93 Papers, One Verdict: What Actually Forecasts Volatility

By the early 2000s, volatility forecasting had become a small industry. Hundreds of papers. Dozens of models. And a great deal of noise, because every paper used its own dataset, its own evaluation criterion, its own definition of what "actual volatility" even means, and reached its own confident conclusion.

Ser-Huang Poon and Clive Granger did the tedious, valuable thing. They read 93 of these papers, laid their findings side by side, and tried to work out what the accumulated evidence actually says.

The result is one of the most-cited survey papers in financial econometrics, and it is genuinely useful, both for its verdict on what works and for its diagnosis of why so many of the comparisons it reviews were not worth much.

The problem: nobody could agree because nobody was asking the same question

Poon and Granger's most important contribution is arguably not their ranking of models but their catalogue of the ways volatility forecasting studies go wrong. Before you can say which model is best, you have to fix a pile of decisions that different authors were making differently and often silently.

• What is "actual" volatility? You cannot observe it, so you need a proxy. Some papers used squared daily returns, which are appallingly noisy. Some used the daily high-low range. Some, by then, used realized volatility from intraday data. Different proxies produce different rankings of the same models.
• What counts as a good forecast? Root mean squared error? Mean absolute error? Some R-squared? The choice matters enormously, and the answer can flip when you change it. Poon and Granger flag this clearly, and Andrew Patton later proved just how bad the problem is.
• What horizon? A model that forecasts tomorrow well may be useless at forecasting the next three months, and vice versa.
• What about extreme values? A single crash day can dominate an error metric entirely, so the ranking ends up being decided by how well models handled one day in 1987.

None of this is glamorous, but it is why the literature was so confused. Many "model A beats model B" findings were artefacts of these choices.

The key idea via analogy: two ways to guess tomorrow's weather

Strip away the model zoo and the literature is really comparing two philosophically different sources of information.

Source one: look at history. Study how the weather has behaved, spot the patterns, note that stormy days cluster, and extrapolate. This is the entire GARCH and time series family, plus simple things like moving averages of past volatility. You are inferring the future from the past.

Source two: ask the people who are betting on it. Option prices contain a forecast. When a trader pays a high price for an option, they are saying they expect big moves. Invert an option price through a pricing model and you extract implied volatility, which is the market's own collective forecast of how volatile things will be between now and expiry. You are not inferring, you are reading a prediction market.

The central empirical question of the review is: which source wins?

The verdict

Poon and Granger's overall reading of the evidence is that implied volatility, taken from option prices, generally provides better forecasts than models built purely from historical returns.

This should not be shocking, and yet it is a genuinely important result. Implied volatility has an unfair advantage: it contains everything the historical models know, since traders can see the past too, plus information the historical models cannot possibly have. It knows there is an earnings announcement next Tuesday. It knows a central bank meeting is scheduled. It knows the election is in three weeks. A GARCH model, staring only at past returns, is structurally incapable of seeing a scheduled event that has not happened yet.

Their findings among the historical models are more nuanced, and importantly, less clear-cut than partisans of any single model would like. The various approaches, GARCH variants, stochastic volatility models, simple historical averages, tend to perform within a fairly narrow band of each other, and the ranking shifts with the market, the horizon, and the evaluation criterion. There is no universal champion. This is a genuinely useful thing to know, and it anticipates the conclusion Hansen and Lunde reached with a formal horse race two years later.

Why it mattered

• It organised a chaotic field. Anyone starting work on volatility forecasting reads this paper first. It is the map.
• It made the case for option-implied information. The finding that market prices beat statistical models pushed practitioners toward using implied volatility as an input, and researchers toward models that combine both sources rather than choosing between them.
• It exposed the evaluation problem. By insisting that the proxy and the loss function are first-order issues rather than technicalities, it set up the methodological cleanup that Hansen, Lunde and Patton would carry out over the following decade.
• It cooled the model arms race. The message that most reasonable models perform within spitting distance of each other was a useful corrective to a literature that was accumulating acronyms faster than insight.

The honest limitations

• It is a review, and reviews inherit the flaws of what they review. If the 93 underlying studies used bad proxies and non-robust loss functions, and many did, then the conclusions drawn from them are correspondingly shaky.
• Publication bias runs through everything. Papers reporting that a new model beat the benchmark get published. Papers reporting that it did not, mostly do not. The surveyed literature is a biased sample.
• Comparing incomparable studies is intrinsically hard. Different assets, periods, frequencies and metrics. Some of the aggregation is closer to careful judgement than to formal meta-analysis.
• Implied volatility is not a free lunch. It typically forecasts risk-neutral expected volatility, which is systematically higher than what actually occurs, because it contains a risk premium. It is a biased forecast, even if it is an informative one. That bias is itself an object of study, and it is what the variance risk premium literature is about.
• It predates the realized volatility revolution's full flowering. Published in 2003, the survey catches the start of the high-frequency era but not the models, HAR above all, that would come to dominate afterwards.

The one-line takeaway

Poon and Granger read two decades of volatility forecasting research and reported two things worth remembering: the option market's implied volatility generally beats models built from past returns, largely because it knows about events that have not happened yet, and most of the sprawling model zoo performs so similarly that the choice of evaluation method often mattered more than the choice of model.