How to pick a loss function

Match the loss to your output type and what "wrong" means for your product.

1. **Name the task** — regression, binary/multi classification, ranking, or generation

2. **Match output shape**

3. **Align with metric** — optimizing log-loss helps probability quality; business may still care about recall at 95% precision

4. **Baseline first** — train with default loss, read confusion matrix or residual plots

5. **Adjust only with evidence** — change loss when errors show a clear pattern (many large outliers → try MAE)

| You predict… | Start with… |

|--------------|-------------|

| House price | MSE |

| Spam / not spam | Binary cross-entropy |

| Image class (10 labels) | Categorical cross-entropy |

| LLM next token | Cross-entropy (handled by framework) |

A loss that does not match your deployment threshold can look great in training and fail in production.

**Try the lesson:** `loss-functions` in Lane D · [Glossary: loss functions](/reference/loss-functions)