MIT OpenCourseWareSession 12: Student Projects (cont’d)
MIT OpenCourseWare
Project focus 0:00
Hazhir Rahmandad opens the last class by saying the session will stay open-ended. The group can work on personal projects, ask for feedback on modeling work, and revisit formulation issues if needed. He also wants time for reflection on what has been learned and what could be changed in the class.
Learning next steps 1:35
He then answers a question about what comes after the course. He points to practice through hard business dynamics problems, old PhD general exam questions, and looking at other people’s models. He also notes that the next version of the class may focus on connecting models with data, since that is important for real applications.
Connecting models and data 7:00
When Maria asks about her project, he says serious empirical work needs data. He explains that good estimation depends on model complexity, process noise, and how much of the dynamics cannot be observed. He recommends matching the model’s level of detail to the data, and including the data-generating process, including delays and measurement bias.
Green startup structure 20:33
The model should keep the main links clear: green startups, brown firms, and the energy use tied to each. The green side matters because it can lower other firms’ energy use and total emissions. The speaker suggests keeping economic capital simple for now, treating it as a parameter unless it becomes central. Subscripts can help later if you want more sectors, but the first model should stay easy to read and simulate.
Build it step by step 33:00
The key mechanism is that green startups change brown firms’ energy intensity over time, and that change can feed back into startup growth. The speaker advises not to split every case, like selling technology versus using it internally, unless the model needs it. Start with exogenous birth and death rates, simulate early, and then add one layer at a time. A contribution comes from making the green-brown connection quantitative and showing a nontrivial dynamic.
Writing to the host language 41:03
A model only counts when it speaks in the field’s own words. You have to read what others in that field have done, then write your paper in their language and show what your model adds. Using system dynamics is not the point by itself. The point is whether it helps that audience understand the problem better.
Reading model feedback 45:10
The project review then turns to a book-reading model with high-frequency and low-frequency readers. The current setup treats the shift as exogenous, while social influence changes desired reading and buying rates. The reviewer asks for a clearer formulation of social influence, because the anchor, the saturation, and the nonlinear shape are tangled together. He also questions whether the story is about buying books from social signals or about adjusting reading time to others’ habits.
Booking curves and assumptions 1:01:03
A booking model needs clear anchors. The low fare and high fare demand are treated as separate groups, with a protected level in between. The point is to make the assumptions visible and check that they are defensible. The same idea applies to the forecast: you need a reference booking curve, not just a raw total, so you can judge whether current sales are ahead or behind.
Protecting fare inventory 1:05:02
The low-fare rate is driven by capacity first. If low-fare seats remain, bookings flow there until that pool is full, and only then do residual shoppers move to the high-fare pool. For multiple fare classes, the same cascade repeats from the cheapest fare upward. The protected level should then adjust over time, but the current rule is still too static and needs a better feedback process.
Pricing and coding choices 1:23:31
The discussion starts with the effect of taking prestige out of the model. Without that effect, the choice is between high prices with low quantities or low prices with high quantities, and the black dot marks the most profitable point. The student then varies the share of customers who are price sensitive, from 100 percent down to 90 percent, and the optimum shifts quickly. Hazhir Rahmandad explains that a flat payoff surface can make small changes move the best point, and that a small prestige-focused group may also trigger a tipping dynamic. He then walks through ways to run Vensim models in Python, export data, and build linear or piecewise linear price and quantity policies.
Screening model issues 1:37:41
The second model tracks low and high smoking history, nodule growth from none to symptomatic, and treatment paths for screened and unscreened cases. Rahmandad checks who can be screened, what happens after detection, and whether treatment follows detection or symptoms. He points out that recovery and death rates should be specified for symptomatic nodules before treatment, that exit rates are needed, and that assuming only lung cancer causes death creates problems. He also questions the use of screening participation in the growth path, since nodules keep growing while people wait to be screened. The student says that participation is still exogenous for now, but will later depend on false results and perceived risk.
Screening stays separate 1:44:33
Screening should not change the flow to the right. If more people are screened, fewer remain exposed to that movement, but the rate itself stays at its baseline. The same idea applies to treatment paths with passing and recovering. If there are only two exits, use one average time and one dying fraction, not extra time constants that make the model harder to read.
Discrete thresholds and data 1:49:30
Changing a threshold in place does not work when the stock definition itself is fixed. A cleaner choice is to use more layers and move people between them with discrete rules. For screening, the flow should depend on tumor growth, not on participation. Data may be sparse, but deaths, screening counts, and detected cases can still be tied to the stocks and their data-generating process.
Habits and pace 2:04:00
People say the course changed how they work. One person now watches units and model versions more carefully. Another says the pace helped them get used to building a model quickly instead of delaying for a week. Hazhir Rahmandad agrees that pressure can help you start, even if some struggle is unavoidable when you are learning.
Need more basics 2:07:31
The group asks for more work on very simple formulas, smooth delays, and examples of a model that is built well from the start. Rahmandad says students come in with different backgrounds, and that the class assumed some basics that were never actually taught. He also says he should prepare complete model examples and more focus on dynamics, not just structure.
Model choice and framing 2:22:32
The discussion turns to the balance between covering many models and giving enough feedback without losing people. The first model gave time to settle in and see the big picture, and later the class moved to the harder question of what to represent at all. That conceptual step, not the mechanics, is the real challenge.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
