3 of 8

Project Workflow

Created baseline model without any features

Optimized dataset

Added key features for modelling
Removed outliers and missing values

Define Problem & Gather Data

Exploratory Data Analysis (EDA)

Data

Preparation

Modeling

Online Research about flights and delays

Used Postgres & SQL to collect data

Started EDA

Created project timeline

Looked for trends within data

Completed exploratory_analysis notebook

Day 1

Day 2 - 3

Day 4 - 5

Day 6 - 7

Submission

Day 7

Models utilized:

ADA boost
Elastic Net
Naive Bayes
Poly Regression
Random Forests
Gradient Descent
SVR
XGBoost

Powerpoint presentation

Finalized submission file

4 of 8

Insights and Key Relationships

Target variable is normally distributed around 0

Strong relationship between dep_dealy and arr_delay

Minor relationship between taxi_out and arr_delay

Arr_delay per month is relatively consistent ever yearly

The longest arr_delays occurred late at night or very early in the morning

Other Insights:

34% of flights arrive late
34% of flights depart late
24-25% of flights that depart late arrive late (airlines make up lost time)
California, Texas, Florida, Illinois & Georgia covered majority of US air traffic (75%)

5 of 8

Feature Importance

Added date time features:

Month
Year
Day�

Other features:

Origin State
Origin City
Destination State
Destination City
Route
Depart hour
Distance

Feature Engineered:

By Route, origin city & month

Dep_delay (mean)
Arr_delay (mean)
Taxi_out_route (mean)
Taxi_in_route (mean)

Total Features Used:

23 features

6 of 8

Best Results - Test data

Model Type:	Basic linear regression	Optimized linear regression	ElasticNet Classifier	Naive Bayes	Poly Regression	ADA Boost Classifier	Random Forestst	SGD Regressor	SVR	XGBoost
R2 Score:	0.015	0.196	0.191	-0.695	0.222	-0.098	-0.645	-1.96	-0.0	0.174
MAE Score:	23.62	10.66	10.68	14.72	10.49	11.89	15.05	1.9*10^15	11.7	14.22

7 of 8

Biggest Challenges

Computation requirement

Difficult time running all the models with large enough data sets
Required to run some models overnight�

Feature Engineering Challenges

Mostly weak relationships with target variable prediction �

Issues with gathering data using postgres

Limitations with the server and/or postgres
Very difficult to use random with join when gathering data

8 of 8

Future of the project

Uses of the project

With the proper equipment, could expand the dataset, refine the features and teach the models with larger datasets to get a more accurate accuracy rate�
Could eventually sell or make the code open source�

Could be used by airlines to give real time updates to consumers on the status on their flight