Back to the log

Maintaining Evidence-Driven Forecast Revisions: Worked Examples‌‌​⁠‌​‌⁠‌​⁠​​⁠​‌​​​‌‍‍⁠‌‌‍‍⁠‌⁠‌⁠⁠‌⁠‍​⁠​⁠‌‍‌​‍​​‌‍‍‌⁠‍⁠⁠​‌​⁠⁠⁠⁠⁠​

All numeric cases below are synthetic teaching examples, not empirical research findings.

Example: honest revision

Prior 40%, then verified information arrives with LR=1.5. Prior odds=2/3; posterior odds=1; new probability=50%. Append the new record and preserve the 40% forecast.‌‌​⁠‌​‌⁠‌​⁠​​⁠​‌​​​‌‍‍⁠‌‌‍‍⁠‌⁠‌⁠⁠‌⁠‍​⁠​⁠‌‍‌​‍​​‌‍‍‌⁠‍⁠⁠​‌​⁠⁠⁠⁠⁠​

Example: decisive evidence

A previously uncertain event is officially completed and the resolution rule is satisfied. Resolve it. Do not move from 40% to 42% because “superforecasters make small updates.”

Example: unchanged review‌‌​⁠‌​‌⁠‌​⁠​​⁠​‌​​​‌‍‍⁠‌‌‍‍⁠‌⁠‌⁠⁠‌⁠‍​⁠​⁠‌‍‌​‍​​‌‍‍‌⁠‍⁠⁠​‌​⁠⁠⁠⁠⁠​

The only new story repeats last week’s announcement. Log that it adds no new evidentiary origin and retain the probability.

Counterexample

Changing a 30% forecast to 90% after the result and presenting it as the original forecast destroys prospective scoring.