ACCT 420: Advanced linear regression


Session 4


Dr. Richard M. Crowley

Front matter

Learning objectives

  • Theory:
    • Further understand:
      • Statistics
      • Causation
      • Data
      • Time
  • Application:
    • Predicting revenue quarterly and weekly
  • Methodology:
    • Univariate
    • Linear regression (OLS)
    • Visualization

Datacamp

  • Explore on your own
  • No specific required class this week

Based on your feedback…

  • To help with replicating slides, each week I will release:
    1. A code file that can directly replicate everything in the slides
    2. The data files used, where allowable.
      • I may occasionally use proprietary data that I cannot distribute as is – those will not be distributed
  • To help with coding
    1. I have released a practice on mutate and ggplot
    2. We will go back to having in class R practices when new concepts are included
  • To help with statistics
    1. We will go over some statistics foundations today

Assignments for this course

  • Based on feedback received today, I may host extra office hours on Wednesday

Quick survey: rmc.link/420hw1

Statistics Foundations

Frequentist statistics

A specific test is one of an infinite number of replications

  • The “correct” answer should occur most frequently, i.e., with a high probability
  • Focus on true vs false
  • Treat unknowns as fixed constants to figure out
    • Not random quantities
  • Where it’s used
    • Classical statistics methods
      • Like OLS

Bayesian statistics

Focus on distributions and beliefs

  • Prior distribution – what is believed before the experiment
  • Posterior distribution: an updated belief of the distribution due to the experiment
  • Derive distributions of parameters
  • Where it’s used:
    • Many machine learning methods
      • Bayesian updating acts as the learning
    • Bayesian statistics

Frequentist vs Bayesian methods

This is why we use more than 1 data point

Frequentist perspective: Repeat the test

detector <- function() {
  dice <- sample(1:6, size=2, replace=TRUE)
  if (sum(dice) == 12) {
    "exploded"
  } else {
    "still there"
  }
}

experiment <- replicate(1000,detector())
# p value
paste("p-value: ",
      sum(experiment == "still there") / 1000,
      "-- Reject H_A that sun exploded")
## [1] "p-value:  0.962 -- Reject H_A that sun exploded"

Frequentist: The sun didn’t explode

Bayes persepctive: Bayes rule

P(A|B) = \frac{P(B|A) P(A)}{P(B)}

  • A: The sun exploded
  • B: The detector said it exploded
  • P(A): Really, really small. Say, ~0.
  • P(B): \frac{1}{6}\times\frac{1}{6} = \frac{1}{36}
  • P(B|A): \frac{35}{36}

P(A|B) = \frac{P(B|A) P(A)}{P(B)} = \frac{\frac{35}{36}\times\sim 0}{\frac{1}{36}} = 35\times \sim 0 \approx 0

Bayesian: The sun didn’t explode

What analytics typically relies on

  • Regression approaches
    • Most often done in a frequentist manner
    • Can be done in a Bayesian manner as well
  • Artificial Intelligence
    • Often frequentist
    • Sometimes neither – “It just works”
  • Machine learning
    • Sometimes Bayesian, sometime frequentist
    • We’ll see both

We will use both to some extent – for our purposes, we will not debate the merits of either school of thought, but use tools derived from both

Confusion from frequentist approaches

  • Possible contradictions:
    • F test says the model is good yet nothing is statistically significant
    • Individual p-values are good yet the model isn’t
    • One measure says the model is good yet another doesn’t

There are many ways to measure a model, each with their own merits. They don’t always agree, and it’s on us to pick a reasonable measure.

Frequentist approaches to things

Hypotheses

  • H_0: The status quo is correct
    • Your proposed model doesn’t work
  • H_A: The model you are proposing works
  • Frequentist statistics can never directly support H_0!
    • Only can fail to find support for H_A
  • Even if our p-value is 1, we can’t say that the results prove the null hypothesis!

OLS terminology

  • y: The output in our model
  • \hat{y}: The estimated output in our model
  • x_i: An input in our model
  • \hat{x}_i: An estimated input in our model
  • \hat{~}: Something estimated
  • \alpha: A constant, the expected value of y when all x_i are 0
  • \beta_i: A coefficient on an input to our model
  • \varepsilon: The error term
    • This is also the residual from the regression
      • What’s left if you take actual y minus the model prediction

Regression

  • Regression (like OLS) has the following assumptions
    1. The data is generated following some model
      • E.g., a linear model
        • Next week, a logistic model
    2. The data conforms to some statistical properties as required by the test
    3. The model coefficients are something to precisely determine
      • I.e., the coefficients are constants
    4. p-values provide a measure of the chance of an error in a particular aspect of the model
      • For instance, the p-value on \beta_1 in y=\alpha + \beta_1 x_1 + \varepsilon essentially gives the probability that the sign of \beta_1 is wrong

OLS Statistical properties

y = \alpha + \beta_1 x_1 + \beta_2 x_2 + \ldots + \varepsilon

\hat{y} = \alpha + \beta_1 \hat{x}_1 + \beta_2 \hat{x}_2 + \ldots + \hat{\varepsilon}

  1. There should be a linear relationship between y and each x_i
    • I.e., y is [approximated by] a constant multiple of each x_i
    • Otherwise we shouldn’t use a linear regression
  2. Each \hat{x}_i is normally distributed
    • Not so important with larger data sets, but a good to adhere to
  3. Each observation is independent
    • We’ll violate this one for the sake of causality
  4. Homoskedasticity: Variance in errors is constant
    • This is important
  5. Not too much multicollinearity
    • Each \hat{x}_i should be relatively independent from the others
    • Some is OK

Practical implications

Models designed under a frequentist approach can only answer the question of “does this matter?”

