Multi-Horizon Joint Forecasting

A team needs volatility forecasts at 1, 5, 10, and 20 days ahead, all for the same underlying risk process. The obvious approach is to train four completely separate models, one per horizon. The problem: those four models can end up telling inconsistent stories — the 5-day forecast implying a calmer path than the 10-day forecast bridges to, simply because each model was fit independently with no constraint that they describe the same evolving reality. Multi-horizon joint forecasting instead trains a single model that outputs predictions for all horizons at once, sharing an underlying representation of the process across horizons rather than treating each horizon as an unrelated problem.

Why sharing structure helps

Separate single-horizon models ("direct" forecasting, one model per horizon) each have to relearn, from scratch, the same underlying dynamics of the series — trend, seasonality, mean-reversion speed — using only the data available for that particular horizon's training examples. A joint model instead learns one shared internal representation of "the state of the process now" and produces a full path of predictions from it:

$$(\hat{y}_{t+1}, \hat{y}_{t+2}, \ldots, \hat{y}_{t+H}) = f(x_t)$$

In plain English: one model, one shared understanding of the current state, many outputs read off simultaneously — as opposed to $H$ separate models each guessing the same "current state" independently and possibly disagreeing about what it is. This tends to pool statistical strength across horizons (a horizon with less data to itself, like a rarely-realized 20-day-ahead label, benefits from sharing structure learned mostly from the more plentiful 1-day-ahead examples) and produces forecasts across horizons that are mutually consistent by construction.

Worked example: consistent versus inconsistent horizon forecasts

Four independently trained volatility models forecast 1-, 5-, 10-, and 20-day-ahead volatility for the same day: 18%, 24%, 19%, 26% respectively — a path that rises, falls, then rises again, which is an implausible shape for a genuinely smooth-evolving volatility process and is more likely an artifact of the four models disagreeing about current conditions than a real forecast of choppy future dynamics. A joint model trained to output all four horizons from the same shared hidden state instead produces 18%, 21%, 23%, 26% — a smoothly increasing path that's both more plausible and easier for a downstream risk system to interpolate between (say, to estimate 7-day volatility) without the interpolation crossing contradictory model outputs. The joint model didn't just save the cost of training four models — it produced a forecast that's internally coherent in a way four independent fits structurally cannot guarantee.

What this means in practice

Multi-horizon joint forecasting is worth reaching for whenever several related horizons are needed simultaneously and internal consistency across them matters for downstream use — risk systems that interpolate between reported horizons, hedging decisions that need a coherent forward path rather than four disconnected snapshots, or any setting where training data for the longer horizons is comparatively scarce and benefits from being pooled with the more abundant short-horizon examples. It's less necessary when only a single horizon is genuinely needed, where a dedicated single-horizon model, tuned specifically for that one target, can still outperform a jointly trained generalist.