Taiwan Ownership Flow — shown side by side, deliberately never summed
The word “flow” invites one question: “so is the smart money buying”. This page does not answer it and no field tries to. What is here is several independent series of shareholding-structure change, drawn from different statistical definitions and updated at different frequencies — some daily, some weekly. Adding them up would require weights with no basis and would first require spreading a weekly figure across days; both steps invent information. So we never sum; we place them side by side.
Every series carries its own definition
Each series is returned with what it measures, its unit, its update frequency, its data date and its coverage. Some count lots, some count accounts, some count percentage of shares — these are not convertible into one another and do not belong on one axis. Series routinely disagree, which is expected: they measure different things. We pick no representative, take no weighted average and publish no blended flow.
Different frequencies are not forced into alignment
A weekly series has one observation per week. Drawing it on a daily chart means either step-holding it or interpolating four days that were never observed; the latter produces smooth daily movement that does not exist. We therefore present each series at its own frequency, label that frequency in the response, and align only on dates where both actually have observations. No interpolation, no smoothing, no resampling — resampling would hand a weekly curve a resolution it never had.
Inflow is not a price forecast
A rise in one holder category's share necessarily means a fall in another's: the share count is fixed and all holders always sum to one hundred percent. “Inflow” therefore describes a redistribution of structure, not fresh buying appearing from nowhere. We claim no causal link between structural change and later prices, and we publish no signal, no price estimate, no buying-leaders ranking and no stock list.
Data state and reproducibility
Every series comes from official publications, gated at its own publication time so a backtest cannot see later revisions. Missing observations are marked undetermined with the period named, never zero-filled or carried forward. Output is deterministic for identical inputs, and the module introduces no formula, weight or threshold of its own, so any series can be recomputed from the same public record.