How Prolepsia makes a forecast.
This page explains how Prolepsia makes a forecast: how it researches a question, how it turns the research into one probability, and how a forecast is checked once the outcome is known.
What a forecast is
A Prolepsia forecast answers a question about the future that has a date it resolves on. A yes/no question gets a probability. A multiple choice question gets a probability for each option. A question about a number or a date gets a forecast range.
Every forecast comes with its reasons, the sources it read and the date it was made. Prolepsia returns a number rather than a paragraph, because a number can be checked after the fact and a confident paragraph cannot.
What the probability means
A probability is Prolepsia’s estimate, on the date shown next to it, of how likely the outcome is. It is not a promise, and it changes as the evidence does.
Read 70% as: of many questions given 70%, about 7 in 10 should come out yes if the forecasts are good. One question cannot show whether a probability was good. A record of many resolved questions can, which is what the track record is for.
Research, with sources
Prolepsia researches each question before it forecasts. It reads the public record, such as filings and registries, price series, official statistics and what the market already thinks. It also reads any sources you provide: uploaded files, notes, and remote sources connected over MCP.
Each source is labelled the same way: what it is, when it was published and how far it is trusted. Every forecast lists the sources it read, so you can open them and check.
While a forecast runs, its page in the app shows each stage, what the forecast is doing now and the sources it has read so far.
Several forecasts, combined into one
A forecast is not one model’s single answer. Several forecasters work on the question separately. Each one does its own research and gives its own forecast.
Prolepsia then combines them into one. For a yes/no question it combines their odds. For a multiple choice question it averages the probability each one gives to each option. For a number, a date or a count it averages their forecast distributions.
A last step reviews the separate forecasts, the combined figure and the reasons behind them, and gives the final forecast. It is told to stay close to the combined figure.
The run mode you pick sets how many forecasters take part and how much research each one does.
The report
After the forecast is made, Prolepsia writes a report that explains the reasoning and cites the sources. The report is written from the finished forecast, so the numbers do not change to fit the prose. You can read it short, medium or detailed.
Historical runs with a cutoff date
Prolepsia can run a forecast with a cutoff date, to see what it would have said on an earlier day. For such a run it turns off open web search and web page reading, because those cannot be limited to a date. It reads only sources that support a cutoff date, so the run does not read what was published after it.
How well this works depends on the source. Not every source keeps its figures as they stood on the day, and uploaded data can reach you later than the dates in its rows, so its publication date matters. Check a source’s limits before relying on a historical run.
Forecasting again when the evidence changes
A forecast is not made once and left. A question can be forecast again on a schedule you choose, or when a condition you set is met, such as a filing appearing or a number passing a threshold. Conditions are checked periodically, not the instant something happens.
A question can also rerun automatically when new information would meaningfully change the answer. Each rerun uses credits.
Each new forecast is added to the question’s history, and the earlier ones are kept. The report says what changed since the last forecast, and every public forecast page shows each forecast with its date.
How resolved forecasts are scored
Once a question’s outcome is established, its forecasts can be scored. Some questions cannot be scored, for example when no outcome is established or the question is annulled.
Prolepsia scores its public yes/no and multiple choice forecasts with the Brier score, which measures how far each probability was from what happened. Only forecasts made before the question resolved count. The track record lists every resolved public question with its outcome and score, and says what was left out and why. It is computed from the forecasts each time the page loads.