Why Machine Learning Predicts Volatility, Not Price Tops
The same model that fails completely at calling the top can be genuinely good at telling you how wide the range will be. The difference is not effort — it is the structure of the data.
Two questions with very different answerability
'Where will the price be?' and 'how much will the price move?' feel like variations of one problem. Statistically they are not remotely alike.
Price direction is close to unforecastable over short horizons, because any reliably exploitable pattern gets arbitraged away by the participants who find it. Volatility is meaningfully forecastable, and that difference is structural rather than a matter of model sophistication.
The property that makes volatility predictable
Returns show very little autocorrelation: knowing today's direction tells you almost nothing about tomorrow's. But the magnitude of returns is strongly autocorrelated. Large moves cluster with large moves; calm follows calm.
This is volatility clustering, and it is one of the most durable regularities in financial data — present across asset classes, countries and decades. It gives a model genuine signal to learn from, which directional prediction simply does not offer.
Why arbitrage removes one and not the other
If a model reliably predicted direction, acting on it would move prices until the edge vanished. Directional predictability is self-destroying, which is the practical content of market efficiency.
Volatility forecasts do not self-destruct in the same way. Knowing next week will be turbulent does not tell you which way to trade, so acting on it does not eliminate the pattern. Volatility can be traded through options, but the forecast remains useful because it is a statement about dispersion, not about direction.
What this should change about how you read models
Treat any claim of directional prediction with heavy scepticism, particularly if it is backtested. Treat volatility and range estimates as the genuinely useful output of quantitative modelling.
This is why a well-built probability tool answers 'what is the range and how confident are we?' rather than 'what is the target?'. The first question has a defensible statistical answer. The second mostly does not, and a model that claims otherwise is usually fitting noise.
The honest limits on the volatility side too
Volatility forecasting is better, not solved. Models trained on calm regimes underestimate turbulent ones, they cannot anticipate scheduled catalysts unless explicitly told about them, and they systematically understate genuinely extreme outcomes because the training data contains too few of them.
See a deliberately zero-drift, volatility-only model — it estimates ranges and confidence, and never claims a direction.
Explore the live probability cone →Quick answers
Why can models forecast volatility but not price direction?
Return magnitudes are strongly autocorrelated — volatility clusters — which gives models real signal. Return direction shows almost no autocorrelation, and any exploitable pattern is arbitraged away.
Why doesn't arbitrage eliminate volatility predictability?
Because knowing that a period will be turbulent does not tell you which way to trade. Acting on a volatility forecast does not remove the underlying clustering pattern.
What should I expect a good quantitative model to output?
A range and a confidence level, not a price target. Range estimation has a defensible statistical basis; short-horizon directional prediction largely does not.