Case Study: Prediction of Heavy Machinery Auction Prices

June 25, 2018

Objective

This Case Study aims at applying Linear Regression algorithms to predict the auction price of heavy machinery equipment.

Findings

• The best predictors for the auction price of a machine are (i) the type of machine, (ii) some measure of weight or size (depending on the machine type), and (iii) the machine's enclosure.
• Even though the age of a machine is strongly correlated with the price, after adding controls the size of the association is low (i.e. small estimated coefficient).
• Statistically-significant predictors related to the market conditions of the transaction, such as the state and the auctioneer, have a small association with price (i.e. small estimated coefficients).

Data

The data is sourced from auction result postings and includes information on usage and equipment configurations. The dataset is sparse and consists of 53 features and >400,000 rows. Many of the features contained mixed datatypes and/or presented a high share of missing values.

Intuitively, important features to determine machinery price (target) should capture:

• function or type of machinery
• depreciation (proxy: age)
• size and/or power
• energy consumption
• comfort of use (proxy: enclosure)

Moreover, a machine's price could be affected by market conditions captured by the dataset, such as:

• state in which the auction takes place
• who the auctioneer is

Data Cleaning & Feature Engineering

• Log(price)
• Fill in NaNs
• Age
• Creation of dummies from categorical variables

Modeling

3 algortihms were tried for the same model specification:

• Linear Regression
• Ridge
• Lasso

The model evaluation was based on Root Mean Squared Log Error.

Results

Model	Error
OLS	0.4083
Ridge (L2)	0.4082
Lasso (L1)	0.4083

Pros

• Low errors
• Intuitive interpretation of estimates (betas)

Cons

• High complexity, i.e. too many features

Team

• Ini
• Cosma
• Olga