Probabilistic Prediction

Probabilistic prediction returns a distribution over possible outputs instead of only a point prediction. It matters when confidence, calibration, uncertainty, abstention, ranking, or decision cost affects the downstream use.

For supervised learning, a probabilistic model estimates:

$$p_{\theta }(y\mid x)$$

where $x$ is the input, $y$ is the target, and $\theta$ are model parameters.

Point Prediction

A point prediction can be derived from the distribution. For classification:

$$\hat{y}=\arg \underset{y}{\max }{p}_{\theta }(y\mid x)$$

For regression with squared-error loss, the optimal point estimate is the conditional mean:

$$\hat{y}={\mathbb{E}}_{p_{\theta }(y\mid x)}[y]$$

For absolute-error loss, the conditional median is often the corresponding decision.

Predictive Distribution vs Score

Not every score is a probability. A model can output:

• Logits: unnormalized scores before softmax or sigmoid.
• Probabilities: normalized values that sum to one for mutually exclusive classes.
• Scores: arbitrary ranking values that may not be calibrated.
• Uncertainty estimates: variance, entropy, ensemble disagreement, or predictive intervals.

The evaluation metric should match the output type. Probability quality needs probability metrics and calibration, not only accuracy.

Training

For a conditional likelihood model, training often minimizes negative log-likelihood:

$${\mathcal{L}}_{\mathrm{NLL}}(\theta )=-\frac{1}{n}\sum _{i=1}^n\log p_{\theta }(y_i\mid x_i)$$

This trains the full predictive distribution, while a downstream decision rule may later turn the distribution into a class, ranking, thresholded action, or abstention decision.

Checks

• Is the output a calibrated probability, a score, a logit, or a hard label?
• Does the loss train the quantity needed by the downstream decision?
• Is the probability evaluated on held-out data under a fixed protocol?
• Are probability metrics separated from thresholded classification metrics?
• Is uncertainty used for a concrete decision such as abstention, triage, or active learning?

Related

• Decision rule
• Classification
• Regression
• Negative log-likelihood
• Cross-entropy loss
• Probability metrics
• Proper scoring rule
• Brier score
• Calibration
• Uncertainty estimation
• Conformal prediction
• Threshold selection
• Softmax