Assignment Task
Part
1. A. Begin to finalize your report by only including report/website quality output. By this, only include stylized output (do not use base R print). Include your code in your report using code_folding: hide. Avoid any package loading or other incidential inclusion of output in your report. This could mean using stargazer to show regressions, DT::datatable to show data.frames, and adding titles/labels to plots. Give all plots and tables appropriate captions. Write two to three sentence introductions for the different sections of your report. Your report should introduce property assessment in Detroit, Michigan (as defined in part 2A), your two prediction models (from 2B and 2C), and a conclusion.
B. Feature engineering. You have now created two base models and evaluation metrics from last week. Investigate creating at least two new predictors and analyze if they improve your model(s). Some possibilities:
- Neighborhood foreclosures/blight ticket
- Previous rates of assessment (part B only)
- Census variables (such as income, race)
- Neighborhood demolition of homes (look at how many class 401 properties w/ nonzero assessment by year)
- Sale price per square foot for neighborhood You may either add one metric to each model or two metrics to one of the two models.
C. Prediction. Create “out of sample” predictions for both models. By this, predict overassessment and assessment/ valuation for homes which did not sell for each model (2016 for B, 2019 for C). Note if you are having trouble with this step that you cannot use any information specific to the sale of a property for out-of-sample prediction. In other words, we use information on sale prices to determine the true value of homes but we do not know this information for homes which did not sell and if we want to make a prediction for these homes we cannot use sale price information as a predictor.
A helpful data framework for this section would be to create a dataset of all properties in 2016/2019. You would then label all the properties which sold (leaving un-sold properties unlabeled) and creating your testing and training data by filtering only to properties which were labeled. After training, you can then augment your full dataset of labeled/ unlabeled data to get out of sample predictions.
D. Model Explanation. Each model type has different tools for explainability and we will discuss this more in class. Undertake this initial work knowing that we will gain more techniques for this later on. Complete one of the following
Either (contextual explanation):
For the classification overassessment model, aggregate your predictions by census tract. Join in a census variable. Create a simple correlation plot and create a representation of the geographic variance in your predictions (this could be a leaflet map by census tract for example).
2. A. Bring together your previous submissions into one cohesive report. This report should offer a brief overview of the problem (assessment), general trends on properties, your model, why your model is better than other models, and any technical or ethical critique.
Your final submission will build upon your part 3 submission by ‘switching out’ the model you use and adding a conclusion.
B. Conclusion & Presentation
Write a four-paragraph conclusion to your file. Include information on your model type, its performance on your chosen objective function, any ethical or implementation issues (e.g. should Detroit use your model?).
In class on the 30th, everyone will give a brief presentation on their work. You may present your knitted Rmd file or pull some of your graphs into a slide deck. Your presentation should be at most five minutes. Broadly look to answer if your model should be implemented by discussing the information in your conclusion and assignment.
