1. Case Study: FAA Air Traffic Data
JMP Explorers Series Seminar
February 6, 2013
Tom Donnelly, PhD
Systems Engineer & Co-insurrectionist
JMP Federal Government Team

2. Outline Discuss the Problem Key Steps Fit and compare models Summary
Get the data
Augment original data
Shape the data for visualization and modeling
Split data into Train, Validate/(Tune) and Test subsets
Fit and compare models
Partition vs. Neural vs. 2nd-Order Polynomial
With and Without Federal Holiday factor in the model
Summary
GOAL…
“Data > Information > Knowledge > Understanding > Action”

3. Problem and Situation
FAA already using a classification and regression tree tool to model air traffic to better understand staffing needs
While comparing results to this tool during evaluation, JMP’s answers were sometimes different. “Why?”
Initial FAA analyses done separately for each Air Route Traffic Control Center (ARTCC Facility) and for one year only – resulting in 23 files of 365 rows of data
Initial FAA focus was on using Weekday and Month as factors with which to model Total Traffic
By using seven years of data we can also learn about the effect and trends of Fiscal Year
Using 7 years of data for 21 ARTCCs (53,697 data rows) makes it easier to study the effect of Federal Holidays

4. Download Air Traffic Data From FAA Site

5. Download ARTCC Data From FAA Site

6. Jacksonville (ZJX) vs. Albuquerque (ZAB)?
ZAB

7. Include Modeling of 10 Federal Holidays
Symbol 1 K W M 4 L C V T X

8. Augment Original Data Use Tables Menu to:
Join Holiday information
Join Latitude/Longitude Information
Stack Air Traffic Type
Use Formula to extract and/or expand information from original m/d/y formatted Date column
Day of Week, Month, Year
Fiscal Year – use conditional & comparison formulas
Use Value Ordering to make better sense of ordinal data
Weekday, Month, Fiscal Year and Federal Holidays

9. Make the Visualization Dynamic
Use Data Filters and Distributions to select on the fly what to view
Use Column Switcher to compare model predictions in Graph Builder

10. Randomly Assign Data into 3 Groups – Train, Validate (Tune), and Test (60%/20%/20%)

11. Actual vs. Predicted Total Traffic for ALL Data in Test Subset

12. Actual vs. Predicted Total Traffic for NO Fed Holidays in Test Subset

13. Actual vs. Predicted Total Traffic for ALL Fed Holidays in Test Subset

14. Actual vs. Predicted Total Traffic for Thanksgiving Holiday in Test Subset

15. Actual vs. Predicted Total Traffic for Christmas Holiday in Test Subset

16. Actual vs. Predicted Total for October 2012 w & w/o “Sandy” Data for 29th, 30th & 31st

17. Actual vs. Predicted Total Traffic - Partition - for 10 Federal Holidays by Holdback Subset

18. Actual vs. Predicted Total Traffic – Neural - for 10 Federal Holidays by Holdback Subset

19. Actual vs. Predicted Total Traffic - 2nd Order - for 10 Federal Holidays by Holdback Subset

20. Fit Models Using Honest Assessment Method i. e
Fit Models Using Honest Assessment Method i.e. Train, Validate(Tune) and Test Subsets
Use JMP’s Random Indicator to assign subsets
In this case used a 60/20/20 split
Original work used just one year of data and fit each facility separately. This meant “outliers” especially Federal Holidays were only represented in one subset – making it impossible to take their behavior into account
Using seven years of data makes these Federal Holidays randomly fall into all three groups – making it possible to account for their behavior
Using seven years of data also makes it possible to model effect of Fiscal Year on Total Traffic.

21. Potential Enhancements to Analysis
Model non-Federal holidays
Model the holidays people really observe
Model SQRT(Total Traffic)
Normalize variance – a usual regression assumption
Prevent nonsensical predictions – negative traffic!
Model Type of Traffic separately or combined
Model impact of major storms

22. Prediction Profiler and Actual Values at 5 ARTCCs for Thanksgiving Day Test Data (Holdback = 2) for Fiscal Year 2010
ARTCC
Total Traffic
ZAB
2126
ZAN
683
ZAU
3699
ZBW
2879
ZDC
4265

23. Prediction Profiler Beside 3-D Response Surface

24. Bubble Plot

25. Q & A