Well... it's finally over! This was an intense competition with strong competitors. I don't want to sound cliché, but I learned a tremendous amount. In fact, this was my first time tackling a time series problem, and I'm now completely captivated by it.

I'd like to invite everyone to share a concise summary of what worked for you and what didn't. The post-competition period is an excellent opportunity to learn from others' perspectives. So, let's share our insights, even if you weren't pleased with your final score!

What worked for me:

• Stratifield K fold with k = 10
• LGBM was my go-to model from day one
• The train data that I used was only when fault occur, so it was about ~90k rows

• Dropped from train data all de NE IDs with more than 5 NaNs
• I introduced seasonal features on data_rate using statsmodels
• Lag features! but 1 to 3. More than 3 the CV decrease.
• I dropped TA, access_success_rate, relation and fault duration.
• Feature engineering with cqi and mcs
• exponentially weighted on data_rate

df['data_rate_rolling'] = df['data_rate_lag'].ewm(span=4).mean()

• I applied an optimal threshold for all folds rather than a unique threshold for each fold

What didn't worked

1. Approaches like regression, interpolation, and imputation for handling NaNs in the test set. Despite my efforts, none of these methods worked well
2. Clustering
3. Trend features
4. Sin and Cos features
5. I struggled to find alternatives to gradient boosting trees, primarily because handling NaNs in the test set was a significant challenge. This limited me to GBDT models. As a result, my ensembles lacked sufficient diversity, which I believe was a major setback.
6. Feature engineering on NE ID: improve CV but decrease LB a lot.

(Bonus) What I couldn't make work:

data_rate has a clear seasonal frequency of 12 hours. Precisely at hour 01, there is a shift in the number of faults x number of faults where data_rate decrease. Every hour 01 this shift happens and my models was unable to capture it. Below, you can see the F1 score per fold. The lowest score corresponds to hour = 01. Unfortunately, I couldn't find a solution to this issue.