Original Research · Taiwan · Data Quality

The Seasonal Lie: what Taiwan's filing calendar does to your data

If you work with Taiwan quarterly financials, part of what looks like signal is filing convention. Here's the measurement — and the honest limits of what it proves.

Helmledgers · 27 October 2026

Taipei 101 at dusk above the city skyline, with an overlaid perspective grid showing a glowing map of Taiwan and a Q1 to Q4 quarterly matrix
Illustrative composite. The quarterly grid graphic is decorative; the measured filing data appears in the charts below.

57 quarterly bundles · 2013–2026

134,096,682 extracted facts

Companies stay in the dataset. The mix of filing documents changes with the quarter.

TWSE + TPExFiling-level provenanceQuarter-type effects

The anomaly that shouldn't be there

Early in a recent data build, a coverage table for Taiwan's banking sector produced a pattern that looked, at first glance, like missing data. In the first and third quarters of every year, the number of banks filing full consolidated statements dropped from the mid-twenties to around twenty. In the second and fourth quarters, it returned. The same sawtooth appeared in total filings across the entire market: roughly 1,500 to 1,900 documents in Q1 and Q3, versus 2,300 to 2,700 in Q2 and Q4.

The first instinct is a data gap. It isn't one. It's the filing calendar — and the difference matters, because the two problems have different fixes.

What the data shows

The numbers come from a complete extraction of Taiwan's public filing record: every TWSE and TPEx listed company, 2013 through 2026, 134,096,682 extracted facts across 57 quarterly bundles, verified against public figures to the unit. (The verification method matters to everything below, and we'll return to it.)

Two cuts make the point. First, filings per quarter across the full market — rising almost monotonically across the years, but oscillating within every year, with Q1/Q3 sitting roughly 25–30% below Q2/Q4:

Chart 1 — Filings per quarter, full market

2,9002,4752,0501,6251,20020132016201920222025
Q1Q2Q3Q4
Source: Helmledgers Taiwan extraction, frozen 26 October 2026. Complete quarters through 2025 plotted. The 2026 Q1/Q2 counts are described in the article; Q3/Q4 are incomplete and excluded.
YearQ1Q2Q3Q4
2,0131,5301,9651,5572,024
2,0141,5692,0371,6052,123
2,0151,6342,3271,6472,393
2,0161,6762,3441,6902,402
2,0171,7032,3541,7172,422
2,0181,7442,3761,7552,449
2,0191,7602,3911,7742,451
2,0201,7802,3931,7832,478
2,0211,7992,4451,8192,545
2,0221,8282,4871,8382,605
2,0231,8602,5451,8852,663
2,0241,9232,6131,9472,721
2,0251,9862,6792,0202,730
2026*2,0492,70752

*2026 Q3/Q4 are live partial quarters as of writing and are deliberately excluded from the chart. Methodology: filings with extracted facts per MOPS quarterly bundle. Source: Helmledgers Taiwan substrate (tw_facts.db), frozen 2026-10-26, SHA-256 prefix 396b7884. Extraction and verification method documented in the runbook referenced in Paul's draft; public link to follow.

Second, the banking cohort — the cleanest possible lens, because every member is regulated to file on the same schedule — oscillates between roughly 20 and 24 members in Q1/Q3 and 22 to 33 in Q2/Q4:

Chart 2 — Consolidated banking cohort per quarter

363126201520132016201920222025
Q1Q2Q3Q4
Source: Helmledgers Taiwan extraction, frozen 26 October 2026. Complete quarters through 2025 plotted. The 2026 Q1/Q2 counts are described in the article; Q3/Q4 are incomplete and excluded.
YearQ1Q2Q3Q4
201322302331
201423312432
201524332433
201624322430
201722292229
201821272128
201922272227
202022262225
202122252225
202222252225
202321242124
202421242124
202521242122
2026*2022——

Methodology: distinct companies with consolidated (cr) filings under the banking-industry taxonomy module, per MOPS bundle. Same source and fingerprint as Chart 1.

Two things are visible in Chart 2. The first is the article's subject: the seasonal square wave. Q1/Q3 values remain between 20 and 24 across the series; Q2/Q4 generally sit higher. This recurs through multiple taxonomy generations, the classic-XBRL to inline-XBRL transition, and the COVID years. The source table also records Q4 2025 and Q2 2026 at 22, so an absolute assertion that every Q2/Q4 has at least 24 would not be accurate.