  • Is this a problem?
  • Often, this is enough

Linear model implementation

What exactly is a linear model?

  • Anything OLS is linear
  • Many transformations can be recast to linear
    • Ex.: log(y) = \alpha + \beta_1 x_1 + \beta_2 x_2 + \beta_3 {x_1}^2 + \beta_4 x_1 \cdot x_2
      • This is the same as y' = \alpha + \beta_1 x_1 + \beta_2 x_2 + \beta_3 x_3 + \beta_4 x_4 where:
        • y' = log(y)
        • x_3 = {x_1}^2
        • x_4 = x_1 \cdot x_2

Linear models are very flexible

Mental model of OLS: 1 input

Simple OLS measures a simple linear relationship between an input and an output

  • E.g.: Our first regression last week: Revenue on assets

Mental model of OLS: Multiple inputs

OLS measures simple linear relationships between a set of inputs and one output

  • E.g.: Our main models last week: Future revenue regressed on multiple accounting and macro variables

Other linear models: IV Regression (2SLS)

IV/2SLS models linear relationships where the effect of some x_i on y may be confounded by outside factors.

  • E.g.: Modeling the effect of management pay duration (like bond duration) on firms’ choice to issue earnings forecasts
    • Instrument with CEO tenure (Cheng, Cho, and Kim 2015)

Other linear models: SUR

SUR models systems with related error terms

  • E.g.: Modeling both revenue and earnings simultaneously

Other linear models: 3SLS

3SLS models systems of equations with related outputs

  • E.g.: Modeling both stock return, volatility, and volume simultaneously

Other linear models: SEM

SEM can model abstract and multi-level relationships

  • E.g.: Showing that organizational commitment leads to higher job satisfaction, not the other way around (Poznanski and Bline 1999)

Modeling choices: Model selection

Pick what fits your problem!

  • For forecasting a quantity
    • Usually some sort of linear model regressed using OLS
    • The other model types mentioned are great for simultaneous forecasting of multiple outputs
  • For forecasting a binary outcome
    • Usually logit or a related model (we’ll start this next week)
  • For forensics:
    • Usually logit or a related model

Modeling choices: Variable selection

  • The options:
    1. Use your own knowledge to select variables
    2. Use a selection model to automate it

Own knowledge

  • Build a model based on your knowledge of the problem and situation
  • This is generally better
    • The result should be more interpretable
    • For prediction, you should know relationships better than most algorithms

Modeling choices: Automated selection

  • Traditional methods include:
    • Forward selection: Start with nothing and add variables with the most contribution to Adj R^2 until it stops going up
    • Backward selection: Start with all inputs and remove variables with the worst (negative) contribution to Adj R^2 until it stops going up
    • Stepwise selection: Like forward selection, but drops non-significant predictors
  • Newer methods:
    • Lasso and Elastic Net based models
      • Optimize with high penalties for complexity (i.e., # of inputs)
      • We will discuss these in week 6

The overfitting problem

Or: Why do we like simpler models so much?

  • Overfitting happens when a model fits in-sample data too well
    • To the point where it also models any idiosyncrasies or errors in the data
    • This harms prediction performance
      • Directly harming our forecasts

An overfitted model works really well on its own data, and quite poorly on new data

Statistical tests and Interpretation

Coefficients

  • In OLS: \beta_i
  • A change in x_i by 1 leads to a change in y by \beta_i
  • Essentially, the slope between x and y
  • The blue line in the chart is the regression line for \hat{Revenue} = \alpha + \beta_i \hat{Assets} for retail firms since 1960

P-values

  • p-values tell us the probability that an individual result is due to random chance

“The P value is defined as the probability under the assumption of no effect or no difference (null hypothesis), of obtaining a result equal to or more extreme than what was actually observed.”"
– Dahiru 2008

  • These are very useful, particularly for a frequentist approach
  • First used in the 1700s, but popularized by Ronald Fisher in the 1920s and 1930s

P-values: Rule of thumb

  • If p<0.05 and the coefficient matches our mental model, we can consider this as supporting our model
    • If p<0.05 but the coefficient is opposite, then it is suggesting a problem with our model
    • If p>0.10, it is rejecting the alternative hypothesis
  • If 0.05 < p < 0.10 it depends…
    • For a small dataset or a complex problem, we can use 0.10 as a cutoff
    • For a huge dataset or a simple problem, we should use 0.05

One vs two tailed tests

  • Best practice:
    • Use a two tailed test
  • Second best practice:
    • If you use a 1-tailed test, use a p-value cutoff of 0.025 or 0.05
      • This is equivalent to the best practice, just roundabout
  • Common but semi-inappropriate:
    • Use a two tailed test with cutoffs of 0.05 or 0.10 because your hypothesis is directional

R^2

  • R^2 = Explained variation / Total variation
    • Variation = difference in the observed output variable from its own mean
  • A high R^2 indicates that the model fits the data very well
  • A low R^2 indicates that the model is missing much of the variation in the output
  • R^2 is technically a biased estimator
  • Adjusted R^2 downweights R^2 and makes it unbiased
    • R^2_{Adj} = P R^2 + 1 - P
      • Where P=\frac{n-1}{n-p-1}
      • n is the number of observations
      • p is the number of inputs in the model

Causality

What is causality?

A \rightarrow B

  • Causality is A causing B
    • This means more than A and B are correlated
  • I.e., If A changes, B changes. But B changing doesn’t mean A changed
    • Unless B is 100% driven by A
  • Very difficult to determine, particularly for events that happen [almost] simultaneously
  • Examples of correlations that aren’t causation

Time and causality

A \rightarrow B or A \leftarrow B?

