ReadMe.md

NKSR Hackathon approaches

https://www.kaggle.com/competitions/nk-iv-prediction/ i didn't include any data files here

• You should start by explore.ipynb

• what worked best is matching.cpp which finds the nearest neighbour (based on rmse) like knn and imputes it's non nan values into - current row. Since the IV values does not change much in a second, it is supposed to find a temporal order. You need to prepare the data into csv with only call put iv columns without any header for this code to work.

• then cubic_spline.ipynb for remaining ipynb (if during matching MAX_ITER is reached or rmse < THRESHOLD)

• finally applying savgol filter or pca smoothing made predictions better (bot in final_day/pca_smoothing.ipynb)

• final_day/lol.csv also works good, it uses xgb with iteratively improving the completely filled data (initially guessed through matching and cubic spline)

• I tried many other things like svi or pchip instead of cubic spline but they did not fit correctly in a many rows.

• I also tried using call put parity, training xgb on train data's each column to predict previous column, it works fine but not as good as cubic spline.