Bundesbank paper guides optimal seasonal adjustment for high-frequency data

Official statistics increasingly use higher-frequency time series, raising a key question: should seasonal adjustment be done at the original high frequency or after aggregating data to a lower frequency?

This Bundesbank discussion paper by Daniel Ollech and Martin Stefan offers a comprehensive framework for this decision, balancing statistical quality and practical utility.

Their methodology uses simulated and real-world time series, evaluating leading seasonal adjustment methods (DSA2, WSA, X-13, TRAMO-SEATS) across daily, weekly, monthly, and quarterly levels.

Diagnostic tools, including tests for residual seasonality, calendar effects, and revision size, assess series quality.

The paper highlights that seasonal patterns and calendar effects vary with temporal resolution.

Daily data captures intricate patterns but is noisier.

The core argument involves a trade-off between informational richness and statistical robustness, noting challenges like identifying calendar effects in quarterly data.

A practical decision tree guides temporal aggregation for seasonal adjustment.

It begins by defining the use case and eliminating unsuitable levels.

For a quarterly target proxy, quarterly aggregation is relevant only if the target series is significantly delayed.

If a target exists, maximize alignment using out-of-sample prediction accuracy.

Without a target or with other objectives, consider series properties.

Avoid quarterly aggregation for critical calendar effects, as their identification is unreliable.

For cross-seasonal patterns, daily or higher frequencies are preferred.

Finally, evaluate remaining resolutions with diagnostics, summarize via mean ranks, and select the lowest-ranked option.