A_t \rightarrow B_{t+1}

  • If there is a separation in time, it’s easier to say A caused B
    • Observe A, then see if B changes after
  • Conveniently, we have this structure when forecasting
    • Recall last week’s model:

Revenue_{t+1} = Revenue_t + \ldots

Time and causality break down

A_t \rightarrow B_{t+1}? \quad OR \quad C \rightarrow A_t and C \rightarrow B_{t+1}?

  • The above illustrates the Correlated omitted variable problem
    • A doesn’t cause B… Instead, some other force C causes both
    • Bane of social scientists everywhere
  • This is less important for predictive analytics, as we care more about performance, but…
    • It can complicate interpreting your results
    • Figuring out C can help improve you model’s predictions
      • So find C!

Today’s application: Quarterly retail revenue

The question

How can we predict quarterly revenue for retail companies, leveraging our knowledge of such companies

  • In aggregate
  • By Store
  • By department

More specifically…

  • Consider time dimensions
    • What matters:
      • Last quarter?
      • Last year?
      • Other timeframes?
    • Cyclicality

Time and OLS

Time matters a lot for retail

How to capture time effects?

  • Autoregression
    • Regress y_t on earlier value(s) of itself
      • Last quarter, last year, etc.
  • Controlling for time directly in the model
    • Essentially the same as fixed effects last week

Quarterly revenue prediction

The data

  • From quarterly reports
  • Two sets of firms:
    • US “Hypermarkets & Super Centers” (GICS: 30101040)
    • US “Multiline Retail” (GICS: 255030)
  • Data from Compustat - Capital IQ > North America - Daily > Fundamentals Quarterly

Formalization

  1. Question
    • How can we predict quarterly revenue for large retail companies?
  2. Hypothesis (just the alternative ones)
    1. Current quarter revenue helps predict next quarter revenue
    2. 3 quarters ago revenue helps predict next quarter revenue (Year-over-year)
    3. Different quarters exhibit different patterns (seasonality)
    4. A long-run autoregressive model helps predict next quarter revenue
  3. Prediction
    • Use OLS for all the above – t-tests for coefficients
    • Hold out sample: 2015-2017

Variable generation

library(tidyverse)  # As always
library(plotly)  # interactive graphs
library(lubridate)  # import some sensible date functions

# Generate quarter over quarter growth "revtq_gr"
df <- df %>% group_by(gvkey) %>% mutate(revtq_gr=revtq / lag(revtq) - 1) %>% ungroup()

# Generate year-over-year growth "revtq_yoy"
df <- df %>% group_by(gvkey) %>% mutate(revtq_yoy=revtq / lag(revtq, 4) - 1) %>% ungroup()

# Generate first difference "revtq_d"
df <- df %>% group_by(gvkey) %>% mutate(revtq_d=revtq - lag(revtq)) %>% ungroup()

# Generate a proper date
# Date was YYMMDDs10: YYYY/MM/DD, can be converted from text to date easily
df$date <- as.Date(df$datadate)  # Built in to R
  • Use mutate for variables using lags
  • as.Date() can take a date formatted as “YYYY/MM/DD” and convert to a proper date value
    • You can convert other date types using the format= argument

Example output

conm date revtq revtq_gr revtq_yoy revtq_d
ALLIED STORES 1962-04-30 156.5 NA NA NA
ALLIED STORES 1962-07-31 161.9 0.0345048 NA 5.4
ALLIED STORES 1962-10-31 176.9 0.0926498 NA 15.0
ALLIED STORES 1963-01-31 275.5 0.5573770 NA 98.6
ALLIED STORES 1963-04-30 171.1 -0.3789474 0.0932907 -104.4
ALLIED STORES 1963-07-31 182.2 0.0648743 0.1253860 11.1
## # A tibble: 6 x 3
##   conm          date       datadate  
##   <chr>         <date>     <chr>     
## 1 ALLIED STORES 1962-04-30 1962/04/30
## 2 ALLIED STORES 1962-07-31 1962/07/31
## 3 ALLIED STORES 1962-10-31 1962/10/31
## 4 ALLIED STORES 1963-01-31 1963/01/31
## 5 ALLIED STORES 1963-04-30 1963/04/30
## 6 ALLIED STORES 1963-07-31 1963/07/31

Create 8 quarters (2 years) of lags

# Custom Function to generate a series of lags
multi_lag <- function(df, lags, var, ext="") {
  lag_names <- paste0(var,ext,lags)
  lag_funs <- setNames(paste("dplyr::lag(.,",lags,")"), lag_names)
  df %>% group_by(gvkey) %>% mutate_at(vars(var), funs_(lag_funs)) %>% ungroup()  
}
# Generate lags "revtq_l#"
df <- multi_lag(df, 1:8, "revtq", "_l")

# Generate changes "revtq_gr#"
df <- multi_lag(df, 1:8, "revtq_gr")

# Generate year-over-year changes "revtq_yoy#"
df <- multi_lag(df, 1:8, "revtq_yoy")

# Generate first differences "revtq_d#"
df <- multi_lag(df, 1:8, "revtq_d")

# Equivalent brute force code for this is in the appendix
  • paste0(): creates a string vector by concatenating all inputs
  • paste(): same as paste0(), but with spaces added in between
  • setNames(): allows for storing a value and name simultaneously
  • mutate_at(): is like mutate but with a list of functions

