Regression:

Prediction
Professor Brett Saraniti
Module 3, 2022
What single factor is the biggest
determinant in the price of a home?
 To use DATA to build a MODEL that predicts the market value of a home in
Wilmette, IL.

 Wilmette is the small suburb just North of Evanston where some Kellogg faculty
live. Also a few derivative traders, big shot attorneys, various tycoons & titans of
industry…

 Most famous recent resident 

Every house sold in Wilmette: April 17, 2018 - April 17, 2019
Excel Spreadsheet - Wilmette.xlsx
 Is there a relationship between the size of the home and its sale price?
 Is this relationship LINEAR

 How precise is the relationship –

 Does it FIT exactly on a straight line?
 How far from “the line” do the data points tend to be?

 Are there other observable/testable traits about the relationship between

Price and Squarefootage?
 What can we do to improve our model?
https://www.redfin.com/IL/Winnetka/56-Indian-Hill-Rd-60093/home/13782109
 Let’s build a model to Predict the Price of a home in Wilmette

 We believe Price can be expressed as an “almost linear” function of squarefeet:

Price = β0 + β1 squarefeet + 𝜖𝜖

Noise or
Error

Linear
 The simple linear regression model

Y = β0 + β1 X + 𝜖𝜖
 Y is called the dependent variable
 X is called the independent variable
 We use Excel (or other software) to estimate β0 and β1 .
 We’ll make a few assumptions about 𝜖𝜖

 Specifically, we’ll assume 𝜖𝜖 ~ Normal (0,𝜎𝜎2)

Our model assumes that the actual price for any home of a
certain squarefootage (sqft), will be “near” the line --

 For a house with 2968 sqft, we assume the price can be

represented by the following model:

Price = β0 + β1 (2968) + 𝜖𝜖
 The error term represents the value of the home that
cannot be explained just by its size…
 Excel and a million other software platforms use a process called the “least
squares method” to estimate the coefficients in the model.

 The method dates back to 1795 and a guy named Gauss

 Some details 
 Excel will do all the hard work for us. It will produce the equation for “the
best” line. In our case, it looks like this:

Price = 14,699 + 273.54 squarefeet

 The coefficient β0 is estimated to be 14,699.

 β0 represents the Y-intercept or constant term in the line.
 Always remember: this is just an estimate of β0. We won’t ever KNOW the
true value of β0
 The value reported by Excel as the intercept is referred to as bo
14,699
 Excel gave us this linear equation:

Price = 14,699 + 273.54 sqft

 The coefficient β1 is estimated to be 273.54

 β1 represents the slope or rate of change term between sqft and price.
 Always remember: this is just an estimate of β1. We won’t ever KNOW the
true value of β1
 The value reported by Excel as the coefficient of X is referred to as b1
Rise

Run
= 273.54
Rise

Run
 The Regression MODEL –
Y = β0 + β1 X + 𝜖𝜖

 The Regression LINE –

ŷ = b 0 + b 1x
 Click Here

Then Here
Input the Column
Info for Y and X.

Tick Boxes for

Residuals and
Line Fit Plots
Wilmette Real Estate

Price = 14699 + 273.54 sqft

 b1 estimates β1
 We can compute 95% Confidence Intervals for β1

We are 95%
Confident that β1 is
in the range
[249.99, 297.09]
𝑋𝑋� ± 𝑡𝑡 𝑠𝑠� 𝑋𝑋� ± 𝑡𝑡 (𝑆𝑆𝑆𝑆𝑆𝑆 𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸) 𝑏𝑏1 ± 𝑡𝑡 𝑆𝑆𝑆𝑆𝑆𝑆 𝐸𝐸𝐸𝐸𝐸𝐸
𝑛𝑛

Confidence Interval for the unknown

Confidence
population mean, μ
Interval for the
unknown model
parameter, β1
 Coefficients Standard Error t Stat P-value Lower 95% Upper 95%
Intercept 14699.43561 37993.96821 0.386888664 0.699283 -60257.72241 89656.59364
SQUARE FEET 273.5377151 11.93702873 22.91505879 6.29E-56 249.9875099 297.0879203