The second is a story hiding in the level: the consolidated banking cohort declines from a 2015 peak of 33 to 22 by 2026, a reduction of one-third in the observed filing cohort. That is consistent with changes in the banking sector over time, but filing counts alone cannot establish how much resulted from mergers or consolidation.

What's actually going on — measured, not assumed

Taiwan's disclosure regime requires listed companies to report quarterly, and the research finds continuity in the covered company population. What changes is the population of documents — and the mechanism was measured across all 57 quarterly bundles, not merely inferred.

Three findings. First, standalone (individual) reports are predominantly semi-annual: roughly 450 filings per Q2/Q4 against 180 per Q1/Q3 across fourteen years. Second, the bill-finance equity report type exists only in Q2/Q4 in this extraction — zero exceptions. Third, even the consolidated tier is roughly 15% lighter in interim quarters (1,804 filings in Q1 2025 against 2,197 in Q2).

What does not change is content density: filings in Q1/Q3 carry as many facts per document as any other quarter — Q3 documents are the densest of the year. The interim packages are not thinner. There are fewer of them.

So the honest statement is precise: the companies are continuous; the documents are seasonal. Any analysis that treats the document population as constant across quarters is comparing different instruments at different times of year.

The demonstrated problem

Computing cross-sectional statistics across differently composed quarters can rank observations partly by filing-cycle position. A company whose March disclosures are represented by a different mix of document types still appears in the data, but the available observations may have changed. Z-scores across those mixed populations may conflate the change in the business with the change in what was filed or extracted.

That is a measurement artifact to test independently of any trading strategy. It exists whether or not anyone trades on it. The observed filing-count asymmetry establishes the potential for this problem; measuring its effect on a specific factor requires a factor-level test.

The open question — stated honestly

How much this moves live strategies is a question, and we do not pretend to have answered it here. A properly matched year-over-year comparison — same quarter type, same filing composition — is not automatically invalid, and we do not claim it is.

What the measurement says is narrower and firmer: any cross-quarter comparison that does not condition on filing type risks mixing two populations, and the burden of proof sits with the unconditioned version. If a factor's spread lives in Q2/Q4 and vanishes within quarter types, the calendar is a candidate explanation worth ruling out before the economics get the credit.

The practice

Three rules, none expensive:

1. Stratify before you score.

Never compute cross-sectional statistics on Taiwan quarters without a quarter-type indicator (interim vs. semi-annual). A factor that is rank-stable within quarter types and noisy across them is telling you it's measuring the calendar.

2. Verify the filing, not just the number.

A panel entry should be traceable to a specific filed document — a specific bundle, with a fingerprint. If your vendor can't show you the filing a number came from, you can't distinguish “the company didn't disclose it” from “the pipeline didn't extract it.” Taiwan's interim quarters force this distinction constantly.

3. Treat seasonality as metadata, not noise.

The filing calendar is the cleanest, most stable structure in the dataset. It should be a column and a documented property — not something every analyst re-discovers and half-handles.

The larger point

Taiwan is not an edge case. Every market has filing conventions that leak into data: semi-annual regimes in Europe, the 10-K/10-Q content asymmetry in the US. The general lesson: the shape of disclosure is itself information — and it is almost never in the vendor's schema.

A dataset that tells you what companies reported, without telling you how they were expected to report, is encoding the regulator's calendar into your measurements. The fix is not a better model. It's better plumbing: filing-level provenance, version-stamped extractions, and the discipline of measuring the instrument before trusting the readings.

The coverage table that started this article is now one of the first quality gates we run on any new market. If the seasonality isn't there, the extraction is wrong. The filing layer, it turns out, is the most honest layer in the whole stack. It always tells you what it is.

Appendix — report-type mix per quarter (selected years)

Document counts by report type: cr = consolidated, ir = standalone (individual), er = bill-finance equity reports.

YearQ1 cr/irQ2 cr/ir/erQ3 cr/irQ4 cr/ir/er
20131,381 / 1491,634 / 295 / 361,414 / 1431,675 / 315 / 34
20181,594 / 1501,948 / 400 / 281,606 / 1492,007 / 414 / 28
20231,688 / 1722,074 / 443 / 281,715 / 1692,170 / 464 / 29
20251,804 / 1822,197 / 454 / 281,831 / 1892,230 / 472 / 28

Full 57-quarter table available on request; same source and fingerprint as the charts.

Next in this series: the filing type that appears in no documentation, found by refusing to average 0.6% of the data into “other.”

Helmledgers reads public filings with forensic-accounting discipline — every figure traceable to its source. This article is a measurement of public data, not investment advice.

One SEC filing finding each week. Source-linked. No fluff.