Continuous questions allow the solicitation of forecasts with a moving end date. For example, while a typical forecast might ask “Will X happen by Dec. 31?”, a continuous question would ask “Will X happen in the next Y months?” At any given moment, a new forecast refers to the next Y months, from the day the forecast is made.

This type of question is a great fit for any topic that isn’t tied to specific dates, such as a moving time window for risk monitoring (“Will China invade Taiwan in the next 6 months?”), and creates a clean dataset that avoids time-based attrition of likelihood of traditional questions. (That is, at the beginning of a traditionally dated question, the likelihood of the event occurring is much greater than a day before the question closes).

Continuous questions use the same Brier and Relative Brier scoring rules as all other scored forecasting questions. What makes them unique is how the forecast and scoring periods interact.

How are continuous questions formatted?

Continuous questions have three components: the forecast period, scoring period, and question length.

• Forecast Period: The forecast period is the period for which a user is making a forecast. For example, in the above example “Will Country X invade Country Y in the next 6 months?”, the forecast period is 6 months. The forecast period can be set to whatever length is most suitable for the topic at hand.
• Scoring Period: The scoring period determines how frequently the question will be scored. For example, a question author could choose to score the above question on a monthly basis. In this case, forecasters would receive a score each month based on the accuracy of their forecasts in each scoring period.
• Question Length: The question author chooses an end date for the question to stop being forecasted (even if the event it asks about has not come to pass).

Here's an example:

“Will Country X invade Country Y in the next 6 months?”

• Forecast period: 6 months
  
• Scoring period: 1 month
  
• Question length: 1 year
  
• Question launch date: January 1, 2022
  
• User’s first forecast: January 1, 2022
  
• First scoring date: August 1, 2022 (for forecasts made between Jan 1 and Jan 31)
  
• Second scoring date: September 1, 2022 (for forecasts made between Feb 1 and Feb 28)
  
• Third scoring date: October 1, 2022 (for forecasts made between Mar 1 and Mar 31)
• Final scoring date: August 1, 2023
  

How are continuous questions scored?

Scenario 1: The event does not happen at any point during the life of the question

Example:

Scenario 1 occurs if Country X does not invade Country Y between Jan 1, 2022 and June 30, 2023. On each scoring date, the question resolves “No” for the relevant scoring period, providing the user with a score. Forecasts are scored using Brier and Relative Briers scores, as normal.

As the event fails to occur, once every forecast period has completed, a forecaster that consistently forecasts from the beginning of the question until the end will receive 12 scores for this question.

Scenario 2: The event occurs within the first forecast period

If the event occurs during the first forecast period, all the scoring periods are combined into a single, longer scoring period. The question resolves to “Yes” and the user receives a single score.

For example:

Country X invades Country Y on May 15, 2022. The user receives a single score calculated from all their forecasts between January 1 and May 15. The user does not receive a score for each month leading up to May 15.

Scenario 3: The event occurs after the end of the first forecast period

If the event occurs after the first forecast period ends, the user will receive a score based on two resolutions: “No” for all forecasts for which the forecast period does not include the event, and “Yes” for forecasts for which the forecast period does include the event. As in Scenario 2, if there are multiple periods that all resolve to “Yes”, they are combined and receive one score.

Example A:

Country X invades Country Y on July 15, 2022. Users who began forecasting prior to January 15 will receive two scores. Forecasts made from January 1 up to and including those made on January 14 will resolve to a “no” as the invasion did not occur within 6 months of those dates. Forecasts made on January 15 and later will resolve to “yes”, as the invasion is within a 6 month window of those dates.

To summarize, the user receives a single score judged against two different outcomes: “No” from January 1-14, and “Yes” from January 15 onward.

Example B:

Country X invades Country Y on August 15, 2022. Users who began forecasting in January will have already received a score for January, with the question having been resolved to “No” for that scoring period. If they continued forecasting into February and beyond, they will also receive a score for the partial period of February 1-14 (resolving to “No”) and a score for forecasts made February 15 and later, resolving “Yes”.

To summarize, the user receives two scores over the course of the question:

Scoring period 1: Forecasts made in January, resolving to “No” (score received August 1)

Scoring period 2: Forecasts made between February 1-14, resolving to “No” and forecasts made between February 15 and August 15, resolving to “Yes” (score received August 15)

Scenario 4: The event becomes unable to occur at some point during the life of the question

In this scenario, something occurs that makes the event impossible. The question is resolved to “No” at the time this occurs, with all remaining scoring periods combined into one.

For example:

Country X and Country Y agree to and implement complete political unification on August 15, 2022. The question is immediately resolved to “No”. Users will have already received a score for January (if they forecasted). They now receive an additional single score for all forecasts made from February 1 through to August 15.

To summarize, the user receives two scores over the course of the question:

• Scoring period 1: Forecasts made in January, resolving to “No” (score received August 1)
  
• Scoring period 2: Forecasts made between February 1 and August 15, resolving to “No” (score received August 15).
  

Important Notes:

Forecasts Expire

Unlike other questions, forecasts on continuous questions expire after one scoring period. Because continuous questions that never resolve positively could theoretically last for years, a user who makes a forecast when the question first launches and forgets about it could receive dozens of scores from a single forecast.

To prevent this from happening, forecasts expire one scoring period after they are made. So make sure to return to the question and update your forecast!

When a forecast expires, the days remaining in the scoring period are treated the same way as the days that elapsed prior to their first forecast. (See how Relative Brier Scores are calculated).

Example A:

A user makes a forecast on January 15, and does not make any updates. Their forecast remains “live” up to and including February 15. The user thus has forecasts in two scoring periods: scoring period 1 (January 1-31) and scoring period 2 (February 1-28).

Example B:

A user makes a forecast on January 15, and updates their forecast on February 2, but makes no further updates. Their updated forecast remains “live” up to and including March 2, 2022. The user thus has forecasts in three scoring periods: scoring period 1 (January 1-31), scoring period 2 (February 1-28), and scoring period 3 (March 1-31).

Example C

Instead of a 1 month scoring period, the question author makes a 2 month scoring period.

A user makes a forecast on January 15, and does not make any updates. Their forecast remains “live” up to and including March 15. The user thus has forecasts in two scoring periods: scoring period 1 (January 1 - February 28) and scoring period 2 (March 1 - April 30).