Catalyst Turnaround

Catalyst Turnaround

_{<a href="README.ko.md">한국어</a>}

Buys a Korea-listed company that broke, only when something specific and dated would repair it — and sells when that thing happens, or stops being coming.

In one paragraph

Catalyst Turnaround looks for Korean companies whose business or balance sheet has deteriorated and where a named, sourced, dated event would repair it: a receivable balance being recovered, a regulated price being normalised, a loss-making division being closed, a refinancing moving from needed to secured. It writes that event down as a record with a source, a window, and the exact number in a future filing that would prove it happened — and then it manages the position by whether that number arrives. If the event lands, it trims into the recovery. If it slips, it may hold once, on stated new evidence and a stated cost. If it slips again and again, or is cancelled, or the company's cash stops covering the wait, it goes to an exit review. It never records a rising share price as a catalyst that worked.

The methodology

The claim being tested is narrow and it is not «this is cheap». It is that a specific, observable event will move this specific company's cash flow, on a schedule, and the market is pricing something else. Everything below follows from having to be able to check that later.

The catalyst ledger

The centre of this manager is a record rather than a score. Every catalyst it registers carries: what kind of event it is, the primary document it came from, when that document was published, whether the date is confirmed or estimated, the window it is expected in, the indicator that will confirm it, and the conditions that would count as success and as failure — all written before the window opens.

While a position is held, that record moves through five states: scheduled → in progress → realised, delayed, or failed. Four rules make it a record rather than a narrative.

1 — A delay costs three things. New evidence, the return still expected from here, and the extra downside now being accepted. Recording the same threat again is not new evidence. After two extensions the question stops being when does it land and becomes does this capital beat the alternative, compared against a benchmark, explicitly.

2 — A delay and a cancellation are different records. One is a window that moved; the other is an event that will not happen. They lead to different decisions.

3 — Realised and failed are final. A record is not edited. A genuinely new catalyst is a new row with its own source, which leaves the old finding readable — and contrary evidence, once filed, cannot be quietly dropped on a later run.

4 — Success is a number in a document. Never the share price, in either direction. The price enters this methodology in exactly two places: the distance to the level that would mean the thesis is wrong, which sets the size, and the progress toward the target, which triggers the trim.

What it deliberately does not do

• It does not rank the market by how far things fell. The deepest fall is not preferred for being the deepest, and there is no single "down-ness" score anywhere in it.
• It does not use a low valuation multiple as a qualifying condition — and this is a decision rather than an oversight. A company recovering from near-zero earnings has an enormous multiple on the way in and a flattering one on the way out; screening on the first would exclude the whole case class. Nor is a momentum oscillator a gate. Price stabilisation is a secondary confirmation: it can make the manager wait, it can never make it buy.
• It does not treat a policy as a thesis. A government programme can be real, funded and announced and still reach a given company only through an assumption nobody wrote down. Without a traced path from the policy to this issuer's earnings, the name stays a research candidate.
• It does not net offsetting effects. In the kind of case this was written from, the receivable balance, the regulated price, the input cost and the overseas result all move at once, sometimes against each other. They are tracked as four separate channels, and an effect that is derived from another — the interest that stops accruing on a falling balance — is recorded and not counted twice.
• It does not confuse "we could not read it" with "this is wrong". A figure that could not be read is an absence, and an absence is never a refutation.

Words this page uses

Catalyst	a specific, dated, sourced event that would change this company's finances — not a theme and not a hope
Confirming indicator	the number, in a named part of a future filing, whose arrival proves the catalyst happened
Invalidation	two things: a price level, which sizes the position, and a business condition, which closes it
Runway	how many months the company's liquid assets cover its cash burn. It has to outlast its own catalyst
Refinancing coverage	cash plus committed refinancing, against debt falling due within a year. Below 1.0 the company needs money nobody has promised
Channel	one route by which the recovery shows up — a receivable balance, a regulated price, an input cost, an overseas result
Staged plan	an entry built in parts, each with its own size, its own date-or-price condition and its own expiry

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 book — holdings, cash,<br/>and every open proposal on this fund"]:::reads
        FILE["Filings and policy documents<br/>receivables, debt, cash flow, notices"]:::reads
        REG["Its own catalyst register<br/>windows, delay counts, contrary findings"]:::reads
        LED["Its own candidate ledger<br/>names in progress, open questions,<br/>the receipt it stopped at"]:::reads
        PX["Adjusted won price history"]:::reads
    end

