Uncovering the Data Generation Process in Econometrics

Have you ever looked at a news report about rising inflation, a sudden dip in the stock market, or a company’s new pricing strategy and wondered, “How do they *know* that?” Economists and data scientists seem to make bold claims based on data, but the data itself is just a collection of numbers. The real magic-and the real challenge-lies in understanding the story *behind* those numbers. In econometrics, we have a name for this hidden story: the Data Generation Process, or DGP. It’s the true, underlying economic reality that churns out the data we observe every day, and trying to uncover it is the ultimate goal of any econometrician.

What is the data generation process (DGP)? 🔗

At its core, the Data Generation Process is the true, unknown, and almost certainly unknowable mechanism that produces the observed data. Think of it as the complete, hyper-complex set of rules that govern an economic phenomenon. If we’re studying crop prices, the DGP isn’t just a spreadsheet of prices and rainfall; it’s the *entire* real-world system, including the physics of weather patterns, the biology of crop growth, the psychology of farmer expectations, the logistics of supply chains, and the political decisions affecting subsidies.

The invisible factory 🔗

You can imagine the DGP as an invisible factory. We stand outside and see the finished products-the data-coming out on a conveyor belt. We see ‘Monthly Inflation: 3%’, ‘GDP Growth: 1.5%’, or ‘Sales: 10,000 units’. Our job is to look at those finished products and try to reverse-engineer the blueprint of the factory. What machines are inside? How are they connected? Which levers are being pulled?

The catch is that we are never allowed inside the factory. We will *never* see the true DGP. It’s simply too complex for any one person or computer to map out completely. So, if we can’t know the truth, what’s the point?

Models as our map 🔗

This is where econometric modeling comes in. Since the DGP is the “territory,” our goal is to build a “map”-an econometric model. As the famous saying goes, “All models are wrong, but some are useful.” Our map will never be the territory, but a good map can help us navigate. We use statistical and mathematical techniques to build a simplified model that we hope approximates the *essential relationships* of the true DGP. We aren’t trying to model every molecule of water in a weather system, but we *are* trying to capture the essential relationship between atmospheric pressure, humidity, and the likelihood of rain.

The central question then becomes: How do we even start building this map? Where do we lay down the first road?

How do we start building the map? 🔗

When an econometrician sits down with a dataset, they face a choice. They need to decide which variables to include in their model. Broadly, two competing philosophies guide this search: starting simple and building up, or starting complex and trimming down.

The ‘specific-to-general’ approach 🔗

This approach is often the most intuitive. You start with a very simple, “specific” model based on a core economic theory. For example, to model demand for ice cream, you might start with a simple model: Ice Cream Sales = f(Price).

You run this model and test it. You’ll quickly find it’s not very good. It doesn’t explain much. So, you add another variable: Ice Cream Sales = f(Price, Temperature). That’s better! Then you test it again and find something else is missing. So you add another: Ice Cream Sales = f(Price, Temperature, Advertising_Budget).

This is like building a house brick by brick, without a complete blueprint. While it feels logical, it’s fraught with peril. The biggest risk is omitted variable bias. If you leave out a critical variable (like consumer income), your model might incorrectly attribute its effect to one of the variables you *did* include (like price). You might conclude that price is hugely important when, in reality, both price and sales were being driven by the third, unobserved variable.

The ‘general-to-specific’ (GETS) approach 🔗

This is the approach now preferred by most modern econometricians. Instead of starting small, you start “general.” You throw in *everything* that could plausibly be relevant, based on economic theory and past research.

Your “general model” might look like: Sales = f(Price, Temperature, Advertising, Income, Competitor_Price, Holiday, Location,…).

This initial model is big, messy, and hard to interpret. But it’s also less likely to suffer from that nasty omitted variable bias. From this complex starting point, you begin a systematic process of simplification. You test each variable to see if it’s statistically significant. Is ‘Location’ actually contributing anything? No? You remove it. Is ‘Holiday’ relevant? Yes? You keep it. You methodically trim the fat, testing at each step to ensure your model remains a valid description of the data.

This is like carving a statue from a large block of marble. You start with the whole block (the general model) and carefully chip away everything that *doesn’t* look like the statue (the irrelevant variables). What’s left, you hope, is a clean and accurate representation of the underlying form. This robust method, often associated with economists at the London School of Economics (LSE), is a powerful way to let the data, guided by theory, reveal the most important relationships.

