Projection: the estimate for the current period
Every hatched estimate on Repleader comes from one published calculation, projection-1.0, in one module. The last bar of every chart is the current week or month: the part observed so far is drawn solid, and the rest of the period, as estimated here, is drawn as a grey hatched block above it.
The calculation
For a period (a week or a month) of a series:
- O, observed so far
- the count observed in the period up to the latest reading.
- f, elapsed share
- calendar days elapsed divided by days in the period, at least 0.05. For a lagging period, one the source delivered but not completely (the archive's newest months), f is the coverage ratio: observed divided by the expected value, at most 1.
- H, history
- the last 6 completed periods' calibrated values, H1 oldest to H6 newest.
- R, run rate
- O divided by f.
- T, pace
- the mean of H1 to H6.
- s, trend
- the mean of H4 to H6 minus the mean of H1 to H3, divided by 3: the change per period.
- P, projected pace
- T plus 3.5 times s, never below zero (3.5 periods is the distance from the middle of the history to the middle of the current period).
- w, weight
- w equals f: the more of the period is seen, the more the run rate counts.
- E, estimate
- the larger of O and w times R plus (1 minus w) times P. It is never below what was observed.
The tooltip on every current bar lists O, f, R, T, s, P, w and E for that bar. Rankings and leaderboards never use E; they use observed and calibrated counts only.
How well it works
Backtest of 2024-10 to 2026-09 (the last 24 completed months), written 2026-11-18T04:23Z: for each month, E was computed from the first third and the first two thirds of the month and compared with the month's final calibrated count. Source: filings_day_obs (ClickHouse raw docket store), scaled to the Repleader counter.
| Series | Seen | Periods | Median error | Within 25 percent |
|---|---|---|---|---|
| National | a third of the month | 24 | 5.9% | 100.0% |
| Each district, each month | a third of the month | 2248 | 10.4% | 84.7% |
| National | two thirds of the month | 24 | 5.1% | 100.0% |
| Each district, each month | two thirds of the month | 2248 | 7.4% | 93.3% |
The national target is met: at a third of the month, 100.0 percent of national months land within 25 percent of the final count (target 70 percent).
Per-district results are in the machine-readable file that the daily backtest writes.