A useful volatility forecast can leave direction unresolved

A market can become predictably more dangerous while its next move remains directionally uncertain. That distinction gives trading-agent research a useful opening: forecast the distribution an existing mandate will encounter, then measure whether that information improves its execution. We should demand directional evidence when a system claims direction, while allowing a risk forecast to earn its own result.

The distinction has an unusually clear historical foundation. Engle's 1982 ARCH paper introduced disturbances with time-varying variance conditional on past information; its empirical application concerned United Kingdom inflation. Bollerslev's 1986 GARCH paper extended the variance equation to include lagged conditional variances, with another inflation-uncertainty application. These are primary sources for a modeling mechanism, rather than tests of present-day crypto trading agents.

Two forecasts can disagree without either being confused

Write a return as a conditional mean plus a scale multiplied by an innovation. The mean asks where the distribution is centered. The scale asks how widely outcomes spread around that center. With zero conditional mean and continuous innovations symmetric about zero, increasing scale leaves the probability of a positive return at one half. It increases the probability of exceeding a fixed absolute threshold. This is a mathematical construction, and a clean counterexample to treating every successful market forecast as a successful sign forecast.

Consider an illustrative pair of zero-mean normal return distributions, one narrow and one wide. A predictor that identifies the wide state may correctly warn that a fixed absolute return threshold is more likely to be exceeded at the forecast horizon. It still has no reason, from that information alone, to prefer a positive rather than negative return. The word “predictable” therefore needs an object: variance, sign, a tail event or an execution cost.

Real returns can have skewness, jumps and changing conditional means. A variance model alone leaves those features unresolved. Even a good variance estimate can produce poor tail probabilities if its assumed distribution understates extreme moves. We would assess the forecasted distribution and its realized coverage rather than infer safety from the presence of a familiar model name.

Preserve the information through the decision

Our research question is whether an agent can use an independently assessed uncertainty forecast without inventing a directional thesis to explain it. That requires separating three components: the market forecast, the mapping from forecast to permitted action and the result of carrying out that action.

A volatility forecast can be statistically useful while an agent ignores it. Conversely, an exposure cap can improve outcomes even when the volatility estimator contributes nothing beyond a simple recent-range baseline. Both possibilities are easy to hide in an end-to-end profit chart. A component comparison should preserve the original direction decisions, change only the permitted risk mapping and record when the forecast actually changes an order.

For the forecast component, compare against a constant-variance model and a time-valid recent-variance baseline on later windows. Evaluate a stated variance loss together with interval coverage, paying particular attention to stressed periods. For the mapping component, inspect whether the same predicted distribution produces consistent restrictions under the same owner mandate. For economics, use actual fill costs and the appropriate margin rules. Each comparison answers a different question.

The important failure case is a forecast that looks informative only after the order is written. Exposure and liquidation distance can predict an order's eventual risk because the order created that risk. Such a result can support a post-proposal review, while a pre-decision market claim requires information available before proposal generation.

We see conditional uncertainty as a substantial research target because it connects statistical modeling to concrete operating decisions. Its value depends on the mandate and the execution path: an owner with fixed exposure may use it only for explanation, while an authorized risk-sensitive policy may act on it. Reporting which path produced the result keeps a useful volatility forecast from being promoted into an unsupported promise about price direction.