MarcusWelcome to VoraPrep Audio’s Office Hours. Today we're tackling a topic that’s critical for risk modeling: time series. Emma, let’s start with the scenario from the exam: a risk manager needs to forecast Value-at-Risk, but the data is tricky. The stock prices are drifting, and the returns have these wild swings followed by calm periods. Where do you even begin?

EmmaYou begin with the most fundamental concept: stationarity. Before you can model anything, you have to know if the statistical properties of your data are stable over time.

MarcusOkay, "stationarity." I see that word everywhere. What does it actually mean for a time series to be stationary?

EmmaThink of a river. A stationary river has a constant average depth and a consistent flow. Its properties are stable. You can take a sample today, and it will look statistically similar to a sample from last year. A non-stationary river, on the other hand, might be flooding—its average depth is rising—or drying up. Its properties are changing over time.

MarcusSo a stationary time series is predictable in its basic character, even if the individual values change. Why is that so important for our models?

EmmaBecause most of the classic models we use—like ARMA models, which we'll get to—are built on the assumption that the data is stationary. They assume a constant mean and constant variance. If you use them on non-stationary data, like raw stock prices, your model will produce nonsense. You'll get what we call spurious regressions and completely unreliable forecasts.

MarcusIt sounds like trying to measure the depth of a flooding river with a fixed ruler. So, how do we spot a non-stationary series?

EmmaThe first step is just to look at a plot of the data. Does it have a clear trend, up or down? That's a huge red flag. Technically, we look at something called the Autocorrelation Function, or ACF plot. For a non-stationary series, that ACF will decay very, very slowly.

MarcusAnd there are formal tests for this, right? The source material mentions Dickey-Fuller.

EmmaExactly. The Dickey-Fuller and Augmented Dickey-Fuller—or ADF—tests are the formal tools. Their main job is to test for something called a "unit root," which is the statistical signature of many non-stationary series, like a random walk. For the exam, just know that an ADF test helps you formally decide if your series is stationary or not.

MarcusOkay, so let's say we run the test and confirm our series is non-stationary. Is the data useless? Do we just give up?

EmmaNot at all. We just have to transform it to make it stationary. The most common method is called differencing. You simply take today's value and subtract yesterday's value. For stock prices, this is almost the same as calculating daily returns.

MarcusAh, and that’s why the source material says prices are typically non-stationary, but returns are usually stationary. We've differenced the price series to get the return series.

EmmaPrecisely. And if a series becomes stationary after differencing it one time, we say it's "integrated of order 1," which we write as I(1).

MarcusGot it. So, step one is always to check for stationarity and difference the data if needed. Once we have a stationary series to work with, how do we start modeling it? I see AR, MA, and ARMA. Let's start with AR models.

EmmaAR stands for Autoregressive. It means the model uses past values of the series itself to predict the current value. An AR(1) model, for example, says that today's value is some function of yesterday's value, plus a random error term.

MarcusSo it's like saying, "If the market was up yesterday, it's slightly more likely to be up today"? It has a memory of its own past.

EmmaThat's a great way to put it. The 'p' in AR(p) just tells you how many past periods, or lags, the model is looking at.

MarcusOkay, that makes sense. Then what about MA, for Moving Average? How is that different?

EmmaThis is a subtle but critical distinction. A Moving Average model, or MA(q), doesn't use past values of the series itself. It uses the past *error terms*—the shocks or surprises from previous periods. It's modeling the current value based on the "memory" of past random shocks.

MarcusSo an AR model is like predicting the weather based on the last few days' temperatures, while an MA model is like predicting it based on the errors in the last few days' forecasts?

EmmaThat’s a fantastic analogy. It captures the difference perfectly. And just like with AR, the 'q' in MA(q) tells you how many past shocks the model remembers.

MarcusAnd I assume ARMA is just putting the two together?

EmmaExactly. An ARMA(p,q) model combines both. It says that today's value depends on its own past values—the AR part—and on past shocks—the MA part. It’s a flexible model, but remember, it’s only suitable for a stationary series.

MarcusWhich brings us to the last one: ARIMA. The 'I' must be important here.

EmmaIt is. The 'I' stands for Integrated, and it brings us right back to our discussion on differencing. An ARIMA model is what we use for a non-stationary series. The model has three parameters: p, d, and q.

MarcusLet me guess. 'p' is the AR part, 'q' is the MA part... so 'd' must be the number of times we had to difference the series to make it stationary.

EmmaYou've got it. An ARIMA(p,d,q) model is simply an ARMA(p,q) model that's been applied to a series that was differenced 'd' times. So if we difference a stock price series once to get returns, and then model those returns with an ARMA model, we've actually built an ARIMA model for the original price series.

