Advisor Meetings

[sheacon]

1/10

• SAINT

• library funding

• asymmetric cost function

  • estimating price too high is different from price too low

• how much do i care about outliers?

• people

  • shotwell (teaching schedule TR 2:15-3:30, catch after)
  • spencer-smith (deep learning)

• tennessee validation

  • inspect 1000 obs sample
  • look at outliers/tails

[sheacon]

1/17

• discussed data collection progress: marketcheck, ohio, texas

[sheacon]

1/25

• data acquisition

  • texas invoice, confirming scope
  • ohio, ftp access / data sample
  • marketcheck, waiting…

• accre skilling up

  • github
  • slurm scripts
  • user data directory?
  • consider google cloud
  • aws/azure can be expensive

• modelling pipeline

  • data load and typing
  • one-hot encoding
  • transformations
    • inventory (think about normalizing, e.g. inventory per capita?)
    • progress metric: does this transformation increase the correlation with the target variable
  • normalize and scale predictive features
  • final: numerical matrix

• find domain knowledge

  • what drives used car prices?
  • find research
  • find expert
  • reach out to carmax pricing team
  • used car salesman
    • “what are people willing pay for?”
    • “what cars do you like having on the lot?”

[sheacon]

2/1

capstone meeting questions

• what historical transaction features are appropriate for my use case?

  • e.g. can “price a similar vehicle sold yesterday” be a feature? assumes real-time access to this kind of data in production use
  • e.g. current inventory

• what am i pricing? trade-in value, private party sale? (9% diff on my car)

• vehicle with listing price but no dmv price?

updates

• still waiting on texas and marketcheck
• reading other research papers
  • market dynamics (new cars)
  • mannheim
• priced car on online services

meeting notes

• FRED car cost data vs COX/Mannheim (govt data more neutral)
• delta (e.g. month-to-month) on prices and incentives is more important than absolute level)
• monthly aggregated features or rolling windows
• regions = m.s.a.
• trying pricing online as buyer

[sheacon]

2/8

• aggregate model or rolling series of models
• spark setup
• data status
  • tx friday
  • marketcheck
• track changes and amend license agreement and send along

[sheacon]

capstone advisor 2/15

• Career advice on move to Richmond with job uncertain vs. Nissan
• JSS for accre advice Friday
• Ask TAs from big data

friday 2/17 dsi office hour suggestions

• pull just vins and dates and dedupe. use index to pull full records
• think about diminishing returns (might not need all data)
• pandas read_csv engine = pyarrow

[sheacon]

capstone meeting 3/2

• 80% match rate!

• 16% mvr records missing price

  • salvage vehicle vs non-salvage? (drop salvage, not my project goal)
  • other brands that should be drops

• is it ok to cap the data at some price? not interested in extremes?

  • windsor-izing or truncating the distribution (big data tends toward truncating)

• use listing price as feature?

  • dealer tool (drop list price)
    • how much to buy it at auction, how much to spend refurbishing
  • consumer (include list price)

• impute missing data or just drop since i have so much?

  • as long as it doesn’t introduce bias
  • packages that help with this (MICE?)

• metric for improving the worst predictions from base models to final

• RMSE, MAE, MAPE

trim feature engineering

• target encoding on make/model level, then trim clustering provides delta
  • how many make/models are too many to keep as categories instead of target encoding?
• big picture: dimensionality reduction of options to replace trim
• in effect, “trim” is a bundle of options. we are going to do the clustering on bundle of options
• drop trim, redundant with standard/optional features

  check whether hvf_features are standardized across model

• e.g. moonroof vs sunroof

  pca for binary features?

• how do the options bundle down to one dimension

[sheacon]

capstone meeting 3/9

• capstone requirements met?
  • **emphasize business case
    • accurate pricing models important
    • pricing used cars is hard at scale (look at carmax cover letter)
    • price many vehicles in auction with VIN decoder
  • why is this approach better than others?
    • transaction price from dmv vs. listing price
    • private listing data, better features
• overpayement = loss, underpayment = low inventory
  • absolute error -> quantiles
• mean (not absolute) error as a overall measure of balance of underbidding vs overbidding
• GO AHEAD AND OUTLINE PAPER DRAFT to make sure i have the analyses i need