Example output

conm date revtq revtq_l1 revtq_l2 revtq_l3 revtq_l4
ALLIED STORES 1962-04-30 156.5 NA NA NA NA
ALLIED STORES 1962-07-31 161.9 156.5 NA NA NA
ALLIED STORES 1962-10-31 176.9 161.9 156.5 NA NA
ALLIED STORES 1963-01-31 275.5 176.9 161.9 156.5 NA
ALLIED STORES 1963-04-30 171.1 275.5 176.9 161.9 156.5
ALLIED STORES 1963-07-31 182.2 171.1 275.5 176.9 161.9

Clean and split into training and testing

# Clean the data: Replace NaN, Inf, and -Inf with NA
df <- df %>%
  mutate_if(is.numeric, funs(replace(., !is.finite(.), NA)))

# Split into training and testing data
# Training data: We'll use data released before 2015
train <- filter(df, year(date) < 2015)

# Testing data: We'll use data released 2015 through 2018
test <- filter(df, year(date) >= 2015)
  • Same cleaning function as last week:
    • Replaces all NaN, Inf, and -Inf with NA
  • year() comes from lubridate

Univariate stats

Univariate stats

  • To get a better grasp on the problem, looking at univariate stats can help
    • Summary stats (using summary())
    • Correlations using cor()
    • Plots using your preferred package such as ggplot2
summary(df[,c("revtq","revtq_gr","revtq_yoy", "revtq_d","fqtr")])
##      revtq              revtq_gr         revtq_yoy      
##  Min.   :     0.00   Min.   :-1.0000   Min.   :-1.0000  
##  1st Qu.:    64.46   1st Qu.:-0.1112   1st Qu.: 0.0077  
##  Median :   273.95   Median : 0.0505   Median : 0.0740  
##  Mean   :  2439.38   Mean   : 0.0650   Mean   : 0.1273  
##  3rd Qu.:  1254.21   3rd Qu.: 0.2054   3rd Qu.: 0.1534  
##  Max.   :136267.00   Max.   :14.3333   Max.   :47.6600  
##  NA's   :367         NA's   :690       NA's   :940      
##     revtq_d               fqtr      
##  Min.   :-24325.21   Min.   :1.000  
##  1st Qu.:   -19.33   1st Qu.:1.000  
##  Median :     4.30   Median :2.000  
##  Mean   :    22.66   Mean   :2.478  
##  3rd Qu.:    55.02   3rd Qu.:3.000  
##  Max.   : 15495.00   Max.   :4.000  
##  NA's   :663

ggplot2 for visualization

  • The next slides will use some custom functions using ggplot2
  • ggplot2 has an odd syntax:
    • It doesn’t use pipes (%>%), but instead adds everything together (+)
library(ggplot2)  # or tidyverse -- it's part of tidyverse
df %>%
  ggplot(aes(y=var_for_y_axis, x=var_for_y_axis)) +
  geom_point()  # scatterplot
  • aes() is for aesthetics – how the chart is set up
  • Other useful aesthetics:
    • group= to set groups to list in the legend. Not needed if using the below though
    • color= to set color by some grouping variable. Put factor() around the variable if you want discrete groups, otherwise it will do a color scale (light to dark)
    • shape= to set shapes for points – see here for a list

ggplot2 for visualization

library(ggplot2)  # or tidyverse -- it's part of tidyverse
df %>%
  ggplot(aes(y=var_for_y_axis, x=var_for_y_axis)) +
  geom_point()  # scatterplot
  • geom stands for geometry – any shapes, lines, etc. start with geom
  • Other useful geoms:
    • geom_line(): makes a line chart
    • geom_bar(): makes a bar chart – y is the height, x is the category
    • geom_smooth(method="lm"): Adds a linear regression into the chart
    • geom_abline(slope=1): Adds a 45° line
  • Add xlab("Label text here") to change the x-axis label
  • Add ylab("Label text here") to change the y-axis label
  • Add ggtitle("Title text here") to add a title
  • Plenty more details in the ‘Data Visualization Cheat Sheet’ on eLearn

Plotting: Distribution of revenue

  1. Revenue

  2. Quarterly growth
  1. Year-over-year growth

  2. First difference

What do we learn from these graphs?

  1. Revenue
    • ~
  2. Quarterly growth
    • ~
  3. Year-over-year growth
    • ~
  4. First difference
    • ~

What do we learn from these graphs?

  1. Revenue
    • This is really skewed data – a lot of small revenue quarters, but a significant amount of large revenue quarters in the tail
      • Potential fix: use log(revtq)?
  2. Quarterly growth
    • Quarterly growth is reasonably close to normally distributed
      • Good for OLS
  3. Year-over-year growth
    • Year over year growth is reasonably close to normally distributed
      • Good for OLS
  4. First difference
    • Reasonably close to normally distributed, with really long tails
      • Good enough for OLS

Plotting: Mean revenue by quarter

  1. Revenue

  2. Quarterly growth
  1. Year-over-year growth

  2. First difference

What do we learn from these graphs?

  1. Revenue
    • ~
  2. Quarterly growth
    • ~
  3. Year-over-year growth
    • ~
  4. First difference
    • ~

What do we learn from these graphs?

  1. Revenue
    • Revenue seems cyclical!
  2. Quarterly growth
    • Definitely cyclical!
  3. Year-over-year growth
    • Year over year difference is less cyclical – more constant
  4. First difference
    • Definitely cyclical!

Plotting: Revenue vs lag by quarter

  1. Revenue

  2. Quarterly growth
  1. Year-over-year growth

  2. First difference

What do we learn from these graphs?

