In Chapter 2, we saw something important:

ordinary linear regression fails whenever data behaves differently from the assumptions of normality and constant variance.

That naturally leads to a question:

If ordinary regression fails, what replaces it?

The answer is: Generalized Linear Models (GLMs).

And Chapter 3 is where everything comes together.

This chapter is the foundation of:

• logistic regression,
• Poisson regression,
• Gamma regression,
• negative binomial models,
• survival models,
• even many machine learning frameworks.

If Chapter 2 explains why ordinary regression breaks,
Chapter 3 explains: how we fix it systematically.

---

The Big Idea of GLMs

GLMs are not “one model.”

They are: a framework.

A framework that allows us to use:

• different distributions,
• different variance structures,
• different link functions,
  while still keeping the spirit of regression.

Core GLM structure:

That single equation powers a huge portion of modern statistics.

---

The Three Components of a GLM

A GLM has three pieces.

---

1. Random Component

This specifies:

the probability distribution of the response variable.

Examples:

Type of Data	Distribution
Continuous symmetric	Normal
Counts	Poisson
Binary outcomes	Bernoulli
Positive skewed data	Gamma

This is the part where we respect:

• the structure of the data.

---

2. Systematic Component

This is the linear predictor:

which means:

This is still the “regression” part.

Examples:

• deployment,
• customer count,
• month,
• discount,
• seasonality.

All influence the expected outcome.

---

3. Link Function

The link function connects:

• the expected value of the response,
  to:
• the linear predictor.

This is the heart of GLMs.

---

Why Do We Need a Link Function?

Because many outcomes cannot directly behave like:

Example:

• probabilities must stay between 0 and 1,
• counts must stay positive,
• Gamma values must stay positive.

The link function transforms the mean into something that can safely behave linearly.

---

Example — Logistic Regression

Probabilities cannot exceed:

• 1,
  or go below:

So instead of modeling probability directly,
we model: log-odds.

The logit link:

transforms:

• probabilities,
  into:
• real numbers.

Now regression becomes possible.

---

Example — Poisson Regression

Counts must remain positive.

So Poisson regression uses: the log link.

Exponentiating gives:

which guarantees:

• positive predicted counts.

Very elegant.

---

Bernoulli Distribution — The Foundation of Logistic Regression

A Bernoulli random variable models: one trial with two outcomes.

Examples:

• buy/not buy,
• churn/no churn,
• sold/not sold.

Possible outcomes:

Outcome	Meaning
1	success
0	failure

---

Mean and Variance of Bernoulli

Mean:

Variance:

This is extremely important.

Notice:

• variance depends on the mean.

That is already different from ordinary regression.

---

A Beautiful Insight About Bernoulli Variance

The variance peaks at:

Why?

Because uncertainty is highest when:

• both outcomes are equally likely.

If:

• (p=0),
  or:
• (p=1),

there is almost no uncertainty.

So variance shrinks again.

This explains why Bernoulli behaves differently from Poisson.

---

Odds — The Key to Logistic Regression

Suppose:

• probability of winning = 0.75.

Then:

• probability of losing = 0.25.

Odds are:

Meaning:

• winning is three times as likely as losing.

Logistic regression models: the log of the odds.

---

Poisson Regression — Modeling Counts

Poisson regression is used for:

• sales counts,
• purchases,
• website clicks,
• arrivals,
• support tickets.

Its defining property:

Variance grows naturally with the mean.

This is exactly what count data often does.

---

Real Business Example

Suppose:

• small retailers sell 2 items/month,
• large retailers sell 100 items/month.

Large retailers naturally fluctuate more.

Poisson understands this.

Ordinary regression assumes:

• constant variability.

That becomes unrealistic.

---

Gamma Regression — Positive Skewed Data

Gamma regression handles:

• positive continuous,
• right-skewed data.

Examples:

• revenue per transaction,
• insurance claims,
• customer spending,
• waiting times.

Variance behaves like:

Var(Y)=\phi\mu^2

As mean grows:

• variability grows even faster.

Very common in business.

---

Canonical Links

Each GLM has a natural or canonical link.

Distribution	Canonical Link
Normal	identity
Bernoulli	logit
Poisson	log
Gamma	inverse

---

Identity Link

For normal regression:

Nothing changes.

This is why ordinary regression is actually: a special case of GLM.

That is a very important insight.

---

Overdispersion — When Poisson Breaks

Real count data often violates:
[
Var(Y)=\mu
]

Instead:

[
Var(Y)>\mu
]

This is called: overdispersion.

---

Why Overdispersion Happens

Because real-world systems are heterogeneous.

Different parts of the data behave differently.

Example:

Retailer Type	Average Sales
Weak	1
Medium	5
Strong	20

Combining them creates:

• extra variability.

That’s why overdispersion often indicates:

• hidden groups,
• omitted variables,
• clustering,
• latent heterogeneity.

---

Negative Binomial Regression

Negative binomial extends Poisson by adding extra variance.

Variance becomes:

where:

• $(\alpha)$ controls extra dispersion.

This is one of the most practical count models in real business analytics.

---

Real sales Example

Suppose you model:

• monthly sales counts,
• by category,
• over time.

You may include:

• month,
• customer count,
• year,
• seasonality.

A realistic model might look like:

where:

• (f(\cdot)) is a spline.

Now we are entering: Generalized Additive Models (GAMs).

---

Splines and Nonlinearity

Not all relationships are linear.

Example:

• adding customers may help sales strongly initially,
• then level off later.

A spline allows:

• smooth nonlinear effects.

Instead of:

we use:

This becomes a GAM.

---

Forecasting Inventory

One fascinating business application discussed was: ideal inventory forecasting.

Using:

• negative binomial counts,
• seasonality,
• customer count,
• splines,
• weighted recent data,
• lead times,
• safety stock.

This is exactly how advanced operational forecasting systems work.

---

Weighted Forecasting

You explored:

• 70% weight on recent 6 months,
• 30% weight on older 18 months.

This is actually very practical.

Why?

Because:

• markets evolve,
• trends shift,
• recent data often matters more.

---

Safety Stock — Beyond Normality

Traditional safety stock formulas assume:

• normal demand.

But negative binomial demand is:

• discrete,
• skewed,
• overdispersed.

So instead of:

• normal approximations,

you can use:

percentiles from the fitted distribution.

For example:

• forecast the next 12 months,
• scale to 4-month lead time,
• calculate 90th percentile demand.

That becomes: a probabilistic inventory target.

This is a far more modern approach.

---

The Deep Philosophy of GLMs

Chapter 3 teaches something profound:

data types matter.

You cannot force:

• counts,
• probabilities,
• skewed revenue,
• overdispersed demand,

into one simplistic framework.

GLMs adapt the model to the behavior of reality.

That is why they are so powerful.

---

Final Thought

Ordinary regression is only one small corner of statistical modeling.

GLMs generalize regression into a flexible system capable of modeling:

• counts,
• probabilities,
• skewed positive data,
• overdispersion,
• seasonality,
• nonlinear effects,
• real business processes.

And once you truly understand GLMs,
you begin to realize:

the distribution is not just mathematics —
it is a description of how reality behaves.