October 2014

Teaching Future Actuaries: Learn. Do. Automate Loss Modeling.

By Yvonne Chueh

"Successful actuaries will be those that optimize the mix of humans, computers, and algorithms." ~revised from MIT Technology Review "Automate or Perish."

Premise: Those who can't do, teach.
Conclusion: Those Who Don't Teach, Can do.

If our future actuaries will be doers, why not let them practice early in the Loss Modeling class?Being a teacher, I had wondered and trusted this logic: Those Who Don't Teach, Can (and must) Do. Thus, I made up my mind to engage my loss models class to do, using instructor's power in the fall of 2013. This article shares a little about how the way I taught Loss Models was influenced by our campus wide slogan: Learn. Do. Live.

This slogan first appeared on my university's website. Not a big surprise. It then populated throughout the campus, the windows of our HR office, across my department, the giant wall-papered mural of the Cascade Ridge Mountains with the valley view overlooking our entrance lobby, bulletins, and arrays of flying greeting flags which printed national and regional distinguished graduates and faculty to ask us "what did you do today?" Not until I noticed the mobile device charging kiosks whispering to bundles of distinct connecting adapters, and our national award winning student union structure resonating these three words while housing campus activities and excitements, did I decide to practice this slogan with my loss modelers—learn, do, and live, in a friendly manner.

Why do most of our students make their way here? To learn, of course. They come from Wenatchee, Yakima, Seattle, Federal Way, Auburn, Tacoma, Olympia, Spokane, from across the Washington state, and some across the country or overseas. The diversity of students is impressive, well respected, and celebrated in the university.

How well did I do my teaching job, besides making sure that my students really learned before and after getting all the boring or fancy formulas recited and practiced for the exams? So, the first message I decided to send out and planned to validate is "You Can Do (Implement) Almost Everything You Will Learn from the Class."

Loss Models class (covering Exam C syllabus) is a great course to start the H.I.P. (High Impact Practice). H.I.P. was created by a new teaching group on campus I joined after the Critical Thinking Fellow grant project mentioned in my MLC article in the last Expanding Horizons newsletter. Thanks to my earlier research and tools building that saved me precious time and quality of life, I can guide projects using the software package AMOOF[1] made by some enthusiastic and capable computer science majors as well as "do" H.I.P. for my instructor's role. Note that H.I.P and AMOOF turned out to be great resources for guiding the Loss Modeling and VEE Regression Model projects.

So we started the projects phase by phase, mimicking my neighborhood home builder Doug who was the son of a physics professor. His strategy was that if the houses don't sell quickly, he slowed down and got the issues identified and then resolved. I wasn't sure if my phase projects would sell quickly so a selling drill was crafted. That part was easy since no one risked to argue with me that actuaries have to apply knowledge to the real-world practice, and that knowledge transfer should be their strength and interest over their career. A complaint was "we only have this much time and we are taking other classes." So, the key to success lies in the student's individual strengths and interests. Provided matching well, the risk of complaint, rebellion, and failure can be mitigated or prevented. An extra-credit award system for exceptional work quality (beyond reasonable credit hours) played well for highly-motivated students who had concern on keeping high (or rescuing low) GPAs by meeting instructor's expectations. Throughout the course, a great deal of ongoing communications occurred concerning my expectations, time frames, trouble shooting, and ability to cope with test schedules and homework requirements. Although the various communications and issues were not always fun to deal with and somewhat stressful for both sides, the communication skills, willingness to try and help, enthusiasm, and patience cultivating for project success were priceless for both the student learning experience and instructor's role.

Synchronizing to the Loss Models textbook chapter content, at the beginning of fall, I posted quarterly projects in our online teaching platform, called Canvas that conveniently allowed me to add updates and resources in responding to questions, feedback, and progress. Canvas provided 24/7 online announcements, discussions, lecture postings, grade updates, video conferences, file uploading/downloading/sharing, project submission, etc. The teaching management platform added efficiency beyond our regular M,W,F class schedule, which was three hours per week for three10-week quarters. Such online communication systems facilitated projects like ours and fostered a learning community.

The Loss Modeling Projects consisted of:

Phase I: Build Probability Model Data Simulator. (Starting at the end of Fall quarter)

Phase II: Build Probability Model Fitter to Fit Simulated Data. (Winter quarter, 10 weeks)

Phase III: Build Probability Model Tester to Test the Model Fitting. (Spring quarter, 10 weeks)

During the process, students "should" think and design a research problem using real-world data interesting to them, and then use the Excel tools they build to model data and present to the Symposium of University Research Creative Expressions. (SOURCE) on May 20.