  1. Revenue
    • Revenue is really linear! But each quarter has a distinct linear relation.
  2. Quarterly growth
    • All over the place. Each quarter appears to have a different pattern though. Quarters will matter.
  3. Year-over-year growth
    • Linear but noisy.
  4. First difference
    • Again, all over the place. Each quarter appears to have a different pattern though. Quarters will matter.

Correlation matrices

cor(train[,c("revtq","revtq_l1","revtq_l2","revtq_l3", "revtq_l4")],
    use="complete.obs")
##              revtq  revtq_l1  revtq_l2  revtq_l3  revtq_l4
## revtq    1.0000000 0.9916167 0.9938489 0.9905522 0.9972735
## revtq_l1 0.9916167 1.0000000 0.9914767 0.9936977 0.9898184
## revtq_l2 0.9938489 0.9914767 1.0000000 0.9913489 0.9930152
## revtq_l3 0.9905522 0.9936977 0.9913489 1.0000000 0.9906006
## revtq_l4 0.9972735 0.9898184 0.9930152 0.9906006 1.0000000
cor(train[,c("revtq_gr","revtq_gr1","revtq_gr2","revtq_gr3", "revtq_gr4")],
    use="complete.obs")
##              revtq_gr   revtq_gr1   revtq_gr2   revtq_gr3   revtq_gr4
## revtq_gr   1.00000000 -0.32291329  0.06299605 -0.22769442  0.64570015
## revtq_gr1 -0.32291329  1.00000000 -0.31885020  0.06146805 -0.21923630
## revtq_gr2  0.06299605 -0.31885020  1.00000000 -0.32795121  0.06775742
## revtq_gr3 -0.22769442  0.06146805 -0.32795121  1.00000000 -0.31831023
## revtq_gr4  0.64570015 -0.21923630  0.06775742 -0.31831023  1.00000000

Retail revenue has really high autocorrelation! Concern for multicolinearity. Revenue growth is less autocorrelated and oscillates.

Correlation matrices

cor(train[,c("revtq_yoy","revtq_yoy1","revtq_yoy2","revtq_yoy3", "revtq_yoy4")],
    use="complete.obs")
##            revtq_yoy revtq_yoy1 revtq_yoy2 revtq_yoy3 revtq_yoy4
## revtq_yoy  1.0000000  0.6554179  0.4127263  0.4196003  0.1760055
## revtq_yoy1 0.6554179  1.0000000  0.5751128  0.3665961  0.3515105
## revtq_yoy2 0.4127263  0.5751128  1.0000000  0.5875643  0.3683539
## revtq_yoy3 0.4196003  0.3665961  0.5875643  1.0000000  0.5668211
## revtq_yoy4 0.1760055  0.3515105  0.3683539  0.5668211  1.0000000
cor(train[,c("revtq_d","revtq_d1","revtq_d2","revtq_d3", "revtq_d4")],
    use="complete.obs")
##             revtq_d   revtq_d1   revtq_d2   revtq_d3   revtq_d4
## revtq_d   1.0000000 -0.6181516  0.3309349 -0.6046998  0.9119911
## revtq_d1 -0.6181516  1.0000000 -0.6155259  0.3343317 -0.5849841
## revtq_d2  0.3309349 -0.6155259  1.0000000 -0.6191366  0.3165450
## revtq_d3 -0.6046998  0.3343317 -0.6191366  1.0000000 -0.5864285
## revtq_d4  0.9119911 -0.5849841  0.3165450 -0.5864285  1.0000000

Year over year change fixes the multicollinearity issue. First difference oscillates like quarter over quarter growth.

R Practice

Forecasting

1 period models

  1. 1 Quarter lag
    • We saw a very strong linear pattern here earlier
mod1 <- lm(revtq ~ revtq_l1, data=train)
  1. Quarter and year lag
    • Year-over-year seemed pretty constant
mod2 <- lm(revtq ~ revtq_l1 + revtq_l4, data=train)
  1. 2 years of lags
    • Other lags could also help us predict
mod3 <- lm(revtq ~ revtq_l1 + revtq_l2 + revtq_l3 + revtq_l4 + 
             revtq_l5 + revtq_l6 + revtq_l7 + revtq_l8, data=train)
  1. 2 years of lags, by observation quarter
    • Take into account cyclicality observed in bar charts
mod4 <- lm(revtq ~ (revtq_l1 + revtq_l2 + revtq_l3 + revtq_l4 +
             revtq_l5 + revtq_l6 + revtq_l7 + revtq_l8):factor(fqtr),
           data=train)

Quarter lag

summary(mod1)
## 
## Call:
## lm(formula = revtq ~ revtq_l1, data = train)
## 
## Residuals:
##      Min       1Q   Median       3Q      Max 
## -24438.7    -34.0    -11.7     34.6  15200.5 
## 
## Coefficients:
##              Estimate Std. Error t value Pr(>|t|)    
## (Intercept) 15.639975  13.514877   1.157    0.247    
## revtq_l1     1.003038   0.001556 644.462   <2e-16 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 1152 on 7676 degrees of freedom
##   (662 observations deleted due to missingness)
## Multiple R-squared:  0.9819, Adjusted R-squared:  0.9819 
## F-statistic: 4.153e+05 on 1 and 7676 DF,  p-value: < 2.2e-16

