The Term Structure of Forecasts

Ask the same model to predict tomorrow's return, next week's return, and next quarter's return, and you get three genuinely different forecasting problems, not one problem run at three settings. Plotting a forecast's information coefficient, confidence, or typical error against horizon length is analogous to plotting the yield curve against maturity — hence "term structure of forecasts."

At short horizons, forecasts are usually sharper in relative terms (a higher information coefficient) because a real short-lived signal — an order-flow imbalance, an earnings surprise not yet fully priced — still dominates. But short-horizon forecasts also decay fast and are noisier in absolute terms, since one day's return is mostly randomness. At long horizons, individual-day noise averages out and the forecast becomes smoother and more stable, but it also captures far less of the idiosyncratic, fast-moving information that made the short-horizon signal attractive in the first place — the correlation between forecast and realized quarterly return is typically much lower than between forecast and realized daily return, even though the quarterly forecast "feels" more confident.

A practical consequence is that different horizons often call for structurally different models rather than the same model applied further out: a short-horizon model tuned to microstructure and order flow, and a separate long-horizon model built on valuation or macro factors, combined rather than one model asked to do both jobs by simply extending its target window.