    HELD{"Is this name<br/>already held?"}:::judges

    subgraph UPDATE["Held: update the ledger first"]
        direction TB
        U1["Re-read every open catalyst<br/>scheduled · in progress · realised · delayed · failed"]:::judges
        U2["Did a window arrive?<br/>adjudicate success, delay or failure"]:::judges
    end

    subgraph SWEEP["Not held: find something, and say what was looked at"]
        direction TB
        S1["Resume — retry the receipt ranges that failed<br/>and finish the research left open"]:::judges
        S2["Declare the universe and the filing window"]:::judges
        S3["Sweep new filings and policy notices<br/>forward from the last successful receipt"]:::judges
        S4["Shortlist at most three, finish at least one,<br/>and record which of the four this run was:<br/>candidates_produced · no_candidate_qualified ·<br/>discovery_not_run · discovery_incomplete"]:::judges
    end

    subgraph NEW["Not held: build the case"]
        direction TB
        N1["Is there a real recovery path,<br/>traced to this issuer's own earnings?"]:::judges
        N2["Register the catalyst<br/>source · window · confirmed or estimated · confirming indicator"]:::judges
        N3["Compare the recovery indicators<br/>between two points in time, by channel"]:::judges
        N4["Can it survive its own catalyst?<br/>runway and refinancing coverage"]:::judges
        N5["Three scenarios, the difference from<br/>consensus, and both invalidations"]:::judges
    end

    SIZE["Size from the loss to invalidation,<br/>capped by the whole-account limit"]:::judges

    subgraph OUT["What it hands back — one of these"]
        direction TB
        BUY["Enter in stages, or add one due stage"]:::proposes
        TRIM["Trim into a realised catalyst"]:::proposes
        HOLD["Hold — including through one stated delay"]:::proposes
        EXIT["Exit review: cancelled · invalidated ·<br/>unfunded · out of extensions"]:::proposes
        WATCH["Research watch, or WAIT with the finding<br/>that stopped it"]:::proposes
    end

    ARM["Arms its own next review:<br/>the window end, and the earnings release"]:::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 --> HELD
    HELD -- yes --> UPDATE --> SIZE
    HELD -- no --> SWEEP --> NEW --> SIZE
    SIZE --> OUT --> ARM --> MAND --> YOU --> ORD

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

Cadence. A risk review after the Korean close on every held position; a WATCH on each open catalyst's window and on the disclosures that carry its confirming indicator; and a review around the earnings release that can confirm it — before, so the scenarios are written down, and after, so they are scored. Every judgement it submits arms the next one, including the judgements that conclude there is nothing to do. Your Aumos may refuse an arming, and the review interval stored for an installed manager is whatever you confirm on the install screen.

What it needs

Market	Korea-listed single equities, in won. Not ETFs, not overseas listings
Data sources	filings and public policy documents through open-dart for the receivable, working-capital, debt and cash-flow lines — read through the host's own OpenDART source cache, both routes of it: the corporate-code list that maps a ticker to the identifier filings are keyed by, and the filings index it then reads. A connected Toss market-data login supplies adjusted won price history and the Korea-listed enumeration this manager declares as its universe before it sweeps, because nothing else in the host can list a market. Web sources — a ministry gazette, a tariff decision — are filed as evidence with their url, publisher and publication date before anything cites them
What a key costs	an OpenDART key is free from Korea's Financial Supervisory Service on registration. The price connection is a login this fund already has; no brokerage account has to be attached for market data
Its own memory	two records, never merged. A catalyst register — open windows, delay counts, and which contrary findings are on the record; without it a position that has slipped three times reads as a first slip, and the manager says so rather than guessing. And a candidate ledger beside it — which names it is researching, what question each one is still open on, the receipt it stopped reading at, and the ranges it still owes a retry. Neither holds copied prices, filings, holdings or cash: those are re-read from your account and the sources every run
Settings	the catalyst horizon, the delay budget, the runway and refinancing floors, the number of improving channels required, the single-name ceiling and the trim threshold are all yours to set. The shape of the method is not: the catalyst record's required fields, the state machine, and the refusal to read a price move as success are the methodology
The book	it reads your holdings and every open proposal on the fund, because a name another manager has proposed and not yet filled is exposure you have already committed to
Your approval	it proposes and never trades. Every judgement is a proposal your Aumos judges against your Mandate and you approve or refuse