Quarter and year lag

summary(mod2)
## 
## Call:
## lm(formula = revtq ~ revtq_l1 + revtq_l4, data = train)
## 
## Residuals:
##      Min       1Q   Median       3Q      Max 
## -20245.7    -18.4     -4.4     19.1   9120.8 
## 
## Coefficients:
##             Estimate Std. Error t value Pr(>|t|)    
## (Intercept) 5.444986   7.145633   0.762    0.446    
## revtq_l1    0.231759   0.005610  41.312   <2e-16 ***
## revtq_l4    0.815570   0.005858 139.227   <2e-16 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 592.1 on 7274 degrees of freedom
##   (1063 observations deleted due to missingness)
## Multiple R-squared:  0.9954, Adjusted R-squared:  0.9954 
## F-statistic: 7.94e+05 on 2 and 7274 DF,  p-value: < 2.2e-16

2 years of lags

summary(mod3)
## 
## Call:
## lm(formula = revtq ~ revtq_l1 + revtq_l2 + revtq_l3 + revtq_l4 + 
##     revtq_l5 + revtq_l6 + revtq_l7 + revtq_l8, data = train)
## 
## Residuals:
##     Min      1Q  Median      3Q     Max 
## -5005.6   -12.9    -3.7     9.3  5876.3 
## 
## Coefficients:
##             Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  4.02478    4.37003   0.921   0.3571    
## revtq_l1     0.77379    0.01229  62.972  < 2e-16 ***
## revtq_l2     0.10497    0.01565   6.707 2.16e-11 ***
## revtq_l3    -0.03091    0.01538  -2.010   0.0445 *  
## revtq_l4     0.91982    0.01213  75.800  < 2e-16 ***
## revtq_l5    -0.76459    0.01324 -57.749  < 2e-16 ***
## revtq_l6    -0.08080    0.01634  -4.945 7.80e-07 ***
## revtq_l7     0.01146    0.01594   0.719   0.4721    
## revtq_l8     0.07924    0.01209   6.554 6.03e-11 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 346 on 6666 degrees of freedom
##   (1665 observations deleted due to missingness)
## Multiple R-squared:  0.9986, Adjusted R-squared:  0.9986 
## F-statistic: 5.802e+05 on 8 and 6666 DF,  p-value: < 2.2e-16

2 years of lags, by observation quarter

summary(mod4)
## 
## Call:
## lm(formula = revtq ~ (revtq_l1 + revtq_l2 + revtq_l3 + revtq_l4 + 
##     revtq_l5 + revtq_l6 + revtq_l7 + revtq_l8):factor(fqtr), 
##     data = train)
## 
## Residuals:
##     Min      1Q  Median      3Q     Max 
## -6066.6   -13.9     0.1    15.1  4941.1 
## 
## Coefficients:
##                         Estimate Std. Error t value Pr(>|t|)    
## (Intercept)            -0.201107   4.004046  -0.050 0.959944    
## revtq_l1:factor(fqtr)1  0.488584   0.021734  22.480  < 2e-16 ***
## revtq_l1:factor(fqtr)2  1.130563   0.023017  49.120  < 2e-16 ***
## revtq_l1:factor(fqtr)3  0.774983   0.028727  26.977  < 2e-16 ***
## revtq_l1:factor(fqtr)4  0.977353   0.026888  36.349  < 2e-16 ***
## revtq_l2:factor(fqtr)1  0.258024   0.035136   7.344 2.33e-13 ***
## revtq_l2:factor(fqtr)2 -0.100284   0.024664  -4.066 4.84e-05 ***
## revtq_l2:factor(fqtr)3  0.212954   0.039698   5.364 8.40e-08 ***
## revtq_l2:factor(fqtr)4  0.266761   0.035226   7.573 4.14e-14 ***
## revtq_l3:factor(fqtr)1  0.124187   0.036695   3.384 0.000718 ***
## revtq_l3:factor(fqtr)2 -0.042214   0.035787  -1.180 0.238197    
## revtq_l3:factor(fqtr)3 -0.005758   0.024367  -0.236 0.813194    
## revtq_l3:factor(fqtr)4 -0.308661   0.038974  -7.920 2.77e-15 ***
## revtq_l4:factor(fqtr)1  0.459768   0.038266  12.015  < 2e-16 ***
## revtq_l4:factor(fqtr)2  0.684943   0.033366  20.528  < 2e-16 ***
## revtq_l4:factor(fqtr)3  0.252169   0.035708   7.062 1.81e-12 ***
## revtq_l4:factor(fqtr)4  0.817136   0.017927  45.582  < 2e-16 ***
## revtq_l5:factor(fqtr)1 -0.435406   0.023278 -18.704  < 2e-16 ***
## revtq_l5:factor(fqtr)2 -0.725000   0.035497 -20.424  < 2e-16 ***
## revtq_l5:factor(fqtr)3 -0.160408   0.036733  -4.367 1.28e-05 ***
## revtq_l5:factor(fqtr)4 -0.473030   0.033349 -14.184  < 2e-16 ***
## revtq_l6:factor(fqtr)1  0.059832   0.034672   1.726 0.084453 .  
## revtq_l6:factor(fqtr)2  0.154990   0.025368   6.110 1.05e-09 ***
## revtq_l6:factor(fqtr)3 -0.156840   0.041147  -3.812 0.000139 ***
## revtq_l6:factor(fqtr)4 -0.106082   0.037368  -2.839 0.004541 ** 
## revtq_l7:factor(fqtr)1  0.060031   0.038599   1.555 0.119936    
## revtq_l7:factor(fqtr)2  0.061381   0.034510   1.779 0.075344 .  
## revtq_l7:factor(fqtr)3  0.028149   0.025385   1.109 0.267524    
## revtq_l7:factor(fqtr)4 -0.277337   0.039380  -7.043 2.08e-12 ***
## revtq_l8:factor(fqtr)1 -0.016637   0.033568  -0.496 0.620177    
## revtq_l8:factor(fqtr)2 -0.152379   0.028014  -5.439 5.54e-08 ***
## revtq_l8:factor(fqtr)3  0.052208   0.027334   1.910 0.056179 .  
## revtq_l8:factor(fqtr)4  0.103495   0.015777   6.560 5.78e-11 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 299.7 on 6642 degrees of freedom
##   (1665 observations deleted due to missingness)
## Multiple R-squared:  0.9989, Adjusted R-squared:  0.9989 
## F-statistic: 1.935e+05 on 32 and 6642 DF,  p-value: < 2.2e-16

