Fundamental Mean Reversion

Fundamental Mean Reversion

_{<a href="README.ko.md">한국어</a>}

Looks for Korean companies whose share price has fallen further than the harm to the business justifies — and refuses to buy any of them until the falling has visibly stopped.

In one paragraph

This manager watches Korea-listed companies for large falls, and then spends its time on the question the fall does not answer: did the business actually get worse by as much as the price says? A price can fall because a company earns less than it used to, or because something happened once and the market treated it as though it would keep happening. Only the second is worth buying, and telling them apart means reading the company's filings rather than its chart. When the answer is that the business is intact, it still waits — because a cheap share that is still falling is not yet an opportunity — and it buys only after a set of conditions, written down in advance, say the fall has stopped. It sizes what it buys by what it loses if it is wrong, sells back into the recovery in pieces, and gives every position a deadline it cannot quietly extend.

The methodology

Mean reversion, with the emphasis on the reason rather than the shape. The claim being tested is that the market sometimes takes a temporary shock and extrapolates it as permanent, and that a price which fell on that extrapolation returns toward its normal range when the shock turns out to have been temporary. Everything here is machinery for telling that case apart from the case where the market was simply right.

Words this page uses, in plain terms:

Adjusted price	a price history corrected for splits and dividends, so that a 2-for-1 split does not look like a 50% crash
Drawdown	how far the price is below its highest point over some period
RSI	a 0–100 number summarising how one-sided recent moves have been. Low usually means "sold hard recently"
200-day average	the average closing price over roughly the last ten months. A common shorthand for where the price has been living
Oversold	sold hard and quickly. It describes the selling; it says nothing about the company
Impairment	a one-time writedown of something the company owns. It lowers the reported profit without lowering the cash the business earns
Invalidation	the condition, written before buying, that would mean the reasoning was wrong
Staged entry	buying a planned position in pieces, each piece waiting for something specific

Six rules do the work.

1 — The chart chooses what to research; the business decides what to buy. A fall of 30% or more, with either a low RSI or a price well under the 200-day average, puts a name on the research list. That is all it does. No position in this methodology has ever rested on those numbers, and a run that offers them as a reason is describing the order it worked in.

2 — Take the fall apart before explaining it. A missed quarter, a whole-sector de-rating, an index exclusion, a regulatory headline and a share issue are five different causes with five different half-lives, and a big fall usually has more than one. The thesis names them, then names the test that separates a one-time charge from a business that now earns less — and the test is named before its answer is known.

3 — Oversold is never an entry. The conditions that open one are fixed in advance: the lowest price of the last 120 sessions must be at least 15 sessions old, the price must be at least 5% above it, and the RSI must have come back above 35. Three separate outcomes are kept apart — still making new lows, a base that has not finished, and the reading could not be taken — because they call for three different things, and only one of them is about the company.

4 — Say which kind of target it is. A recovery target is a historical price band, a moving average, or a valuation range built from normalised earning power — and the package derives the kind of claim from the basis rather than accepting a label. A bounce's plausibility is not an independent valuation result, and this is the one confusion that makes a technical guess sound like arithmetic. It also never assumes the old high is recovered.

5 — A stop price is not a fill. Position sizes come from what the position loses on the way to its invalidation, plus a haircut for the session that opens straight through it. On the Korean market the daily price limit is the mechanism rather than the protection: a limit-down session is one in which a stop is a wish, and a trading halt is one in which it is not even that.

6 — Every position has a deadline, and it cannot be dissolved. If the recovery has not happened by the date the thesis named, the position is re-judged in writing — ended, or restated with a new deadline and a reason. What it may not become is a long-term holding, which is the same position with the stop and the deadline removed.

What it deliberately does not do

• It does not average down on price alone. A staged addition needs the thesis and the stabilisation to still hold; a rung whose only trigger is a lower price is refused outright, because that is how a bounded position becomes an unbounded one.
• It does not treat a fall below the 200-day average as disqualifying. That is the ordinary condition of the thing it is looking for. What it does separate out is a pullback inside an uptrend — price above a still-rising average — which is a real trade belonging to a different method.
• It does not read a price series it cannot vouch for. An undeclared adjustment basis, an unadjusted series with a dividend in the window, or a history that steps by a factor are all refused as missing data. They are never read as a finding about the company.
• It does not run a screen and stop. A candidate is carried to a finished thesis or left with a stated reason and a condition for coming back to it.

