Home / Market distributions
Taiwan market cross-sectional distributions
Decile structure across the whole market for P/E, P/B and cash dividend yield, frozen once per trading day. It exists for one purpose: to let any cross-sectional number you are holding find its place in the market.
Snapshot date 2026-09-17. 中文版 (Chinese)
The three yardsticks side by side
Snapshot date 2026-09-17. This table puts the key percentiles of all three metrics on one screen, so any cross-sectional number you are holding can be placed by order of magnitude first, then read in full on its own decile page. Sample sizes differ by row because blanks stay blank rather than being filled in, so the denominators are not meant to match.
| Metric | Valid sample | Minimum | 10th pct | Median | 90th pct | Maximum |
|---|---|---|---|---|---|---|
| P/E ratio | 1,542 | 0.89 | 8.98 | 17.99 | 68.89 | 2,048.00 |
| P/B ratio | 1,078 | 0.16 | 0.70 | 1.64 | 5.23 | 64.69 |
| dividend yield | 1,734 | 0.00% | 0.00% | 2.71% | 6.76% | 113.94% |
How the three yardsticks are computed
Each metric is sampled independently, not forced onto one shared universe
A metric includes only the tickers with a valid reading for that metric on the day; the three are not intersected. Intersecting them would turn every percentile into a percentile of "companies that happened to disclose all three", which is a population nobody is asking about. The cost is that the rows carry different denominators, so the valid sample size is printed on the table for you to judge representativeness.
What gets dropped is meaningless readings, not unflattering ones
P/E and P/B include only values above zero: at or below zero means a trailing-four-quarter loss or negative book value, which is not the cheapest bucket but the ratio losing its meaning. Zero yields, by contrast, are kept, because paying no dividend is a fact rather than a gap. The two rules point opposite ways for the same reason — so the percentiles describe the market as it actually is.
Linear interpolation, matching the default in common statistical packages
Percentiles are taken by linear interpolation over the sorted sample, so the figures can be recomputed independently from the same data. Any metric with fewer than fifty valid observations is not published at all: deciles of a small sample describe coincidence, not market structure.
Frozen once per trading day, never rewritten
Boundaries move with each new snapshot, but every day's version is kept separately. A later restatement produces a new day; it does not reach back and edit a day that has already been frozen.
Three terms to read this page
Where to look next
Frequently asked questions
What are the median P/E, P/B and dividend yield in the Taiwan market?
As of 2026-09-17: the median P/E is 17.99 across 1,542 valid observations, the median P/B 1.64 across 1,078, and the median cash dividend yield 2.71% across 1,734. The sample sizes differ by design — a loss-making company has no meaningful P/E, while a company that paid no dividend still has a yield observation of zero. Different criteria, different denominators.
Which metric should I use?
It depends on where the value of the business sits. Use P/E for companies with stable earnings; use P/B for asset-heavy industrials and financials, where most of the value is in plant, land and balance-sheet assets rather than current earnings; use dividend yield when the holding is there for cash return. Reading all three beats reading one, and the cases where the three disagree are usually the ones worth more of your time.
How is a cross-sectional percentile different from a time-series one?
The cross-section asks where a number ranks against the whole market on one day. The time series asks where it ranks against that ticker's own history. The same P/E can sit high in the cross-section and low in its own history at once — which says the company has long carried a premium rating and is currently cheap relative to itself. This page is the cross-section; the time-series percentiles live on the ticker pages and the valuation river chart.
Can these percentiles be used directly in a backtest?
Yes, and that is what they exist for. Every trading day's boundaries are frozen as their own point-in-time vintage, so the ones shown for 2026-09-17 are the ones that were genuinely computable that day and are never rewritten when a later restatement lands. Slicing a sample with retroactively corrected percentiles plants look-ahead bias straight into the model.
Why only three metrics?
A percentile yardstick is only worth publishing when the population is large enough and the definition is not open to interpretation. P/E, P/B and cash dividend yield each carry market-wide disclosure across more than a thousand tickers and leave no room for definitional drift. Other measures such as return on equity or revenue growth vary far more in coverage and in how they are stated; folding them into a market-wide percentile would produce a number that looks precise while its denominator is unclear. Multi-metric conditional filtering belongs on the screener page.
This is data and tooling, not investment advice. The table is a cross-section on one snapshot date and carries no judgement about whether a level is cheap: a low decile may correctly price a decline, and a high one may be fair for growth. To judge a single ticker you want it against its own history, which is what the valuation river page does.