Testing out of sample

  • RMSE: Root mean square Error
  • RMSE is very affected by outliers, and a bad choice for noisy data that you are OK with missing a few outliers here and there
    • Doubling error quadruples that part of the error
rmse <- function(v1, v2) {
  sqrt(mean((v1 - v2)^2, na.rm=T))
}
  • MAE: Mean absolute error
  • MAE is measures average accuracy with no weighting
    • Doubling error doubles that part of the error
mae <- function(v1, v2) {
  mean(abs(v1-v2), na.rm=T)
}

Both are commonly used for evaluating OLS out of sample

Testing out of sample

adj_r_sq rmse_in mae_in rmse_out mae_out
1 period 0.9818514 1151.3535 322.73819 2947.3619 1252.5196
1 and 4 periods 0.9954393 591.9500 156.20811 1400.3841 643.9823
8 periods 0.9985643 345.8053 94.91083 677.6218 340.8236
8 periods w/ quarters 0.9989231 298.9557 91.28056 645.5415 324.9395

1 quarter model

8 period model, by quarter

What about for revenue growth?

Backing out a revenue prediction, revt_t=(1+growth_t)\times revt_{t-1}

adj_r_sq rmse_in mae_in rmse_out mae_out
1 period 0.0910390 1106.3730 308.48331 3374.728 1397.6541
1 and 4 periods 0.4398456 530.6444 154.15086 1447.035 679.3536
8 periods 0.6761666 456.2551 123.34075 1254.201 584.9709
8 periods w/ quarters 0.7758834 378.4082 98.45751 1015.971 436.1522

1 quarter model

8 period model, by quarter

What about for YoY revenue growth?

Backing out a revenue prediction, revt_t=(1+yoy\_growth_t)\times revt_{t-4}

adj_r_sq rmse_in mae_in rmse_out mae_out
1 period 0.4370372 513.3264 129.2309 1867.4957 798.0327
1 and 4 periods 0.5392281 487.6441 126.6012 1677.4003 731.2841
8 periods 0.5398870 384.2923 101.0104 822.0065 403.5445
8 periods w/ quarters 0.1563169 714.4285 195.3204 1231.8436 617.2989

1 quarter model

8 period model

What about for first difference?

Backing out a revenue prediction, revt_t= change_t+ revt_{t-1}

adj_r_sq rmse_in mae_in rmse_out mae_out
1 period 0.3532044 896.7969 287.77940 2252.7605 1022.0960
1 and 4 periods 0.8425348 454.8651 115.52694 734.8120 377.5281
8 periods 0.9220849 333.0054 95.95924 651.4967 320.0567
8 periods w/ quarters 0.9397434 292.3102 86.95563 659.4412 319.7305

1 quarter model

8 period model, by quarter

Takeaways

  1. The first difference model works about as well as the revenue model at predicting next quarter revenue
    • From earlier, it doesn’t suffer (as much) from multicollinearity either
      • This is why time series analysis is often done on first differences
        • Or second differences (difference in differences)
  2. The other models perform pretty well as well
  3. Extra lags generally seems helpful when accounting for cyclicality
  4. Regressing by quarter helps a bit, particularly with revenue growth

What about for revenue growth?

Predicting quarter over quarter revenue growth itself

adj_r_sq rmse_in mae_in rmse_out mae_out
1 period 0.0910390 0.3509269 0.2105219 0.2257396 0.1750580
1 and 4 periods 0.4398456 0.2681899 0.1132003 0.1597771 0.0998087
8 periods 0.6761666 0.1761825 0.0867347 0.1545298 0.0845826
8 periods w/ quarters 0.7758834 0.1462979 0.0765792 0.1459460 0.0703554

1 quarter model

8 period model, by quarter

What about for YoY revenue growth?

Predicting YoY revenue growth itself
adj_r_sq rmse_in mae_in rmse_out mae_out
1 period 0.4370372 0.3116645 0.1114610 0.1515638 0.0942544
1 and 4 periods 0.5392281 0.2451749 0.1015699 0.1498755 0.0896079
8 periods 0.5398870 0.1928940 0.0764447 0.1346238 0.0658011
8 periods w/ quarters 0.1563169 0.3006075 0.1402156 0.1841025 0.0963205

1 quarter model

8 period model

What about for first difference?

Predicting first difference in revenue itself

adj_r_sq rmse_in mae_in rmse_out mae_out
1 period 0.3532044 896.7969 287.77940 2252.7605 1022.0960
1 and 4 periods 0.8425348 454.8651 115.52694 734.8120 377.5281
8 periods 0.9220849 333.0054 95.95924 651.4967 320.0567
8 periods w/ quarters 0.9397434 292.3102 86.95563 659.4412 319.7305

1 quarter model

8 period model, by quarter

Case: Advanced revenue prediction

RS Metrics’ approach

Read the press release: rmc.link/420class4

  • How does RS Metrics approach revenue prediction?
  • What other creative ways might there be?

