Why log loss penalizes a wrong 99% forecast so heavily
Log loss scores the probability assigned to the actual outcome. If Yes has probability p, the loss is −ln(p) when Yes occurs and −ln(1−p) when No occurs.
Full analysis selected. Choose Quick brief to read the key points first.
WFN QUICK BRIEF · UNDER 1 MIN
The short version
MODE 2 OF 2
What the primary record establishes
Log loss scores the probability assigned to the actual outcome. If Yes has probability p, the loss is −ln(p) when Yes occurs and −ln(1−p) when No occurs. Lower average loss indicates a better score. Gneiting and Raftery on proper scoring rules ↗
What the distinction changes
Suppose seven of ten outcomes are Yes. Two models always predict Yes, one at 70% and one at 99%. Both pick the winning side seven times, while their average log losses are about 0.611 and 1.389 respectively.
What to check before drawing a conclusion
Evaluate predictions captured before resolution at the same cutoff. An incorrect 0% or 100% forecast gives infinite theoretical loss. If probabilities are clipped for computation, define the limits consistently in advance. The score evaluates forecasts, while trading results additionally depend on prices and execution.
Quick Brief is a reading aid. The full article and source trail stay on this page.
WFN STORY COMPASS
Choose your reading angle
Switch between the reported record, market context and open questions. Status labels distinguish claims from findings.
All context layers are visible. Choose a lens to jump to that part of the story.
Original WFN vector illustrationWFN · 6 November 2026
Editorial noteSource-based explanation with original WFN analysis. Historical records retain their dates; numerical examples are illustrative. The original vector illustration is conceptual and is not documentary evidence.
WFN REALITY LEDGER
What is on the record?
Primary-source explainer
Primary-source explainer
Primary record / reference periodGneiting and Raftery on proper scoring rules
The linked source supports the documented record; WFN interpretation and illustrative examples are separate.
01 · VERIFIED RECORD
What the primary record establishes
Log loss scores the probability assigned to the actual outcome. If Yes has probability p, the loss is −ln(p) when Yes occurs and −ln(1−p) when No occurs. Lower average loss indicates a better score. Gneiting and Raftery on proper scoring rules ↗
02 · WFN EDITORIAL ANALYSIS
What the distinction changes
Suppose seven of ten outcomes are Yes. Two models always predict Yes, one at 70% and one at 99%. Both pick the winning side seven times, while their average log losses are about 0.611 and 1.389 respectively.
03 · REVIEW POINTS
What to check before drawing a conclusion
Evaluate predictions captured before resolution at the same cutoff. An incorrect 0% or 100% forecast gives infinite theoretical loss. If probabilities are clipped for computation, define the limits consistently in advance. The score evaluates forecasts, while trading results additionally depend on prices and execution.
