History.txt

1) Feature cleaning and transforming.
   Duplicate features, constant features.

2) Random Forest -> poor performance roc_auc ~0.52 (on test data)
   Problem: Unbalanced data set. Only ~3-4% of observations are unhappy

3) To deal with unbalance, undersampling of the majority class. Randomly sample as many happy observations as the total number of unhappy observations.
   Performance boost -> ~0.77 (on test data)
   Problem: overfitting.
   
4) To deal with overfitting, sample a fraction of the unhappy class (25%) and complete with as many happy observations. Train multiple random forests with thise method (60) and ensemble the results in a majority vote. Performance boost -> ~0.82 (on test data)

5) Used more forests (90) but with less trees (150). Instead of majority vote, we used mean of predicted probabilities. Performance boost -> ~.83

6) Use geometric mean improved a little bit.

7) Creating mean_balance, mean_mean_balance, sum_zeros, num_assets improved performance -> ~.84