Weekly revenue prediction

Shifted to week 5

End matter

For next week

  • For next week:
    • First individual assignment
      • Finish by the end of Thursday
      • Submit on eLearn
    • Datacamp
      • Practice a bit more to keep up to date
        • Using R more will make it more natural

Packages used for these slides

Custom code

# Brute force code for variable generation of quarterly data lags
df <- df %>%
  group_by(gvkey) %>%
  mutate(revtq_lag1=lag(revtq), revtq_lag2=lag(revtq, 2),
         revtq_lag3=lag(revtq, 3), revtq_lag4=lag(revtq, 4),
         revtq_lag5=lag(revtq, 5), revtq_lag6=lag(revtq, 6),
         revtq_lag7=lag(revtq, 7), revtq_lag8=lag(revtq, 8),
         revtq_lag9=lag(revtq, 9), revtq_gr=revtq / revtq_lag1 - 1,
         revtq_gr1=lag(revtq_gr), revtq_gr2=lag(revtq_gr, 2),
         revtq_gr3=lag(revtq_gr, 3), revtq_gr4=lag(revtq_gr, 4),
         revtq_gr5=lag(revtq_gr, 5), revtq_gr6=lag(revtq_gr, 6),
         revtq_gr7=lag(revtq_gr, 7), revtq_gr8=lag(revtq_gr, 8),
         revtq_yoy=revtq / revtq_lag4 - 1, revtq_yoy1=lag(revtq_yoy),
         revtq_yoy2=lag(revtq_yoy, 2), revtq_yoy3=lag(revtq_yoy, 3),
         revtq_yoy4=lag(revtq_yoy, 4), revtq_yoy5=lag(revtq_yoy, 5),
         revtq_yoy6=lag(revtq_yoy, 6), revtq_yoy7=lag(revtq_yoy, 7),
         revtq_yoy8=lag(revtq_yoy, 8), revtq_d=revtq - revtq_l1,
         revtq_d1=lag(revtq_d), revtq_d2=lag(revtq_d, 2),
         revtq_d3=lag(revtq_d, 3), revtq_d4=lag(revtq_d, 4),
         revtq_d5=lag(revtq_d, 5), revtq_d6=lag(revtq_d, 6),
         revtq_d7=lag(revtq_d, 7), revtq_d8=lag(revtq_d, 8)) %>%
  ungroup()
# Custom html table for small data frames
library(knitr)
library(kableExtra)
html_df <- function(text, cols=NULL, col1=FALSE, full=F) {
  if(!length(cols)) {
    cols=colnames(text)
  }
  if(!col1) {
    kable(text,"html", col.names = cols, align = c("l",rep('c',length(cols)-1))) %>%
      kable_styling(bootstrap_options = c("striped","hover"), full_width=full)
  } else {
    kable(text,"html", col.names = cols, align = c("l",rep('c',length(cols)-1))) %>%
      kable_styling(bootstrap_options = c("striped","hover"), full_width=full) %>%
      column_spec(1,bold=T)
  }
}

Custom code

# These functions are a bit ugly, but can construct many charts quickly
# eval(parse(text=var)) is just a way to convert the string name to a variable reference
# Density plot for 1st to 99th percentile of data
plt_dist <- function(df,var) {
  df %>%
    filter(eval(parse(text=var)) < quantile(eval(parse(text=var)),0.99, na.rm=TRUE),
           eval(parse(text=var)) > quantile(eval(parse(text=var)),0.01, na.rm=TRUE)) %>%
    ggplot(aes(x=eval(parse(text=var)))) + 
    geom_density() + xlab(var)
}
# Density plot for 1st to 99th percentile of both columns
plt_bar <- function(df,var) {
  df %>%
    filter(eval(parse(text=var)) < quantile(eval(parse(text=var)),0.99, na.rm=TRUE),
           eval(parse(text=var)) > quantile(eval(parse(text=var)),0.01, na.rm=TRUE)) %>%
    ggplot(aes(y=eval(parse(text=var)), x=fqtr)) + 
    geom_bar(stat = "summary", fun.y = "mean") + xlab(var)
}
# Scatter plot with lag for 1st to 99th percentile of data
plt_sct <- function(df,var1, var2) {
  df %>%
    filter(eval(parse(text=var1)) < quantile(eval(parse(text=var1)),0.99, na.rm=TRUE),
           eval(parse(text=var2)) < quantile(eval(parse(text=var2)),0.99, na.rm=TRUE),
           eval(parse(text=var1)) > quantile(eval(parse(text=var1)),0.01, na.rm=TRUE),
           eval(parse(text=var2)) > quantile(eval(parse(text=var2)),0.01, na.rm=TRUE)) %>%
    ggplot(aes(y=eval(parse(text=var1)), x=eval(parse(text=var2)), color=factor(fqtr))) + 
    geom_point() + geom_smooth(method = "lm") + ylab(var1) + xlab(var2)
}
# Calculating various in and out of sample statistics
models <- list(mod1,mod2,mod3,mod4)
model_names <- c("1 period", "1 and 4 period", "8 periods", "8 periods w/ quarters")

df_test <- data.frame(adj_r_sq=sapply(models, function(x)summary(x)[["adj.r.squared"]]),
                      rmse_in=sapply(models, function(x)rmse(train$revtq, predict(x,train))),
                      mae_in=sapply(models, function(x)mae(train$revtq, predict(x,train))),
                      rmse_out=sapply(models, function(x)rmse(test$revtq, predict(x,test))),
                      mae_out=sapply(models, function(x)mae(test$revtq, predict(x,test))))
rownames(df_test) <- model_names
html_df(df_test)  # Custom function using knitr and kableExtra