Bad News Hurts More: Nelson's EGARCH

Engle's ARCH and Bollerslev's GARCH gave the world a way to forecast market jumpiness: big moves today mean a jumpy tomorrow. It worked, and it still works. But it carries a hidden assumption that anyone who has watched a market crash knows is false.

In standard GARCH, volatility responds to the size of a move, not its direction. A day where the market drops 5% and a day where it jumps 5% feed the model exactly the same number, because the model squares the return and the minus sign disappears. Real markets are not so even-handed. A 5% crash terrifies people. A 5% rally makes them relax. In 1991, Daniel Nelson wrote the paper that fixed this, along with a couple of other awkward things about GARCH, and called his model exponential GARCH, or EGARCH.

The problem: GARCH is blind to direction, and awkward to constrain

Nelson pointed out three specific complaints about the standard setup.

• It cannot see the difference between good news and bad news. Because GARCH feeds on squared returns, the sign is thrown away. Yet decades of stock data show that a fall in prices raises future volatility more than an equal-sized rise does. Traders call this the leverage effect, and GARCH rules it out by construction, not because the data said so.
• It needs ugly guardrails. Volatility is a variance, so it can never be negative. To guarantee that a GARCH model never spits out a negative variance, you have to force all of its parameters to be positive. Those constraints get in the way when you fit the model. Sometimes the data wants to say "this piece is slightly negative," and you have to tell it no.
• Talking about "persistence" was confusing. In GARCH, it was surprisingly hard to say cleanly how long a volatility shock lasts and what happens when the model sits close to the edge of stability.

The key idea via analogy: model the volume dial, not the volume

Here is the trick, and it is elegant.

Imagine volatility as the volume knob on a stereo. Volume can never be negative, so if you build a machine that pushes the knob around directly, you have to keep checking it never shoves the knob below zero. Annoying.

Nelson said: do not model the volume. Model the logarithm of the volume. The logarithm can be any number at all, positive or negative, and when you undo the logarithm at the end (that is what the "exponential" in the name means) you always get back a positive volume. The constraint enforces itself. No guardrails needed. Parameters are free to be positive or negative, and you can just let the data speak.

With that freedom in hand, Nelson added the second ingredient: he let the model respond to the sign of today's return as well as its size. In plain terms, his volatility update has two separate levers:

1. A magnitude lever: any large move, up or down, cranks tomorrow's volatility up.
2. A sign lever: a negative move gets an extra kick on top, so that bad news raises volatility by more than good news of the same size.

That second lever is the whole point. When you fit EGARCH to stock market data, the sign lever consistently comes out saying what practitioners already believed: falling markets are scarier than rising ones, and volatility reacts asymmetrically. GARCH could never have found that, because it was not allowed to look.

Why it mattered

• It made asymmetry a standard feature, not a footnote. After Nelson, "does your volatility model handle the leverage effect?" became a question every serious model had to answer. A whole family of asymmetric models followed, including the GJR variant published two years later.
• It priced options better. The fact that crashes raise volatility more than rallies is a big part of why out-of-the-money puts on stock indices are expensive relative to calls, the pattern known as the volatility skew. A volatility model that captures the asymmetry gets closer to the world options traders actually live in.
• It freed the estimation. Dropping the positivity constraints made these models easier to fit and let researchers explore richer specifications without fighting the optimizer.
• It sharpened the persistence conversation. Working in logarithms gave a cleaner language for describing how long a volatility shock echoes forward.

The honest limitations

• The exponential can explode. Because you exponentiate at the end, an unusually huge return can produce a forecast that is enormous, far more so than in a plain GARCH. On very fat-tailed data, EGARCH forecasts can spike in ways that look unreasonable.
• The theory is fussier. The mathematical conditions under which the model is well behaved are more delicate than for GARCH, and some of them were only worked out carefully long after the paper.
• Interpretation is less intuitive. In GARCH you can read the parameters as "how much of yesterday's shock carries over." In EGARCH everything lives in logarithms, so the coefficients do not map to plain-English quantities as directly.
• It does not always win. Asymmetry is real, but it does not automatically translate into better out-of-sample forecasts. Later horse races found that simple GARCH is a stubbornly hard benchmark to beat, especially for exchange rates, where the leverage effect is weak or absent because there is no obvious "down" direction for a currency pair.
• The author did not get to develop it. Nelson died in 1995, only a few years after this paper, which is part of why EGARCH's theory took a long time to be fully filled in by others.

The one-line takeaway

Nelson modelled the logarithm of volatility instead of volatility itself, which removed the need for awkward positivity constraints and, more importantly, let the model finally admit what every trader knows: a market falling 5% frightens people far more than a market rising 5%, and volatility responds accordingly.