/used-car-price-prediction

Primary LanguageJupyter Notebook

Predcit a used car price

Notebook

Please check detail test results

Goal

Understand what factors make a car more or less expesive

  • State (region), manufacturer and odometer affect price of the used car

Provide clear recommendations to car dealship - as to what consumers value in a used car

  • New business model buy used cars lower price in Delware and sell the cars in Vermont to maximize sales revenue

Business Ideas

1. Leverage Machin Learnig model to argument pricing for used cars

We have a machine leaerning model which can argument sales represnetitive pricing work. There is some margin which could be +/- 7050

2. Pricing is strongly correlated to

2.1. State

Buy cars from Delaware and sell in Vermont will give us maximize sales revenue.

For example, Ford White SUV good condition and clean titled car can be sold 35,597 in Vermont whereas the same car can be aquired 4,875 in MD.

The gap between the two state is 30,722.

2.2. Car manufacturer and odometer

Factors that are affecting price of used car are that

  • manufacturer - Ferrari, Tesla or Aston-martin, the car brand contribute used car price because retail price (MSRP) is expensive
  • odometer - Very strong negative correlrated to the price of used car

Exporatory Data Analysis

Quality issues

  • Columns; manufacturer, condition, Cylinders, fuel, title_status, transmission, drive, size, type, paint_color, state are fine
  • Issue with unclear model names
  • Price range of car is beteen 0 and 3,736,928,711 (3B?)
    • Data quality need to be reviewed
    • 360,000 for Ferrari looks reasonable, but price for Toyota 3.7B isn't reasonable
    • 0 for a car isn't reasonable
  • Around 4 % of data have missing value

Insights

  • id and VIN are unique random number

  • 42 different car brand (= manufacturer)

  • Pricing factors

    • title_status - if the status is not clean then the price can be lower
      • For example, the 25 years old car with clean title more valuable than the 22 years old with salage titled car

    • year - postive correlation with price
    • odometer - negative correlation with price

    • Price of similar cars are different in differnt state

    • good condition car better for our business than excellent condition car

Questions

  • What is currency of car price? Is it all USD or mix of USD and other
  • Is this list price, traded in/out or sell in/out price?

Feature Engineering

Categorical features

Drop columns

  • id, VIN, model due to its uniquness and randomness thoes will be dropped during the trainng

Select data set in range

  • Only the car price between 500.00 and 400,000.00 will be used for training

Train model

Experiment with three algorithms
Also changed data set to avoid over/under fit
Use the parameter, `Positive = True` to prevent negative price prediction
- Linear Regression
- Ridge
- Lasso
Used verious technique to find the best model and hyper parameter of it

Result of model trainings

Model Data set Hyper Params MSE MAE
Linear Regression Defaults 500 < car price < 120000 121928253.36 8198.16
Ridge Defaults 500 < car price < 120000 114423615.06 7606.56
Lasso Defaults 500 < car price < 120000 98000493.63 7008.86
Ridge Grid Search CV 500 < car price < 120000 108301123.53 7435.62
Linear Regressopn Defaults 500 < car price < 120000 without Sate, Region 126978774.65 8191.15
Ridge Defaults 500 < car price < 120000 without Sate, Region 126978774.65 8191.15
Lasso Defaults 500 < car price < 120000 without Sate, Region 56446418.64 5037.57
Lasso Deafults 500 < car price < 120000 Ford in CA 172136863.09 9562.40
Lasso Deafults 500 < car price < 120000 Ford only 135039985.93 8415.82
Lasso Deafults 500 < car price < 100000 84699147.65 7050.70

Hyperparameter tunning result

Tune Model Data set Hyper Params MSE MAE
Before Lasso Deafults 500 < car price < 100000 84699147.65 7050.70
After Lasso alpha=0.1, fit_intercept=False 500 < car price < 100000 51987298.13 4857.292936419671

Model

Prediction result

Summary

We have a model that can predict a car between 500 to 100000
Note that it may not be able to provide estimate price for a luxary car such as Ferrari

Feature importance

regions are very important factors for pricing

Following shows how much regions are impoart (the higher number the more important)

region_detroit metro 0.002 +/- 0.000
region_delaware 0.002 +/- 0.000
region_houston 0.004 +/- 0.000
region_huntsville / decatur 0.004 +/- 0.000
region_anchorage / mat-su 0.026 +/- 0.001
region_albuquerque 0.035 +/- 0.001
region_eugene 0.045 +/- 0.001
region_humboldt county 0.061 +/- 0.001
region_kalispell 0.061 +/- 0.002
region_asheville 0.063 +/- 0.001
region_SF bay area 0.068 +/- 0.001
region_akron / canton 0.096 +/- 0.001
region_austin 0.106 +/- 0.001
region_charlotte 0.434 +/- 0.001