How a run works

One run, one proposal. Nothing below places an order.

flowchart TB
    classDef reads fill:#1e2a44,stroke:#6f9bf0,color:#cfe0ff
    classDef judges fill:#2f2f38,stroke:#9aa0b4,color:#e8eaf2
    classDef proposes fill:#1b4332,stroke:#40916c,color:#d8f3dc
    classDef person fill:#5c4813,stroke:#f6a609,color:#ffe8b0

    WAKE["Aumos wakes it<br/>after the Korean close, or on a review it armed"]:::reads

    subgraph IN["What it reads"]
        direction TB
        BOOK["Your holdings, cash and<br/>proposals still awaiting approval"]:::reads
        LEDG["Its own candidate ledger:<br/>unfinished research and where the last sweep stopped"]:::reads
        ROSTER["The Korea-listed roster,<br/>from the Toss connection"]:::reads
        PRICE["Adjusted daily price history"]:::reads
        FILE["Company filings from OpenDART"]:::reads
    end

    RESUME["Resume: retry the ranges that failed,<br/>pick the unfinished names back up"]:::judges
    SWEEP["Sweep the roster through the gate,<br/>and say which lanes were actually open"]:::judges
    CUT["Cut to a shortlist small enough<br/>to finish — never ranked by depth of fall"]:::judges
    WRITE["Write the ledger and the cursor back,<br/>so the next run resumes instead of restarting"]:::proposes

    HELD{"Does the fund already<br/>hold this name?"}:::judges
    REVIEW["Review the thesis it was bought on:<br/>invalidation, target, deadline, business"]:::judges

    subgraph JUDGE["What it works out"]
        direction TB
        INTEG["Can this price history be trusted?<br/>adjustment basis, splits, completed bars"]:::judges
        GATE["Has it fallen far enough<br/>to be worth a day's research?"]:::judges
        WHY["Why did it fall — and is the<br/>damage one-time or permanent?"]:::judges
        TGT["What range does it return to,<br/>and what kind of claim is that?"]:::judges
        STAB["Has the falling stopped?<br/>three conditions, fixed in advance"]:::judges
        SIZE["How much, given what it loses<br/>if it is wrong — and gaps and halts"]:::judges
    end

    CANT["Missing data<br/>said as missing data"]:::proposes
    NOPE["Not a case for this method<br/>damaged business, or still falling"]:::proposes
    BUY["BUY<br/>one rung of a planned position"]:::proposes
    TRIM["Stage out<br/>the target was reached"]:::proposes
    REJUDGE["Re-judge it in writing<br/>invalidated, or out of time"]:::proposes
    ARM["Arms its own next review,<br/>its price levels and its deadline"]:::proposes

    subgraph HUMAN["Where a person decides"]
        direction TB
        MAND["Aumos judges it against your Mandate"]:::person
        YOU["You approve, or you do not"]:::person
        ORD["Only then does an order exist"]:::person
    end

    WAKE --> IN --> RESUME --> HELD
    HELD -- yes --> REVIEW
    REVIEW --> TRIM
    REVIEW --> REJUDGE
    HELD -- no --> SWEEP
    SWEEP --> CUT
    CUT --> INTEG
    INTEG -- "cannot be read" --> CANT
    INTEG -- ok --> GATE
    GATE --> WHY
    WHY -- "damage is permanent" --> NOPE
    WHY -- "business intact" --> TGT
    TGT --> STAB
    STAB -- "still falling, or unconfirmed" --> NOPE
    STAB -- confirmed --> SIZE
    SIZE --> BUY
    CANT --> ARM
    NOPE --> ARM
    BUY --> ARM
    TRIM --> ARM
    REJUDGE --> ARM
    ARM --> WRITE
    WRITE --> MAND --> YOU --> ORD

Legend — 🟦 what it reads · ⬜ what it works out on its own · 🟩 what it hands back · 🟧 where a person decides.

Cadence. It asks to run after the Korean close on weekdays, and it always reads completed daily bars — a partial session is refused rather than warned about, because every reading it takes is off the newest rows. Every decision it submits, including the ones concluding there is nothing to do, arms the next review, the price levels that matter, and the deadline. Your Aumos may refuse an arming, and the interval stored for an installed manager is whatever you confirm on the install screen.

