Time Series

Most analytical errors involving historical data are not arithmetic mistakes. They are the result of using information that was not available at the time being analysed. The time series layer exists to make that failure mode structurally difficult.

Point-in-time semantics#

Every series is stored bitemporally: each observation carries both the period it describes and the timestamp at which it became known. A query therefore has two dates, and conflating them is the single most common source of backtests that cannot be reproduced in live conditions.

Revisions#

Reported figures change after publication, sometimes substantially and sometimes years later. A series that shows only current values is a reconstruction of the past that never existed. The revision history is retained in full, and the gap between an original print and its final value is itself queryable — it is frequently more informative than either number alone.