Why the Forecast Horizon Shouldn't Be Fixed

Most quantitative models pick a fixed forecast window — 21 days, 60 days, one quarter — and never revisit the choice. We show why that's wrong, derive a horizon-from-data rule, and explain when stable regimes earn longer forecasts while fast-rotation regimes earn shorter ones.

The conventional choice

Open almost any quant fund's documentation. You'll find a forecast window mentioned — "21-day VaR," "60-day backtest," "one-quarter expected return." The number is rarely justified beyond convention or convenience. Twenty-one days happens to be one trading month. Sixty days happens to fit nicely between earnings cycles. Three months happens to match quarterly reporting.

None of those reasons connect to the underlying mathematics of forecast accuracy. They connect to calendar artifacts. A model that forecasts 21 days because "that's the convention" carries no theoretical justification — and worse, it predicts the same window in a stable market as in a chaotic one.

The actual question a forecast model should answer is: how long is my calibrated state valid? Beyond that window, the model is no longer the right object — extrapolating past it produces predictions the model wasn't built to honour.

What "calibrated state" means

Every quant model carries an internal state — a covariance matrix, a factor decomposition, a learned set of weights, a fitted distribution. That state is calibrated from recent data and used to make forecasts. The state evolves slowly when the underlying market structure is stable, and quickly when it isn't.

The relevant measure of "stable structure" is how fast the dominant directions of the covariance matrix rotate. If today's principal component (call it the "growth factor") is essentially the same one as last week's, the state is stable. If the principal component has rotated by 30° in the past five days, the state is unstable and recalibrating against it will produce stale forecasts the moment they're computed.

Mathematically, the rotation rate of the principal eigen-subspace has a natural timescale — the Grassmannian half-life. We'll call it τ. It's an empirical measurement, not a chosen parameter: τ is what the data actually tells us about its own stability.

The derivation

Once you accept that the right forecast horizon depends on how long the calibrated state is valid, the derivation falls out almost immediately:

where k is a "trust factor" — how far into the future you're willing to extrapolate beyond one half-life. We use k = 2, meaning we forecast roughly two half-lives ahead. Past that, the eigenstructure has rotated enough that the calibration we used is no longer in force.

The floor Tmin and ceiling Tmax exist for practical reasons:

• T_min = 5 days. Even if τ collapses to half a day (extreme rotation), forecasting at sub-week intervals invites short-window noise and transaction-cost overhead that outweigh the marginal information gain.
• T_max = 60 days. Even if τ is measured at 100 days, no quant model's calibration genuinely survives a quarter of trading time. The ceiling acknowledges this honestly rather than letting a stale calibration extrapolate further than its empirical evidence supports.

What this looks like in practice

We observe three regimes regularly in live operation:

Stable regime. When the factor structure is rotating slowly — τ on the order of 30 days — the horizon binds to its ceiling at 60 days. The system reads, accurately, "I have a long-shelf-life calibration; extend forecasts to where my evidence supports them, then stop." The 60-day cap is binding rather than the calibration: we're choosing not to forecast further despite the math allowing it.

Typical regime. τ around 10-20 days. The horizon scales proportionally: 20-40 days. Forecasts adjust naturally as the market's own evidence shifts.

Fast-rotation regime. τ collapsing under 5 days — usually a regime transition or crisis onset. The horizon floors at 5 days. The system reads, accurately, "my calibration is barely holding; I can only honour very near-term forecasts, and I'll recompute more often." This is exactly when an adaptive horizon earns its keep — fixed-horizon models would be forecasting 21 days ahead with a state that's only valid for the next two.

Why this matters operationally

Three concrete consequences:

One. Risk numbers (Value at Risk, expected drawdown, tail metrics) inherit the horizon. A VaR computed at the wrong horizon is misleading in both directions: too long in a stable regime overstates risk; too short in a fast-rotation regime understates it. Adaptive horizons give VaR numbers that scale with the actual stability of the system being measured.

Two. Position sizing inherits the horizon. The Kelly fraction depends on expected return over a window; if the window is wrong, the fraction is wrong. We use the dynamic horizon directly in the position-sizing calculation, which means the system automatically scales back its appetite during regime instability — not because it sees danger, but because it sees a shorter-lived calibration.

Three. Backtest interpretation changes. A model that "outperformed at 21 days" might have done so only in stable regimes where the conventional choice happened to align with what the data wanted anyway. Reporting performance at adaptive horizons separates the model's signal from its window-choice luck.

The cost of getting this wrong

Fixed-horizon quant funds will dispute that any of this matters at scale. Their argument: a misspecified horizon is just one source of error among many; the noise from the wrong window averages out over thousands of forecasts.

That argument has empirical content we can check. In our observation — and we'll have more rigorous backtest data over the coming weeks — the misspecification is asymmetric. A horizon set too long in a fast-rotation regime doesn't just add noise; it systematically commits the model to predictions its calibration cannot honour. The errors are not zero-mean; they're directionally biased toward overconfidence during transitions, which is exactly the time when overconfidence is most expensive.

An adaptive horizon doesn't eliminate forecast error. It moves the system into a place where the error structure is the responsibility of the calibrated state itself, not of an arbitrary choice you made years ago and forgot to revisit.