𝑏𝑏1 ± 𝑡𝑡 𝑆𝑆𝑆𝑆𝑆𝑆. 𝐸𝐸𝐸𝐸𝐸𝐸 273.54 ± 1.973 11.94

t is computed using
=T.INV.2T(.05,185) 273.54 ± 23.55

[249.99 , 297.09]
 The degrees of freedom used
to compute t-values for
confidence intervals in
regression equals: n – k – 1
where k is the number of X’s

n=

=T.INV.2T(.05,185) = 1.973
 We ALWAYS run this particular test:
𝐻𝐻0 : β1 = 0
𝐻𝐻𝑎𝑎 : β1 ≠ 0
 Recall our Model: Y = β0 + β1 X + 𝜖𝜖
 Suppose we find that the p-value for this test is quite small.
 Then we say …
 the coefficient β1 is statistically significant or
 X has a statistically significant relationship with Y
 Coefficients Standard Error t Stat P-value Lower 95% Upper 95%
Intercept 14699.43561 37993.96821 0.386888664 0.699283 -60257.72241 89656.59364
SQUARE FEET 273.5377151 11.93702873 22.91505879 6.29E-56 249.9875099 297.0879203

𝐻𝐻0 : β1 = 0
𝐻𝐻𝑎𝑎 : β1 ≠ 0

𝑏𝑏1 − 𝛽𝛽1 273.54 − 0

𝑡𝑡 = = = 22.915
𝑠𝑠𝑏𝑏1 11.94
p = T. DIST. 2T(22.915,185) = 6.29 x 10-56 ≈ 0.000000000
  Statistical Significance ≠ Economic Significance

“We can be very very very sure that something matters a teeny
teeny tiny bit”
-Professor Brett Saraniti

𝑠𝑠
𝑠𝑠𝑏𝑏1 =
𝑠𝑠𝑥𝑥 𝑛𝑛 − 1
 We want to say something about Professor Schummer’s house
 We know sqft = 2968. How do we use our regression to estimate the selling price?

 Price = 14699 + 273.54 sqft 

 Price = 14699 + 273.54 (2968) = $826,559

 Next Question: How GOOD is our prediction?

Y = β0 + β1 X + 𝜖𝜖 The Standard Error of the
regression estimates the σ for the
𝜖𝜖 ~ Normal (0,𝜎𝜎2) error term.
  We want to add a sense of precision to our Point Estimate: $826,559
 A Confidence Interval is a range or interval estimate for the MEAN
 We don’t want to be 95% confident about the average price of houses with 2968
sqft of living space.
 We want to be 95% confident about the price of ONE PARTICULAR house: Jim’s

 The tool for this job is called a Prediction Interval

Confidence Interval for the mean Prediction Interval for one instance
 (b0 + b1x*) +/- 1.96(se-regression)

 $826,559 +/- 1.96 ∙ $209,344

 We are 95% confident that Jimmy’s home price will be in this
range:
$826,559 +/- $410,314

 Why is this “approximate” ?

 We should use t, not 1.96
 There is also error in the coefficients which we are ignoring (for now.)
  (b0 + b1x*) +/- (t-value)(st. err. of an individual prediction at x*)

1 (𝑥𝑥 ∗ −𝑥𝑥)̅ 2
st. err. of an individual prediction at x*= 𝑠𝑠𝑠𝑠𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟 + +1
𝑛𝑛 (𝑛𝑛−1)𝑠𝑠𝑥𝑥2

Q: What happens when n

A1: This standard error converges to the 𝑠𝑠𝑠𝑠𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟
is really really really large?

A2: The t-value converges to 1.96

 The value of the “± part” depends
1 (𝑥𝑥 ∗ − 𝑥𝑥)̅ 2
on x*, the particular value of X we are 𝑡𝑡 ∗ 𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟 + 2 +1
𝑛𝑛 (𝑛𝑛 − 1)𝑠𝑠𝑥𝑥
using to make our prediction of Y.

 When x* is closer to 𝑥𝑥̅ the interval will be tighter

 When x* is further from 𝑥𝑥̅ the interval will be wider

 In our example, x* is 2968 sqft.

$826,559 +/- $410,314

This is disappointing…
Can Zillow do better?
 86% are within 20% of sale price

 Our Model -- 95% are within 50% of sale price

 They WIN!! Do we surrender?

 Analysts often talk about how well (or not so well) their models “fit the data”
 The best known (but not the only) statistic that is used to measure regression
performance is the R2
 Some analysts put enormous weight on R2
 A former adjunct professor at Kellogg once offered “Rules of Thumb” along the lines
of “if your R2 is below 0.30, your work is trash; if it above .7, your work is
excellent”
 Most seasoned analysts, and virtually everyone who truly understands statistics,
usually make no more than a passing reference to R2
 I would like to explain why R2 is usually not an important statistic and, in
the process, dissuade you from paying it much attention
 Wilmette home prices vary.