MarcusThat connects everything perfectly. So, that whole family of models helps us predict the conditional mean, or the expected value. But the initial problem also mentioned volatility clustering. What is that?

EmmaVolatility clustering is the observation that in financial markets, big changes tend to be followed by more big changes, and small, quiet periods are followed by more quiet periods. Think about a market crash. The day of the crash is hugely volatile, but the next few days or weeks are usually also very choppy. That's volatility clustering.

MarcusSo the variance isn't constant. It changes over time.

EmmaExactly. This is called conditional heteroskedasticity. And we need specific models to handle it. The two big ones for the exam are EWMA and GARCH.

MarcusLet's take them one at a time. What’s the idea behind the EWMA model?

EmmaEWMA stands for Exponentially Weighted Moving Average. It forecasts tomorrow's variance by taking a weighted average of today's variance and today's squared return, or shock. The key is the weighting, controlled by a parameter called lambda.

MarcusAnd a higher lambda means you're putting more weight on past variance, and less on the most recent shock?

EmmaCorrect. Typically lambda is high, like 0.94. This means the model is anchored by past volatility but will still react to new information. Its main advantage is that it’s simple and responds very quickly to recent events.

MarcusWhat's the catch? It sounds pretty useful.

EmmaThe major drawback is that EWMA has no concept of a long-term average volatility. It's purely adaptive. If you have a big shock, the volatility forecast spikes, and it will only come down if you have a series of very small shocks. It implies volatility can drift up or down forever. It doesn't have a gravitational pull back to some normal level.

MarcusAnd I'm guessing that gravitational pull is what GARCH provides.

EmmaPrecisely. GARCH stands for Generalized Autoregressive Conditional Heteroskedasticity. The most common version, GARCH(1,1), models tomorrow's variance as a function of three things.

MarcusOkay, let's break those three down.

EmmaFirst is a constant term, omega. This is the anchor. It's directly related to the long-run average variance that the model will always be pulled towards.

MarcusSo that's the mean reversion component right there. The gravitational pull.

EmmaYes. The second part is the previous day's shock, or squared return. This is the 'ARCH' term, and its coefficient, alpha, tells us how much we react to new information.

MarcusAnd the third piece?

EmmaThe third is yesterday's variance estimate itself. This is the 'GARCH' term, and its coefficient, beta, tells us how persistent our volatility is. A high beta means that high volatility will take a while to die down.

MarcusSo GARCH is like a rubber band. A shock can stretch it, but that constant term, omega, is always pulling it back to its long-run average size.

EmmaThat is the perfect analogy. And that mean reversion is the single most important difference between GARCH and EWMA. For the variance to be stable, the sum of alpha and beta must be less than one.

MarcusLet's apply this. The source gives an example: an analyst sees volatility clustering, but also notices that high volatility periods eventually calm down and return to a historical average. Which model fits, and why?

EmmaThe key phrase is "return to a historical average." That is the definition of mean reversion. Volatility clustering tells us we need either EWMA or GARCH. But the mean reversion part explicitly rules out EWMA and points directly to GARCH(1,1) as the more appropriate model.

MarcusSo the exam will give us these keywords to look for. What's the most common way candidates get tripped up by these concepts?

EmmaThe biggest trap is confusing the roles of EWMA and GARCH. They both capture volatility clustering, but only GARCH incorporates mean reversion. If a question gives any hint that volatility returns to a long-run average, the answer is GARCH. Don't overthink it.

MarcusSo if I'm heading into the exam, what’s the one-sentence takeaway for choosing between them?

EmmaIf the question mentions volatility reverting to a long-term average, think GARCH. If the focus is just on a simple, adaptive model that responds quickly to recent news without that pull-back, think EWMA.

MarcusThis kind of focused breakdown is exactly what we aim for in all our lessons. If you're finding this helpful and want to go deeper, VoraPrep is a full exam-prep app — lessons, practice questions, and an AI tutor in one place. You can get started and try it out at Vora Prep dot com.

EmmaIt really is about building that mental model, step by step. First, is the series stationary? If not, difference it. That decides between ARMA and ARIMA. Then, does the volatility cluster? If so, does it mean-revert? That decides between GARCH and EWMA.

MarcusA clear decision framework. Thanks, Emma. That was incredibly helpful.

EmmaMy pleasure. It’s all about understanding the 'why' behind each model. Keep practicing, focus on these core distinctions, and you’ll be in great shape. We’ll see you at the next Office Hours.