What it is bad at

• The value trap with a date on it. Everything this methodology asks for can be satisfied by a company that is simply declining slowly: a real event, a real window, a real improving quarter, and a business that keeps getting worse underneath. The invalidation conditions are the defence and they are not a complete one.
• Waiting. A catalyst that slips twice costs a year of capital, and nothing charges the position for that on a statement. The delay budget is a blunt instrument aimed at a real failure mode, and it will sometimes force a review on a thesis that was about to be right.
• Policy. Half of what it can find in Korea depends on a decision by a ministry or a regulator, and that decision is not forecastable from a filing. It can measure the execution; it cannot anticipate the reversal.
• Dilution. The companies this finds are frequently the companies that need money. A correct business thesis funded by an equity raise is a wrong per-share thesis, and the manager states the risk rather than pricing it.
• Finding things at all. Its entrance is an event in a filing or a gazette, and that sweep is incremental, budgeted and resumable rather than exhaustive. ⛔ **An unread filing window is reported as discovery_not_run, never as «no candidate was found».** Those two sentences produce the same empty answer and mean opposite things, and the only one this manager may write about the market is the one where the universe was declared, every required lane answered, and no range failed. Expect runs that honestly report having looked nowhere — a run that spent its whole filing budget reviewing what you already hold is one of them.
• The quiet one: it is at its least comfortable when it is working. The run that adjudicates a deadline as a failure and proposes an exit is the run that will look worst if the event arrives two months later.

Do not install this if you want a manager that will hold something indefinitely because the story is still good, one that reacts to price, or one that will treat a policy announcement as a position.

Notes

**The maintainer-facing document is ARCHITECTURE.md (ARCHITECTURE.md)** — what is in the deterministic core and why, which parts were derived from the evidence-gated package, which fixtures stand behind which rule, and what the host owns rather than this package.

This package shows no returns, and there is a specific claim it is refusing to make. The methodology is ported from a private research harness whose author reported a result on the reference case. That result was never re-audited in this design; it is not a validated edge, it is not this package's track record, and no threshold in this package was chosen to reproduce it. A forward track record takes calendar time rather than compute, and nothing on this page is one.

The reference case is a shape, not a rule. The 미수금 회수 / 요금 정상화 thesis that this methodology was written from is a policy-and-balance-sheet turnaround, and it is used in this package's fixtures to check that such a case classifies correctly and is not excluded for having an unattractive valuation multiple. There is no rule anywhere in this package that mentions any specific company, and the fixtures deliberately include a variant of the same case that does not reach a purchase.

Sectors: which one this package computes, and which one it only receives. This manager makes no judgement about what business a company is in — its case work is about an event and a balance sheet, not an industry — so it has no view of its own to offer on sector classification. What it does now do is read a sector ceiling if your Mandate states one. It used to read only the single-name limit, and "this package has no sector concept, therefore it cannot break a sector limit" was never a proof of anything: the host does not enforce that ceiling and this package was not receiving it, so under such a Mandate a correct-looking answer could take the account through a limit you had declared. If a sector ceiling is stated and the account's sector classification is incomplete — the candidate, or any position or open proposal in the book — this manager opens nothing and adds no stage. It waits, names the row it could not classify, and records it as missing data. **And if the total can be formed and is already past the ceiling, it opens nothing and adds no stage either — it says which sector, what it holds and what you declared, and the answer is about the account rather than about the company. For a while only the first of those two was wired: a ceiling this manager could not check withheld the entry, and a ceiling it checked and found exceeded did not. Closing out, reducing on an invalidation, resizing to a risk limit and holding through a delay are untouched by either. What a sector ceiling withholds is an increase**; it never decides how much of its own position this desk sells, because a ceiling other names filled says nothing about which name should come down.

Running it beside other managers. Exposure is measured across the whole account — real holdings and open proposals together — and per-strategy limits do not add up into a larger account limit. If you run this alongside another Korea manager, the single-name ceiling you set on the account is the one that binds, and this package will refuse a position rather than take its own share of a limit somebody else has already used.

Provenance. This is a port; NOTICE.md (NOTICE.md) carries the upstream repository, the pinned commit and the copyright line. No account data, credentials, order implementation or private ledger from that repository is included, and every figure in this package's fixtures is illustrative and was written here.