Doing projects sounds challenging because the Exam C syllabus is large for a three-credit class. In fact, it is workable if we weave the textbook content into the three phase projects to be developed with the right timing and computer lab exercises. To lay the groundwork and get it going, we spent the fall quarter on Chapter 3, 4, 5, 8, 20 (simulation), and then reduced the lecture hours to two hours per week to cover the textbook in the winter and spring quarters. This gave one hour weekly for the Phase Projects from January to June.

In the fall, we cruised through chapters to the policy limits and deductibles by first reviewing relevant probability topics covered in Exam P using worksheets and then connecting each topic to the loss models. The loss variables are essentially conditional or transformed/shifted/translated/censored random variables with corresponding/censored probability mass points. Doing a series of problems organized in worksheet format helped brush up calculus and probability skills and is a very effective approach to smooth out different student knowledge levels without taking much class time. We covered the Bayesian estimation and posterior distributions for the final exam to wrap up the fall quarter.

Using the analog of building a house, laying a solid foundation early in the fall is crucial. It was astonishing how a friend's house was built on a rising water level in crawl spaces above the house foundation. (Luckily, the water was clean and chilled in summer and went down below the ground weeks later.) Another house had manually shoveled in tons of rocks to elevate the crawl space level before passing the house inspection. Fortunately my house and my class were lucky to avoid such issues. The builder added additional layers of crushed rocks in my house building lot to elevate the foundation and block the underground irrigation water from running above the crawl space.

In the last week of the fall quarter, we did simulation lectures and then proceeded to the random number generator in the PC lab. After the theory was delivered and proved, students were ready to do project 1 in the lab and upgrade the user interface during winter break.

To earn extra credit, students explored the construction of a user friendly interface (UI) for running random data simulations. A class leader who took VBA programming class volunteered to investigate over the winter break. My VBA skills were not up to date, so his work was a great help to the class. Using our campus online teaching platform, the class became connected as a discussion group who would share files and thoughts outside the classroom. The first user interface by Christian Chemilewski was a success even though the complicated programming to simulate six different probability distributions assigned to each student with a user interface was a tremendous workload. Christian ran an online workshop via the Canvas Conference which was attended by and video recorded for those who didn't join in time. It turned out the entire class finished their own Phase I project and everyone earned the extra credits. The simulators with various UI design impressed and pleased me.

In the early winter quarter, all students signed off Phase I so we started Phase II. We had one hour each week in the PC lab for the Model Fitter tool and solving textbook exercises such as MME, PME, MLE, and LLE problems using Excel. The textbook and lab exercises together facilitated the building process of the Excel tool for Model Fitter. Extra credit for UI was offered and oral demonstration of UI and fitting results became a new requirement. It turned out LLE was too much to do so we dropped it. In this project the special functions such as the Gamma, Incomplete Gamma, Beta, and Incomplete Beta were frequently encountered and programmed into the Excel cell formulas and VBA language.

During spring break, students were busy with job interviews and prepared for the SOURCE presentation abstract submission. This year, unlike in the MLC class last year, only three projects in loss models generated public presentations. One was on the Seattle Seahawks after their Super Bowl championship, the second was on our campus dining service operations, and the third was the national math modeling competition involving highway traffic rule modeling. These projects used the probability models, estimations, and testing. The testing of model fitting is the Phase III Project with data analysis and fitting in the PC Lab using the textbook data and examples. Once the results were verified and corrected, students proceeded to the Model Tester.

The AMOOF3 software project was close to completion in spring, so the class used their Phase Projects to test AMOOF3 and investigated extra features such as the VaR and TVaR models to compare with the empirical data VaR and TVaR. The measures for goodness of fit extended from the classical KS, Anderson Darling, Chi-Square to Likelihood function values and VaR and TVaR (CTE). The AMOOF3 free software and Excel Phase Projects we collected from this class then can be shared with the actuarial students and practitioners today.

For further information or inquires about our Loss Model projects and AMOOF3 free software, please contact Yvonne Chueh at chueh@cwu.edu.

[1] AMOOF 3.0 Actuarial Model Outcome Optimal Fit 3.0 is partially sponsored by the Actuarial Foundation and CWU Computational Science Master Program. It is available for free at BitBucket.org/AMOOF3/amoof-3.0/downloads. A Prezi presentation is made available at Prezi.com/kxtvzv58evlp/amoof2-update-sept-1st.

Yvonne, Chueh, Ph.D., is professor and director of the actuarial science program at Central Washington University in Ellensburg, Wash. She can be reached at chueh@cwu.edu.