 Some of the variance in Price can be explained by the square footage.

“That house sold for a high Price because it was enormous”
“That other house sold for a small Price because it was tiny”

 R2 measures the percentage of the variance of Y that can be explained

by the value of X.
 Notation

𝑦𝑦�𝑖𝑖
𝑒𝑒𝑖𝑖
𝑦𝑦𝑖𝑖

𝑦𝑦�

𝑒𝑒𝑖𝑖 = (𝑦𝑦𝑖𝑖 −𝑦𝑦�𝑖𝑖 )

𝑥𝑥𝑖𝑖
 𝑒𝑒𝑖𝑖 = 𝑦𝑦𝑖𝑖 − 𝑦𝑦�𝑖𝑖 
 𝑦𝑦𝑖𝑖 = 𝑦𝑦�𝑖𝑖 + 𝑒𝑒𝑖𝑖 
 𝑉𝑉𝑉𝑉𝑉𝑉(𝑦𝑦𝑖𝑖 ) = 𝑉𝑉𝑉𝑉𝑉𝑉 𝑦𝑦�𝑖𝑖 + 𝑉𝑉𝑉𝑉𝑉𝑉(𝑒𝑒𝑖𝑖 )

𝑛𝑛 2 𝑛𝑛 2 𝑛𝑛
∑
 𝑖𝑖=1 𝑦𝑦𝑖𝑖 − 𝑦𝑦 ∑
� = 𝑖𝑖=1 𝑦𝑦�𝑖𝑖 − 𝑦𝑦� + 𝑖𝑖=1 𝑦𝑦𝑖𝑖 − 𝑦𝑦�𝑖𝑖 2
∑
Total SS = Regression SS + Residual SS
 We define R2 as follows:

2
𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅 𝑆𝑆𝑆𝑆
𝑅𝑅 =
𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 𝑆𝑆𝑆𝑆
R2 is a measure of relative precision. It tells us the fraction of the total
variance that can be explained by our linear model.

Total SS = Regression SS + Residual SS

Explained
Unexplained
by Model
 Regression Statistics
Multiple R 0.859926325
R Square 0.739473284
Adjusted R Square 0.738065032
Standard Error 209343.68
Observations 187

ANOVA
df SS MS F Significance F
Regression 1 2.30124E+13 2.30124E+13 525.0999 6.29254E-56
Residual 185 8.10758E+12 43824776351
Total 186 3.112E+13

𝑹𝑹𝑹𝑹𝑹𝑹𝑹𝑹𝑹𝑹𝑹𝑹𝑹𝑹𝑹𝑹𝑹𝑹𝑹𝑹 𝑺𝑺𝑺𝑺 2.30𝑒𝑒+13

R2 = = = 0.7395
𝑻𝑻𝑻𝑻𝑻𝑻𝑻𝑻𝑻𝑻 𝑺𝑺𝑺𝑺 3.11𝑒𝑒+13
 R2 is a relative measure of accuracy.
 If R2 is very high, it means a very specific thing:

The regression equation does a much better job

predicting y as compared to the sample mean, 𝑦𝑦.
�

 If the variance of Y is extremely high, then even an R2 close to ONE does not
guarantee useful prediction intervals.
 R2 is one way (but not the only way) to measure regression performance
 R2 is the fraction of the variation in y explained by the x-variables in your
regression model
 A regression with a high R2 can be uninformative
 A regression with a low R2 can be highly informative
 The specific questions motivating your analysis usually suggest better, more
relevant tools than R2
Practice:
Prediction
Professor Brett Saraniti
 Movie studios often budget tens or even hundreds of millions of dollars
towards marketing a film…

They can reallocate funds or other assets up to the very last minute

 The critical point for any movie is the opening weekend. After that there is
little more that the studios can do to improve the financial performance of a
movie

 Predicting opening weekend gross is worth a great deal to the studios

 Until the 2000’s the studios used mall surveys to forecast revenues…

 These all sucked. Really bad.

 Any surveys from more than a week or two away from opening were total
noise.
 HSX sold forecasts to Hollywood studios for a six-figure fee.

 The studios want to receive their forecasts on Thursday morning which is the
last time they have any ability to add or subtract support for the film.

 HSX can take the Wednesday closing price to utilize the most recent
information when making its forecasts.

 So can we…
 Grab the data from canvas: HSX.xlsx
 Grab the homework instructions on sasinware