119 A-share strategy write-ups, stored one per folder

Every article is a folder with the same five entries

The unit of the repository is a folder, and the folder layout is identical everywhere, which is what makes a corpus of this size navigable without any tooling. Each article directory holds content.md, the body text with the original structure preserved, including headings, lists and code blocks. Beside it sits a code folder for the code snippets, an assets directory with an images subdirectory for figures pulled from the article, downloads.md with notes about attachments, and meta.json holding metadata. Three of those five are conditional in the documentation itself, with qualifiers for images and attachments, so the layout is honest about not every article having pictures or downloads. Reading is direct rather than through an interface: you open articles/*/content.md. There is no index page to render, no viewer and no search, so finding the strategy you want means reading the catalogue first.

GitHub calls it Python, and there is no package here

The language GitHub reports for this repository is Python, and that classification comes from the code snippets sitting inside the article folders rather than from anything at the top level. The root holds five entries and not one of them is source: a .gitignore, a README, a CATALOG.md file, an articles directory and a catalog.json. There is no requirements file, no pyproject or setup script, no package manifest and no tests. That distinction matters for anyone who arrives from the language badge expecting a library. The Python here is quoted material inside prose, in the same way a tutorial repository quotes code, and the project describes itself as an offline research copy rather than a distribution. Its own framing is a research library for learning, and the accompanying note is explicit that it is for study and research and does not constitute investment advice. Fork behaviour matches that framing, with 286 forks against 526 stars, which is what a corpus people copy out of looks like.

The same 119 articles are catalogued in three places

Three separate indexes describe the collection, and they do not obviously derive from one another. The README carries its own numbered list running through the strategies, one line per article with a link to that article's content.md. CATALOG.md sits at the root as a separate markdown catalogue, and catalog.json holds the same thing in machine-readable form. Per-article metadata lives in a fourth place, inside each folder as meta.json. Nothing in the repository describes which of these is authoritative or how they are kept in sync, so if you script against catalog.json you are trusting a file whose update discipline is undocumented. The practical effect is that a tool built on this corpus has two index sources to reconcile before it can trust its own listing, and the README's inline list is the one a human reader sees first. The count itself is stated once, as 119 articles, and that number is the one to check the indexes against.

The titles cluster into a few families, and momentum dominates

Reading the catalogue is more informative than it first appears, because the titles state each strategy's mechanism. Multi-factor work is the largest cluster: factor screening, composite scoring, factor selection with momentum tracking, dynamic factor selection, smart dynamic factor rotation and dynamic rebalancing all appear as separate articles. Momentum and rotation are the second: multi-asset momentum rotation, all-momentum screening, momentum timing, momentum with dividend yield, momentum timing ETF rotation, and a stock and futures blend mixing momentum with reversal. Timing shows up on its own through multi-dimensional momentum timing and an economic cycle plus moving-average regular investment plan. Then the outliers: a short-term low-open limit-up arbitrage, an intraday dynamic screen with pullback buying, a small-cap selection strategy, a high dividend low-volatility approach, and a dynamic mean reversion. Two articles reach further than the rest, one pairing multi-factor machine learning stock selection with Fisher Transform timing and one using deep reinforcement learning for momentum.

Attachments are netdisk links, and the repository warns about them

Where the articles carry more than text, the extra material arrives through a channel that has an expiry date nobody controls. content.md is documented as including complete netdisk links where the original post had them, downloads.md carries notes about the attachments, and the project states plainly that the validity of those attachments depends on the original share. That is the honest version of a mirror: the text is preserved and durable, the binaries are borrowed. Two things follow for anyone building on this. A citation to an article will keep resolving while a link to its dataset will not, so archive anything you intend to keep. And an article whose strategy depends on files you cannot fetch is still worth reading as a description of the method, just not as something you can reproduce tonight. The formatting note in the same README says the body text is laid out as close to the original structure as possible, with headings, lists and code blocks, which is the right priority for a mirror to take.

The project names its own staleness risk

Two disclaimers appear in the README and both are load-bearing rather than decorative. The first is that the material is for study and research and does not constitute investment advice. The second is that some strategies may already be outdated. Given what the catalogue contains, that warning is doing real work. Articles built on limit-up arbitrage, intraday screening and pullback buying encode assumptions about trading rules, price limits and settlement behaviour that change, and an article that never states its data vintage gives a reader no way to tell a dated mechanism from a current one. The same applies to the machine learning and reinforcement learning articles, whose results depend on a period, a universe and a feature set that the titles do not record. The last commit landed on 2026-07-14, the repository is not archived, and there are no published releases, so nothing here is versioned in the sense you could pin.

Editorial conclusion

This repository is worth an afternoon of reading if you want to see how Chinese retail quant writing approaches factor selection, momentum rotation and rebalancing cadence, because the value is in the breadth of examples rather than in any one strategy being usable. It is not a codebase: there is nothing to install, no dependency file and no entry point, so treat the code folders as reading material rather than as something you can run against live data. Three things to weigh before you lean on it. That the corpus is an offline copy rather than a maintained distribution, so upstream changes are not tracked. That the attachments behind several articles are netdisk links whose validity depends on the original share, and the repository says so. And that the project itself states it is for study and research only and that some strategies may be outdated, which matters most for the intraday and limit-up articles where the market rules they assume have moved on.

Frequently asked questions

Official sources

1. Issues
2. README
3. WAYLON/ashare-quant-strategies on GitHub