1. Definition
- A quantile forecast predicts a specific quantile (percentile) of the future outcome distribution, instead of just the mean (point forecast).
- For quantile level $\alpha \in (0,1)$:
$\hat{q}_\alpha(t+h) \;\; \text{is the forecasted value such that} \;\; P(Y_{t+h} \leq \hat{q}_\alpha) = \alpha$
Example:
- 0.5 quantile = median forecast.
- 0.1 quantile = pessimistic lower bound.
- 0.9 quantile = optimistic upper bound.
2. Why Quantile Forecasts Matter
- Unlike point forecasts (single mean estimate), quantile forecasts give different possible scenarios.
- Provide uncertainty information without requiring the full distribution.
- Useful for prediction intervals:
- A 90% prediction interval = [0.05 quantile forecast, 0.95 quantile forecast].
3. Examples
Weather
- Deterministic: “Tomorrow = 25°C.”
- Quantile forecasts:
- 0.1 quantile = 22°C (10% chance temp ≤ 22).
- 0.5 quantile = 25°C (median).
- 0.9 quantile = 28°C (90% chance temp ≤ 28).
Retail Demand
- 0.1 quantile forecast = 450 units (conservative, safe stock level).
- 0.5 quantile forecast = 500 units (expected demand).
- 0.9 quantile forecast = 560 units (upper bound for safety stock).
4. Methods for Quantile Forecasting
- Quantile Regression: directly estimates conditional quantiles.
- Quantile Regression Forests: ensemble tree methods for quantiles.
- Gradient Boosting (LightGBM, XGBoost) with quantile loss.
- Deep learning models (e.g., DeepAR, TFT) that predict quantile forecasts.
5. Loss Function: Pinball Loss
Quantile forecasts are typically trained using pinball loss:
$L_\alpha(y, \hat{q}_\alpha) = \begin{cases} \alpha \cdot (y – \hat{q}_\alpha) & \text{if } y \geq \hat{q}_\alpha \\[6pt] (1-\alpha) \cdot (\hat{q}_\alpha – y) & \text{if } y < \hat{q}_\alpha \end{cases}$
- Asymmetric: penalizes underestimation and overestimation differently depending on quantile.
- Ensures the model targets the correct quantile.
6. Relation to Other Forecasts
- Point forecast = mean or median.
- Quantile forecast = specific percentiles of the distribution.
- Prediction interval = constructed from two quantile forecasts.
- Probabilistic forecast = entire distribution (quantile forecasts approximate it).
Summary:
A quantile forecast predicts a chosen percentile of the outcome distribution (e.g., 10th, 50th, 90th). It’s useful for risk-aware decision-making, constructing prediction intervals, and understanding uncertainty. Trained using pinball loss, quantile forecasts are widely used in retail, energy, finance, and weather forecasting.