What it needs

Market	Korea-listed individual shares, in won. No ETFs, no baskets, no other market
Connection	a Toss login linked to this fund. It reads three things there: the Korea-listed roster it sweeps, the adjusted daily bars for each name on it — paged, because the readings need at least 300 completed sessions — and the market calendar. A market-data-only login is sufficient; no brokerage account has to be attached
Data source	open-dart for company filings, joined to every name the price sweep turns up — the separation of a one-time charge from a genuine decline is made there and nowhere else. It reads them through Aumos's own source cache rather than keeping a copy of its own
What a key costs	an OpenDART key is free with automatic approval from Korea's Financial Supervisory Service portal
Settings	how much of the fund one thesis may risk, how long a thesis may wait, how many it carries at once, and how many candidates one pass researches. ⚠️ Each of these can only be made stricter: the entry conditions themselves are not settings, and a value looser than the built-in one is refused and reported
Memory	it keeps its staged-entry ledgers, its candidate ledger and the cursor its next sweep resumes from in its private folder. Prices, filings, your holdings, cash and fills are re-read from their owners on every run and are never copied there
Your approval	it proposes and never trades. Every buy, trim and exit is a proposal your Aumos judges against your Mandate and you approve or refuse

What it is bad at

• A market that keeps falling. This method buys things that have already fallen a long way, which is the population most likely to fall further. The stabilisation conditions delay every entry — including the good ones — and in a market that only bounces briefly they will produce entries that are then wrong.
• Being right slowly. A business can be intact and the price can decline to notice for longer than any deadline this method would accept. It will close positions that would have worked, and record that it ran out of time.
• Falls that are correct and look temporary. Structural decline arrives disguised as a one-off charge for two or three quarters. This is the failure it is most exposed to, and the one its damage test exists to slow down rather than eliminate.
• Survivorship, and it cannot be corrected here. The roster it sweeps is today's list. A company that was delisted is not on it — and companies whose price fell a very long way are exactly the population delisting selects from. Any pass rate it reports is a fact about today's survivors and never a historical base rate.
• Reading nothing, said plainly. A run whose universe was never declared, whose connection was down, or whose budget went on reviewing what you already hold reports discovery_not_run — not "no candidates". The two are opposite facts, and in prose they come out as the same sentence, which is why the distinction is arithmetic rather than a sentence. Expect to see it: it is a working run reporting honestly, not a broken one.
• The quiet one: it is least comfortable when it is working. Refusing to buy something that is still falling looks like timidity right up until the last leg down.

Do not install this if you want a manager that buys weakness quickly, that will keep averaging down while a thesis is losing, that reacts within the day, or that you expect to override when its own stabilisation conditions say to wait — an overridden condition is the one part of this method with no value at all.

Notes

**Maintainer-facing material is in ARCHITECTURE.md (ARCHITECTURE.md)** — what its deterministic core computes, which modules it was derived from, the fixture inventory, and the list of things only a running host can verify.

Provenance. This is a port of a private research harness, morethanmin/trading-harness, at the commit named in aumos.json, published with its author's permission. NOTICE.md (NOTICE.md) carries the copyright line and states what was deliberately left behind: no account data, no credentials, no order implementation, no personal ledger.

Thresholds are pre-registered and were not fitted. The 30% fall, the 15 sessions, the 5% reclaim, the RSI floor of 35 and the risk budget are fixed in the pull request that added this package, and the repository's own check pins them as literals so that moving one shows up as a diff in a file whose only subject is that they did not move. They were not tuned so that any historical case would come out as a buy. The methodology's reference case — a large-cap Korean internet name whose fall was widely argued to be excessive — classifies as mean reversion in the package's own fixtures and reaches WATCH, because the stabilisation evidence available at that instant was not there. That is the intended behaviour and it was not adjusted.

Withdrawn claim. The harness's author reported a favourable outcome on the trades this methodology was drawn from. That figure was never re-audited — no fill or exit ledger was recomputed for this port — so it is not an edge this package has been shown to have, and it appears nowhere in it. Aumos shows no returns for this package until it has measured some from actual runs.

Sectors: this package has no concept of one, and that is not the same as being safe under a sector limit. Nothing in this methodology reads what industry a company is in — the fall's causes, the damage test and the stabilisation evidence are all about one company — and it does not classify one. But "no sector concept, therefore no conflict" is a conclusion about the code and not about your account: the host does not enforce a Mandate's sector ceiling and this package was not receiving one, so under such a Mandate a correct-looking answer could take the account through a limit you had declared. It now reads sectorCap. Declared and checkable, it is one more ceiling on the position. Declared and uncheckable — the candidate, or any holding or open proposal in the book, carries no sector — no entry is opened and no stage fires, and the answer names the row it could not classify and records it as missing data. Undeclared, the axis simply does not apply. Reviewing, trimming, re-adjudicating and exiting a position you already hold are untouched.

Known limits. The storage of a proposal, the re-arming of a review and the link from a decision to a real fill are all the host's, and none of them is exercised by this repository's checks.

Duplicate-exposure attribution across several managers on one book was on that list and is not any more: untilled/aumos#789 verified it against a running kernel, and this package is now told both what every other manager has proposed and who each position belongs to. You may run this alongside another manager that can hold the same name, on one fund. What is still unmeasured anywhere is a live vendor session and a live broker fill.