15 days for their DaysInHospital. If you look at data from previous years, you can tell that this approach will not perform well.