1. Definition

  • A continuous probabilistic forecast predicts a full probability distribution (continuous) for a future outcome, not just a point or a few quantiles.
  • Instead of outputting “next week’s demand = 500,” the model outputs something like:
    • $Y_{t+1} \sim \mathcal{N}(\mu=500, \sigma^2=30^2)$
  • That means: the forecast is a continuous distribution describing all possible future values and their likelihoods.

2. Contrast with Other Forecast Types

  • Point forecast: “500 units.”
  • Quantile forecast: “10th = 450, 50th = 500, 90th = 560.”
  • Continuous probabilistic forecast: “Next demand follows Normal(500, 30²), so I can compute any quantile or interval.”

3. Characteristics

  • Outputs a full PDF (probability density function) or CDF for the future outcome.
  • Allows computing:
    • Prediction intervals (e.g., 95% range).
    • Quantiles (e.g., 5th, 50th, 95th).
    • Event probabilities (e.g., $P(Y > 550)$).

4. Examples

Weather

  • Deterministic: “Tomorrow = 25°C.”
  • Continuous probabilistic forecast:
    • $Y \sim N(25, 2^2)$ → ~68% chance between 23–27°C, ~95% chance between 21–29°C.

Retail Demand

  • Forecast: $Y \sim \text{LogNormal}(\mu=6.2, \sigma=0.3)$.
  • You can compute:
    • Median demand = ~500 units.
    • 90th percentile = ~650 units.
    • $P(\text{demand > 700}) = 0.08$.

Finance

  • Returns forecast: $Y \sim \text{Student-}t(\nu=5, \mu=0.01, \sigma=0.02)$.
  • Allows risk managers to compute Value-at-Risk (VaR) and Expected Shortfall.

5. Methods for Continuous Probabilistic Forecasting

  • Parametric models: Assume a distribution family (Normal, Lognormal, Student-t, etc.).
    • ARIMA with Gaussian errors.
    • GARCH models (finance).
  • Bayesian models: Naturally produce posterior distributions.
  • Simulation/Ensembles: Generate sample paths → approximate distribution.
  • Deep Learning models:
    • DeepAR (Amazon): autoregressive RNN with likelihood outputs.
    • Temporal Fusion Transformer (TFT): outputs full predictive distributions.
    • Mixture Density Networks: combine neural nets + probabilistic output.

6. Evaluation Metrics

Because we’re dealing with distributions, not single values, evaluation uses strictly proper scoring rules:

  • CRPS (Continuous Ranked Probability Score) — generalization of Brier score.
  • Log Score (Negative Log-Likelihood) — rewards high probability density at the observed outcome.
  • Calibration tests — check if predicted distributions match reality.

7. Why It’s Powerful

  • Provides full uncertainty quantification.
  • Supports risk-aware decisions (inventory buffers, financial risk limits).
  • Flexible: from one forecast, you can extract point estimates, intervals, or event probabilities.

Summary:
A continuous probabilistic forecast predicts an entire probability distribution (e.g., Normal, Lognormal, Student-t) for future outcomes. Unlike point forecasts, it quantifies uncertainty fully, enabling computation of intervals, quantiles, and event probabilities. It’s evaluated with CRPS, log score, calibration, and is central in modern forecasting (weather, finance, retail, energy).