Flight delay insurance marketplace powered by delay prediction risk score
delayStats.ipynb
This notebook explores a dataset of 2016 flights, filtering for those that originated from Boston, and visualizing the features and overall statistics of delay times. Given the distribution of delay times, the 4 "delay blocks" were chosen for the final risk score calculation:
- Less than 30 min
- Greater than 30 min and Less than or equal to 60 min
- Greater than 60 min and Less than or equal 120 min
- Greater than 120 min
Other statistics were calculated such as number of total and delayed flights by month, carrier, and destination airport, to understand importance of various features.
riskCalc.ipynb
This notebook was an early test to compare a regression model of predicting the exact flight delay time or a classification model to group the delay into one of the 4 "delay blocks." While some delay time prediction results from a Random Forest Regressor looked promising, the error rate was significant enough that ultimately the decision was made to use the delay blocks in the final model.
delayPred.ipynb
This notebook focuses on testing different models to predict a delay block for Boston origin flights. KFold method (4 splits) was used to partition training, validiation, and testing sets. Logistic Regression, Bagging Classifier, Gradient Boosting, AdaBoost, Multilayer Perceptron (MLP), K Nearest Neighbors, and Random Forest Classifiers were tested, with Gradient Boosting Classifer producing the best result, with a score of 0.862. However, in further testing, the Random Forest model proved to be more robust and was chosen as the final model.
