Quantile Forecasts and Asymmetric Loss

A forecast of "expected return: +0.3%" says nothing about whether the downside on a bad day is a manageable -1% or a portfolio-threatening -8%. Two stocks can have the exact same expected return and wildly different risk of a bad tail outcome, and a single-number forecast can't tell them apart. A quantile forecast fixes this by predicting not one number but several — the 5th percentile outcome, the 50th, the 95th — giving a picture of the whole distribution of what might happen, not just its center.

Why asymmetric loss changes the target

Ordinary regression minimizes squared error, which treats an overestimate and an equally-sized underestimate as equally bad. But a risk manager sizing a position to survive a bad day does not care symmetrically — underestimating the downside tail is far more costly than overestimating it. Quantile regression handles this directly by minimizing a loss function that penalizes over- and under-prediction differently, depending on which quantile is being targeted:

$$L_\tau(e) = e \times (\tau - \mathbb{1}[e < 0])$$

In words: for a target quantile $\tau$ (say, the 5th percentile, $\tau = 0.05$), this loss penalizes underestimating the outcome nineteen times more heavily than overestimating it — because at the 5th percentile, being told the risk was smaller than it really was is the dangerous kind of error, and the loss function is built to make the model care about that asymmetry.

Worked example

A researcher fits two models to the same return data: an ordinary regression predicting the mean, and a quantile regression targeting $\tau = 0.05$, the 5th-percentile outcome. On a stock with historical daily returns roughly normal with mean 0.02% and standard deviation 1.8%, the mean model naturally predicts close to 0.02% every day. The 5th-percentile quantile model, fit directly on the loss above rather than derived from a normal-distribution assumption, predicts -3.1% on calm days and -5.4% on days following high recent volatility — capturing that the tail risk itself moves around, something a single mean forecast never shows. A risk desk sizing positions to survive a plausible bad day uses the quantile forecast directly; using the mean forecast alone would give no information about the tail at all.

Adjust the mean and standard deviation above — a quantile forecast is choosing a specific point along a curve like this one (say, the far left tail) and predicting how that point itself moves through time, rather than only tracking where the peak sits.

What this means in practice

Value-at-risk and expected-shortfall figures used in risk management are quantile forecasts by another name. Options desks price tail risk using quantile-like reasoning routinely, because option payoffs are themselves asymmetric — a hedge against a large downside move needs the tail forecast, not the mean forecast, to be sized correctly.