Modeling as a journey: the iterative process 🔗

Building an econometric model isn’t a “one-and-done” task. You don’t just run the GETS approach once and publish your results. The search for the DGP is an iterative process-a continuous loop of refinement.

Guess, test, revise, repeat 🔗

The model-building process is a cycle that looks something like this:

1. Guess the DGP: You start by postulating a “general” model. This is your first, best guess at what the unknown DGP looks like, based on all available economic theory.
2. Assume a probability structure: You make some technical assumptions (e.g., “we assume the errors are normally distributed”).
3. Test the model: You confront your model with empirical evidence (the data). You run diagnostic tests to see if your assumptions hold and if the model is statistically adequate. (More on this in a moment).
4. Revise the model: Your tests will almost certainly reveal problems. Your model is “falsified” by the data. So, you go back to step 1, revising your model based on what you’ve learned. You repeat this loop-guess, test, revise, repeat-until you arrive at a “satisfactory” model that is statistically sound, makes economic sense, and can’t be easily rejected.

A real-world example: forecasting inflation in India 🔗

This process isn’t just an academic exercise; it has massive real-world consequences. Think about the Reserve Bank of India (RBI). One of its primary jobs is to manage inflation. To do this, it operates under a flexible inflation-targeting (FIT) framework. This framework means the RBI’s decisions on interest rates-which affect your car loan, home mortgage, and savings account-depend heavily on its *forecasts* of future inflation.

Those forecasts are generated by complex econometric models. The economists at the RBI don’t just build a model and let it run forever. They are in a constant iterative loop. They build a general model (Step 1), test it against the latest data (Step 3), and when it (inevitably) shows errors, they *revise* it (Step 4) to incorporate new information-like a sudden jump in oil prices or a weak monsoon. This constant process of testing and revising is essential for steering the national economy.

The engine of progress: hypothesis testing 🔗

So, what does it mean to “test” a model? This is the engine of the entire iterative process, and it’s built on a powerful idea from the philosophy of science: falsification.

You can’t prove it right, but you can prove it wrong 🔗

The philosopher Karl Popper had a profound insight into how science works. He argued that you can never *prove* a scientific theory is true. For centuries, people in Europe observed millions of white swans, leading to the “truth” that “all swans are white.” This theory was “verified” millions of times. But it wasn’t true. It only took the discovery of *one* black swan in Australia to completely falsify and destroy the theory.

Popper argued that a theory is only “scientific” if it is falsifiable-that is, if there is some conceivable observation that could prove it wrong. We don’t make progress by “proving” our models right, but by relentlessly trying (and failing) to prove them wrong.

The null hypothesis: our scientific ‘punching bag’ 🔗

In econometrics, we put Popper’s idea into practice using hypothesis testing. We don’t try to prove that our variable (like ‘Temperature’) *does* affect ice cream sales. Instead, we do the opposite: we set up a “punching bag” called the null hypothesis (H0).

The null hypothesis is the boring, “nothing is happening” theory.

• Our Theory: Temperature has a significant effect on ice cream sales.
• The Null Hypothesis (H0): Temperature has *no effect* on ice cream sales.

We then use our data to attack the null hypothesis. We’re looking for evidence that is so strong, so unlikely to have occurred by random chance, that it “rejects” the null hypothesis. If we can confidently *reject* the idea that temperature has no effect, we find support for our alternative theory that it *does* have an effect. By falsifying the “no effect” theory, we provide evidence that our model’s claims are supported by the data. This is the logical tool we use at every step of the general-to-specific process to decide which variables to keep and which to discard.

The Data Generation Process, that true, hidden reality, will always remain a mystery. But through a smart strategy-starting general, trimming down, and being relentlessly critical through iterative hypothesis testing-we can build models that are more than just “wrong.” We can build maps that are useful, insightful, and help us navigate a complex economic world.

What do you think? When you read an economic forecast in the news, does thinking about the unknown “DGP” and the iterative modeling process make you see that forecast differently? If you were trying to build a “general” model for something like student exam scores, what are the first 10 variables you would throw in before you started trimming?

References

1. https://www.investopedia.com/terms/e/econometrics.asp
2. https://www.federalreserve.gov/pubs/ifdp/2005/838/ifdp838.pdf
3. https://iegindia.org/upload/publication/Workpap/WP461.pdf
4. https://www.simplypsychology.org/karl-popper.html