This page explains exactly how the numbers and the English text on this site are produced — what is measured, what is generated by a model, and what nobody has checked by hand. If you are going to rely on any of it, you should be able to see where it came from.
InsightIntoChina is built for readers outside mainland China who follow — or want to follow — China A-shares, Hong Kong listings and China-related US listings, and who cannot read Chinese-language broker research in the original. That is the whole audience. If you already read Chinese research directly, this site adds nothing you cannot get faster at the source.
Every brief starts from one Chinese-language research note or market commentary item, collected through third-party data channels. A model reads that note and writes a fixed set of English fields — the same fields for every brief, which is what makes two different notes comparable:
Ticker symbols are extracted from the note, which is what allows a brief to be reached from the ticker page it affects. A note mentioning the same ticker twice counts once; a note whose tone cannot be mapped to one of the five sentiment labels still counts as coverage, but contributes nothing to any average.
The English fields above are written by a model from a Chinese source note and are labelled with a machine-assigned confidence level. None of them is checked by a human before publication. There is no editor, no analyst and no reviewing desk behind this site.
What is not model-generated is the market data layer — prices, returns, moving averages, cross-sectional ranks and the breadth index. Those are arithmetic over price series, recomputed by a scheduled job. A model never contributes a number to a price, a return, or a rank.
Raw daily bars are not copied onto this site. Only a daily snapshot is stored, and it keeps the unadjusted close together with the official cumulative adjustment factor where one exists.
Returns are then computed on the adjusted series, as (close₁ / close₀) × (factor₁ / factor₀) − 1. This matters more than it sounds: on a small sample we measured four of seven names distorted by more than 0.5 percentage points when dividends are ignored, the worst by 7.4 points. The 20-day moving average and the sparklines are computed on the same adjusted series, so the last point of a sparkline equals the latest raw close shown next to it.
Limit: no comparable official adjustment factor exists for Hong Kong and US listings here, so for those names a return is a raw price change that does not include dividends. The field is left empty rather than filled with a reconstructed factor.
Returns are measured over trading bars, not calendar days. A one-day move smaller than 0.05% either way is counted as unchanged, so that floating-point noise is not reported as a direction.
The board on the divergence page is the one number here that is ours rather than the broker’s. It measures a single thing: does the tone of the research disagree with the price?
What it is not: it is not a prediction, not a signal, and not a rating. It says the two inputs disagree. It does not say which of them is right.
Each trading day, every ticker in the universe is classified by its one-day move (up, down, or flat under the 0.05% tolerance), and whether its latest adjusted close sits above its own 20-day moving average. Those counts feed a single 0–100 number:
Weights are renormalised if a component cannot be computed, rather than substituting a value. Labels follow fixed bands: below 25 Extreme Fear, 25–45 Fear, 45–55 Neutral, 55–75 Greed, 75 and above Extreme Greed.
Six indices (CSI 300, CSI 500, SSE Composite, SZSE Component, Hang Seng, S&P 500) take part in the breadth statistics but never appear on the divergence board. A breadth reading needs the index alongside its constituents; a note-level sentiment score does not apply to an index.
Market data is rebuilt by a scheduled job targeting each trading day after the close. The as-of date and the last sync time are read from the database and shown next to the data itself, and the site flags the figures once they are more than seven days old rather than presenting stale prices as live. If the pipeline stops, you will see an older date — not a page pretending otherwise.
Briefs publish on their own schedule, following whatever the upstream channels publish. Coverage is uneven by design: heavily covered sectors produce many briefs, thinly covered ones produce few or none.
Basic funnel events — a page view, a paywall impression, a checkout start — are recorded against a random identifier that this site generates and stores in your browser. It is not linked to your account, your email, or any third party, and there is no cross-site tracking, no advertising pixel and no data broker involved. Details are in the privacy policy.
A brief is a translation and a summary of someone else’s argument. Following it is not a substitute for reading the original or making your own decision.
If a brief gets something wrong, or a figure on the divergence board looks impossible, tell us which one: [email protected]. Rights holders can request removal of a source note at [email protected].
The free preview of a brief includes what happened, why it matters, the bottom line, the tickers and the sentiment. PRO is $9.90 per month and removes the truncation on the four withheld sections across the entire archive — key numbers, hidden logic, risk analysis and the full translation.
PRO does not add new research, new coverage, or advice. It removes a cutoff.
Institutional research on China A-shares & HK stocks — Ticker Coverage · Macro Themes · Daily Digests · Divergence Board · Methodology · How to Cite · About Us · Legal & Privacy