Conditional Forecasts on State Variables

A value signal's unconditional average return across twenty years of history might look modest and unremarkable. But split that same history by the level of interest rates at the time, and the signal might show a strong, consistent edge in high-rate regimes and essentially nothing in low-rate ones. The unconditional average was quietly blending two very different worlds together. Conditioning a forecast on a state variable — some observable measure of the market's current condition — is how researchers surface that structure instead of averaging it away.

Building a conditional forecast

The simplest version interacts the signal with a state variable in a regression:

$$r_{t+1} = \alpha + \beta_1 \, s_t + \beta_2 \, (s_t \times z_t) + \varepsilon_t$$

In words: the forecast $s_t$ still predicts next period's return through $\beta_1$, but a second term lets that relationship change depending on the state variable $z_t$ — a measure like volatility level, the slope of the yield curve, or market-wide liquidity — captured through the interaction term's own coefficient $\beta_2$. When $\beta_2$ is statistically meaningful, it says the signal's effective strength itself depends on the state.

Worked example

A researcher tests a short-term reversal signal (buy names that fell sharply yesterday) and finds an unconditional average next-day return advantage of 4 basis points across 3,000 trading days — a small but real edge. Splitting the sample by realized market-wide volatility (using a state variable $z_t$: 1 if the prior day's index move exceeded 1.5%, else 0) reveals the average edge is 14 basis points on the high-volatility days (about 400 of the 3,000) and only 1.5 basis points on the rest. The unconditional 4-basis-point average was a weighted blend that understated the signal's true strength in the state where it actually matters, and overstated it everywhere else.

Picture two of these normal curves side by side with different means — one for the "high-volatility state" subsample, one for the rest. Conditioning a forecast is the act of recognizing the data really came from two different curves, not one.

What this means in practice

Conditional forecasting lets a desk trade a signal selectively — sizing it up when the state variable says conditions favor it, and sizing it down or turning it off otherwise — rather than running it at a constant size regardless of context. It also explains why a signal can look "broken" in live trading after backtesting well: if the backtest period happened to sit mostly in the favorable state and live trading moved into the unfavorable one, the signal didn't break, the state changed.