Data · Methodology

Register the Forecast
Before You See the Answer

Write the three-month forecast down, then put it next to the actual value when it arrives. Misses stay on the page. Cells with too small a sample show a blank instead of a number. These are the pre-registration and scoring rules used on a small site that aggregates Seoul apartment transactions.

Before asking whether a forecast is right, ask whether a miss would be visible. If not, only the hits survive and the forecast becomes advertising. The three rules below exist for that one question.

Once registered, never edited

Each month a three-month change is registered for Seoul and its sub-regions as a median with an interval. The record keeps the origin month, the target month, the interval, and any later adjustment with the date it was added. When the target month arrives, the official index goes next to it and the only thing written is whether it fell inside. Misses stay. Editing a registration after seeing the answer deletes the comparison itself.

Scoring rules are committed before the first score

The ledger that scores whether a complex's price range contained the deals that followed starts with a rules document, effective only from its commit date. Snapshots on the first of each month. A cell is the same complex and floor area with at least three deals in the previous 183 days. Follow-up windows of 30, 60 and 90 days are all published; no window is chosen afterwards. Only deals past the reporting deadline count. A hit is inclusive at both bounds. The nominal rate and the refutation threshold are written in advance and not changed after results. Changes add a new version and keep the old values. The first number needs three snapshots, so it arrives in January 2027 at the earliest; until then the table says "collecting (n deals)".

Short cells stay blank

Only three public units exist: district totals, price bands, and sample-size bands. Any row with fewer than 100 scored deals shows no number. Cells with fewer than three deals enter no aggregate at all. A blank is not a failure mark; it is the answer "this sample cannot say". Rows that do show a number carry n and a binomial confidence interval, with a note that deals in the same complex or region are correlated, so the independent sample is smaller than n.

Why

A forecast is worth what its visible misses are worth. Write before the answer, rules before results, blank when short. Keep those three and a bad result can be published, and because it can be published it is worth reading.

Further reading

monthly-report-2026-10 (the first monthly aggregate of Seoul apartment transactions for October 2026 produced under these rules, free) · Questions go to Discussions; it is a public board, so the post and your GitHub handle are visible to everyone.