MIT OpenCourseWareSession 10: Student Projects with Nelson Repenning
MIT OpenCourseWare
Modeling purpose 0:00
Nelson Repenning asks what a good system dynamics model should give you. Students say it can surface hidden parts of a problem, test mental models, reveal feedback loops and leverage points, and calm a tense conversation by making the issue make sense. He agrees, but says the real test is the internal logic of the assumptions, not proof about the world.
Defining the puzzle 3:00
He urges you to be clear about why the model exists. Either you want a policy answer or you want to explain a puzzle, such as why theory predicts one thing and the data show another. He warns against saying only that you are modeling a system, since that leaves no boundary or end point for the work.
Publishing angles 16:31
Nelson says the book-reading idea is fine as a first model, but if you want to publish it you may need to tie it to a broader question a journal cares about. He suggests possible frames in literacy, strategy, or public interest, without forcing one literature. The student then explains a social media story: people may buy a book to signal taste or identity, while only skimmed summaries supply the needed content.
Keeping the model spare 21:31
Nelson pushes for a simpler setup. He says to keep two types, not many, and to start with the most basic demand story before adding more detail. He points to two usable hypotheses: cheaper ways to consume books, such as audiobooks and e-books, and a feedback loop where what friends read makes a book more appealing. He says the feedback version may fit better than a stock-and-flow model.
Model size and insight 33:31
Nelson Repenning presses the students to keep the model small enough to understand. He says you should not add every possible feedback loop just because it seems important. If you do, you end up with a model that fits data but teaches you little. He prefers a model that explains why it behaves as it does, and he says model size should grow with your own understanding.
Brand hype and decline 44:00
The discussion then turns to hype brands like Supreme, luxury goods, and language. A brand can become cool because an in-group uses it, then lose that status when an out-group adopts it. Price can rise as the brand becomes scarcer and more exclusive, but that same rise can push the original group away. Repenning says this creates a cycle of rise, spread, and collapse that a model might help you manage.
Managing hype and base 49:00
Harley-Davidson is used as a case where demand surged, but the company feared pricing out its original blue-collar base. The same problem shows up in luxury and street brands. The point is to sell into new demand without losing the core group that gave the brand its value. That balance is hard to keep.
Two groups and price 1:02:31
The model is then sharpened around two groups: cool kids and everybody else. The student explains hypebeasts as buyers who will pay anything, while in-group buyers care about signaling and can be shut out by higher resale prices. Nelson says the cleanest first step is to separate demand for each group, then let diffusion and coolness interact.
Brand strategy choices 1:05:31
Focus on one brand and even one product first. Build the core dynamic with two groups, then add the other feedbacks one by one. The key is how the cool kids and the non-cool kids shape demand for the same item. Pricing can help, but it can also kill the brand if popularity spreads too far.
Airline problem framing 1:13:01
The airline model needs a clearer puzzle. The student wants to track booking from launch to departure, with customers waiting when fares are high and a revenue manager changing prices as demand appears. Nelson pushes for the question underneath the structure: whether bad forecasts and weak feedback handling distort demand and leave the system suboptimized.
Forecasts and missing checks 1:24:30
The forecast is used every day by the revenue manager, and the manager’s fare changes feed back into future forecasts. What is missing is any check on forecast accuracy. That leaves a system that can drift into bad forecasts that nobody trusts. The answer is to name the dynamic hypothesis clearly and decide whether the problem really matters.
Model the feedbacks 1:29:30
You need to say what is going wrong and why. A model is useful because it can test whether the balancing loops are fast enough to correct the error you think is there, or whether a hidden reinforcing effect and path dependence keep the forecast fixed in the wrong place. The same logic applies when you think about distributions: one average stock may not capture the shape you care about.
Film production feedback 1:41:00
The student describes a model in which a tax credit outside Los Angeles pulls productions away over time. Without that incentive, Los Angeles keeps gaining local talent, experience, and infrastructure. Nelson Repenning asks for evidence that other places are building up the needed capacity, and the student points to Canada, the UK, and New Zealand. Repenning suggests using a multinomial logit form for attractiveness and starting with most production still in LA.
Heat mortality structure 1:49:01
Another student turns to heat and health, starting from the rise in mortality on very hot days despite warnings and awareness campaigns. The model begins by splitting people into young, adult, and older groups, then by income and housing quality, with mock numbers used only to test the structure. The group debates whether income belongs in the model, since housing quality and age seem to carry most of the risk, and agrees the time horizon should be decades because aging changes the picture.
Simpler age and housing 2:00:32
Use a simpler stock-flow chain for the people. Start with three age groups: 0 to 15, 15 to 65, and 65 plus. The main point is that the population is aging, which can raise mortality because older people are more vulnerable. Do the same for housing, with older buildings in separate age groups, so you can estimate how many vulnerable people live in outdated homes and test policy choices.
Delay in governance 2:09:01
The student’s model links shared resources to self-governance, where people punish overuse to protect the resource. The key hypothesis is that these controls work when delays are short, but can fail or even worsen the problem when delays are long. The example is fisheries, where licenses can be revoked. Repenning says the work should model the perception of scarcity and the punishment signal more carefully.
Observing use and blame 2:18:30
A model of water use needs a simple aquifer stock, visible thresholds, and some way to assign blame when use seems too high. Nelson says the hard part is not the stock itself but the different cases created by observability, delay, and uncertainty. He suggests starting with a specific setting and a simple bargaining model for the monthly negotiation.
Screening and ecosystem metrics 2:23:30
For lung cancer screening, the puzzle is low follow-up. Nelson says the issue may be dread, bias, or bad reputation from false positives, not a clear feedback loop. For the ecosystem project, he presses for a final measure. If collaboration matters, you need to name what improves, such as sustainability, GDP, or environmental impact.
Closing thanks 2:35:31
He thanks everyone for their patience and says they have run out of time, though they got through it. He notes the session was a little disorganized, but hopes there were still useful nuggets along the way. A student thanks him, and Nelson Repenning answers, "OK," before the exchange ends.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
