The Project
My class is an intro to stats so the project is mainly lit review/light implementation. We get to choose a problem in any topic and research the current statistical analysis techniques. The goal is relate the techniques back to the topics that we have learned so far.
It was actually difficult for me to narrow down a topic! It is probably backwards but I thought about the techniques I wanted to use and then determined how the data should look. Here are concepts I want to hit in this project:
- MLE
- order statistics
- hypothesis testing
- distribution types/model fitting?
- boostrap
My plan is to get some data, fit it to a model, analytically and numerically calculate the MLE. In the process, I could bootstrap a sampling distribution of a mean/\(Y_i\) order statistic. I could also set up a hypothesis test for the MLE.
In addition, I also wanted to use the EM and ranking/sorting algorithms that I wrote last semester!
Topic/ Data
The obvious first choice would be covid: I could build a predicting model with the top symptoms. Another idea was college admission data: California state universities no longer require SAT/ACT scores link. Again, I could build a model for the top indicators of student success. I found a really comprehensive dataset by College Score Card.
Summer is fire season in California so wouldn’t it be great if I could create a predictive model for wildfires??!? Um well just a quick lit review tells me that modeling fires is extremely complex. In fact, it seems like wildfires are chaotic? Here is a model for forest fires that follows self organized criticality: 3D Abelian sandpile simulation.
This article goes over fire model CalFire uses called ‘Wildfire Analyst’ which takes in realtime data to make predictions. The model was built around system of equations published in 1972 by Richard Rothermel. Wildfire Analyst is much more complex compared to Rothermel’s 1972 system the director Mark Finney of the US Forest Service’s Missoula Fire Lab says,
Like any model, Wildfire Analyst is only as good as the data that feeds it—and that data is only as good as our scientific understanding of the phenomenon in question. When it comes to the mechanics of wildland fire, that understanding is “medieval”
Doing more review, there are many parameters that affect wildfire risk and movement. MORE HERE
I do not hope to create a model for fire, only practice the concepts we are using in class to analyze fire data!
Here is a wildfire dataset hosted on Kaggle.
The Plan
I think it would be intersting to focus on the start of fire season. Maybe in order to make sure your property is fire safe, or plan parties? I plan on calculating the expected start of fire season with orders statistics. I could sample the data and create a $Y_{min}$ distribution with the earliest fire dates. Once I have this first-ordered distribution, I can ask questions such as
- do small/medium/large fires have the same start dates?
- is the start of fire season getting earlier?
- is there a noticeable location where fires begin?
More links on fire
Technology Review article on Fire Analyst