Match the market forecast to the clock that can use it

A useful market forecast arrives while the decision it serves is still available. That creates a stronger design question than whether a model can predict “the market”: which target, at which horizon, can this runtime observe and act on? We use a target-and-horizon map to connect market research to the agent's actual clock. The map identifies candidate tasks; each cell still needs its own evidence.

Seconds favor questions close to execution, such as the cost of an arriving order or the probability that a quoted opportunity survives. Hours permit slower information assembly and ask about price, funding or risk over a defined holding interval. Days create room for event resolution and changing portfolio exposure. Months can connect valuation, cash flows and changing economic conditions. These are research orientations, rather than a law that assigns each signal to one exclusive horizon.

Horizon	A concrete target	Information that must exist at forecast time	Decision constraint
Seconds	Cost of a specified arriving order	Received quotes and the order specification	Feed, inference and transport delay
Hours	Probability of a defined return or risk event	Completed observations and timestamped releases	Fees, funding and an admissible holding period
Days	Outcome around a scheduled event	The announcement calendar and its current wording	Gaps, revisions and overlapping positions
Months	A distribution of cash flows or total returns	Contemporaneously available fundamentals	Capital commitment and few independent episodes

A clock changes the source of an apparent advantage

Cont, Kukanov and Stoikov studied same-interval order flow and price movement in U.S. stocks on a ten-second grid. That mechanism belongs near the execution end of the map. A slow assistant can explain the mechanism well while receiving the relevant state too late to use it. Explanation quality and executable information value can therefore diverge.

At the other end, Campbell and Shiller's 1988 study examined aggregate U.S. stock-market data from 1871 through 1986. Historical earnings averages helped forecast the present value of future real dividends. Its valuation setting supplies a different target and information structure from an intraday sign classifier. The result gives us a reason to define cash-flow questions carefully, rather than transplant a historical aggregate relation into a monthly crypto trade.

Between these endpoints, an LLM can gather public releases and connect them to a thesis over hours or days. The unresolved empirical question is whether that synthesis improves a time-valid forecast over simpler information summaries after its latency and cost. A persuasive narrative can remain useful for research organization while providing no measurable forecast increment.

Forecast horizon and action horizon can separate

Imagine an illustrative agent with a day-long event thesis and a seconds-long execution decision. It may preserve the thesis while delaying or changing an order because immediate execution conditions are poor. Conversely, an urgent owner-authorized reduction may proceed despite an unattractive short-term cost estimate. The longer forecast supplies a belief about the asset; the shorter forecast supplies a consequence of carrying out an action now.

A benchmark that forces both into one direction label loses this structure. We would retain a forecast record for each target and then record how the mandate resolved their competing implications. A correct short-horizon cost warning can coexist with an incorrect day-long thesis. A profitable position can coexist with a poor execution forecast. Their joint outcome does not identify which component deserves credit.

Longer horizons also change the quantity of evidence. Reissuing a three-month prediction every hour creates many records with heavily shared future outcomes. It provides frequent decision snapshots, but those snapshots do not create equally many independent market episodes. Calendar separation and outcome overlap must shape the uncertainty calculation.

Our practical design rule is to specify a forecast's latest useful arrival time alongside its resolution time. The former asks whether the runtime can still affect the action. The latter defines when the claim can be scored. That pair keeps a seconds-level opportunity from being sold through an hours-level interface, and a months-level thesis from being judged by tomorrow's price. It also helps researchers choose a tractable first target before spending on a larger model.