---
title: 'Used Cars: Homework 01'
author: 'Chicago Booth ML Team'
output: pdf_document
fontsize: 12
geometry: margin=0.6in
---


# Load Libraries

```{r}
library(data.table)
library(ggplot2)
library(kknn)
```


# Data Import

```{r}
# download data and read data into data.table format
used_cars <- fread(
  'https://raw.githubusercontent.com/ChicagoBoothML/DATA___UsedCars/master/UsedCars_small.csv')
# count number of samples
nb_samples <- nrow(used_cars)
# sort data set by increasing mileage
setkey(used_cars, mileage)
used_cars
```


# Plot $x = mileage$ vs. $y = price$

```{r fig.width=8, fig.heigth=6}
plot_used_cars_data <- function(used_cars_data,
                                title='Used Cars: price vs. mileage',
                                plot_predicted=TRUE) {
  g <- ggplot(used_cars_data) +
    geom_point(aes(x=mileage, y=price, color='actual'), size=1) +
    ggtitle(title) +
    xlab('milage') + ylab('price')
  
  if (plot_predicted) {
    g <- g +
      geom_line(aes(x=mileage, y=predicted_price, color='predicted'), size=0.6) +
      scale_colour_manual(name='price',
                          values=c(actual='blue', predicted='darkorange'))
  } else {
    g <- g +
      scale_colour_manual(name='price',
                          values=c(actual='blue'))
  }
  
  g <- g +
    theme(plot.title=element_text(face='bold', size=24),
        axis.title=element_text(face='italic', size=18))
  
  g
}

plot_used_cars_data(used_cars, plot_predicted=FALSE)
```

The relationship looks downward-sloping, with cars with more mileage having lower prices. This is expected.


# Linear Regression

```{r fig.width=8, fig.heigth=6}
linear_model <- lm(price ~ mileage, data=used_cars)
used_cars[, predicted_price := predict(linear_model, newdata=used_cars)]

plot_used_cars_data(used_cars, title='Linear Model')
```


# KNN Regressions for various $k$

```{r}
k <- 5
knn_model <- kknn(price ~ mileage,
                  train=used_cars, test=used_cars[ , .(mileage)],
                  k=k, kernel='rectangular')
used_cars[, predicted_price := knn_model$fitted.values]

plot_used_cars_data(used_cars, title=paste('KNN Model with k =', k))
```

With $k =$ `r k`, the predictor suffers from **high-variance**, being overly-sensitive to small changes in the _mileage_ variable.


```{r}
k <- 300
knn_model <- kknn(price ~ mileage,
                  train=used_cars, test=used_cars[ , .(mileage)],
                  k=k, kernel='rectangular')
used_cars[, predicted_price := knn_model$fitted.values]

plot_used_cars_data(used_cars, title=paste('KNN Model with k =', k))
```

With $k =$ `r k`, the predictor seems to be a bit too simple and does not do well in the highest and lowest extremes. This is a **high-bias** predictor.


```{r}
k <- 30
knn_model <- kknn(price ~ mileage,
                  train=used_cars, test=used_cars[ , .(mileage)],
                  k=k, kernel='rectangular')
used_cars[, predicted_price := knn_model$fitted.values]

plot_used_cars_data(used_cars, title=paste('KNN Model with k =', k))
```

$k =$ `r k` seems to produce a reasonable, **low-bias**, **low-variance** predictor.


# Predicted Price of Used Car with 100,000 Miles

```{r}
test_case <- data.table(mileage=1e5)

knn_model <- kknn(price ~ mileage,
                  train=used_cars, test=test_case,
                  k=k, kernel='rectangular')
```

The Linear Model predicts price of $**`r formatC(predict(linear_model, newdata=test_case), format='f', digits=2, big.mark=',')`**.

The KNN Model with $k =$ `r k` predicts price of $**`r formatC(knn_model$fitted.values, format='f', digits=2, big.mark=',')`**.