hsnuhow Choosing a framework for a Taiwan business question
A framework is a way of being wrong in an organised fashion. Its only job is to turn a vague question into a claim that data can kill. Pick the wrong one and you produce a deck that is well-structured and says nothing.
This skill picks. It is deliberately thin on what each framework is — that is a search away — and thick on which question it answers, what number feeds it, and when it is the wrong tool.
The failure mode this exists to prevent
Framework theater: applying a framework because it is the framework people apply, filling its boxes with whatever is at hand, and presenting the filled boxes as an argument. The tell is that every box has content and none of it changes the recommendation. A five-force analysis where all five forces are "moderate" cost a week and decided nothing.
The rule: a framework earns its place only if a plausible fill of its boxes would flip the recommendation. If you cannot say in advance what result would change your mind, don't run the framework — go get a number instead.
Route the question first
Start from what was actually asked, not from what is fashionable.
| The question actually asked | Framework | What feeds it (tw-data) |
|---|---|---|
| How big is this? Is it worth entering? | Bottom-up sizing (TAM/SAM/SOM) | household.py --tam 類別 · --mix — grain check first, see below |
| Can this business model sustain itself? (訂閱/平台/內容/freemium) | Unit economics — mostly no data | ARPU, conversion, churn, cost structure need first-party research. 付費意願 and ad-market size are snapshotted with retrieval methods in tw-data's references/external-sources.md (Reuters DNR: TW paid-news rate 10% in 2026; DMA: digital ad market 636.83億 in 2024) |
| Is now the time? Should we wait? | Cycle timing | ndc_signal.py · moea_orders.py --orders |
| Is this industry structurally attractive? | Porter 五力 (partial — see below) | twse_revenue.py --industry · moea_orders.py --production --sector · mof_industry.py --industry (SME-inclusive size/trend) |
| Who are we up against, and who's winning? | Competitor roll-up | twse_revenue.py --industry · --top / --bottom |
| Who is the customer? Which segment? | Segmentation | household.py --quintile · --spend (by 縣市) |
| Where do we grow from here? | Ansoff | household.py --mix (adjacent categories) |
| What is the macro backdrop? | PEST, macro half only | dgbas_macro.py --gdp --cpi --unemployment --wage |
| Why did the number move? | Decomposition — not a framework, arithmetic | whichever series moved; ndc_signal.py components |
If the question isn't in this table, either it is not a strategy question, or — the common case — it is one this data layer cannot feed (unit economics is the classic). Say which, rather than reaching for the nearest matrix. "This case is undecidable on free data; here is the cheapest thing that would decide it" is a deliverable, and often the most valuable one.
The frameworks, briefly, and when not to
Bottom-up sizing (TAM/SAM/SOM)
households × spend per household × category share, then haircut to what you
could serve and what you could win. household.py --tam does the TAM line.
Grain check first. --tam resolves to the survey's 10 top-level
categories and nothing finer. If the target market is narrower than a whole
category — a magazine inside 休閒、運動、文化及教育, which also bundles tuition,
cram schools and package tours — the output is an upper bound of an upper bound,
off by orders of magnitude (that bucket is ~NT$5,000億; Taiwan's entire magazine
industry is ~百億級). Use it as a ceiling, never as the TAM, and never let a
failed narrow lookup get "fixed" into a category-level number.
Share is not level. A category's falling share of the household budget does
not mean falling NT$, and vice versa — always compute the absolute series too
(share × per-household total spend, CPI-deflated via dgbas_macro.py --cpi)
before calling a category "declining". And mind attribution: a bundled category
can move for reasons unrelated to your slice — 少子化 mechanically shrinks the
education sub-item inside 休閒、運動、文化及教育 regardless of what media
spending does. A bundle trend is a proxy for your slice, not evidence about it.
Use it for any consumer category at category grain. It is the strongest thing in this plugin: every line is defensible and a reader can attack the assumptions rather than the number, which is what you want.
Don't for B2B, government, or infrastructure demand — 家庭收支調查 is household consumption only. Sizing a B2B market from it is not conservative, it is unrelated. For B2G use PCC 政府採購 (not yet wrapped; see tw-data's endpoints reference).
The SOM is where sizing dies. TAM is arithmetic; SOM is a claim about your company. Never present a SOM as if the data produced it.
Cycle timing
The 景氣對策信號 tells you where the economy is; 外銷訂單 tells you where it is
going, because orders are booked before they ship. ndc_signal.py gives the
score, the light, and the component breakdown.
Use it when the recommendation is when, not whether, and when the answer plausibly changes with the cycle.
Don't let it decide a structural question. A red light does not make a bad market good. Timing arguments are only interesting once the market has already survived the sizing question — and the light is a national signal, so it is close to irrelevant for a niche that moves on its own drivers.
The trap: Taiwan's cycle light is currently dominated by export electronics. A 紅燈 while 零售營業額 is +3% and 民生工業 production is negative means the economy is booming somewhere else. If your case is domestic consumer, the headline light is actively misleading. Decompose before you cite it.
Porter 五力 — usable at about 40% in Taiwan
Honest accounting of what free data can and cannot support:
| Force | Feasible? | With what |
|---|---|---|
| 現有競爭 | Yes, if the industry's top players are listed | twse_revenue.py --industry — concentration, growth dispersion. TWSE's 33 產業別 have no 出版/媒體/文創 (those are TPEx or private); for media, education and professional services this cell is No |
| 新進入者 | Partial | GCIS 公司登記 (not wrapped) for entry rates |
| 供應商議價 | No | Needs cost structure — not in any free source |
| 買方議價 | Partial | household.py --quintile for consumer; nothing for B2B |
| 替代品 | No | Requires judgement, not data |
So: run the rivalry force properly and say the other four are judgement. A five-force where four boxes are opinion dressed as analysis is worse than one force with a number. Never present the five as equally evidenced.
Don't use Porter for "how do we win" — it is an industry-attractiveness tool. Firm-level questions need a different lens.
Competitor roll-up
twse_revenue.py --industry 半導體業 gives every listed company's monthly
revenue and YoY in that industry. Rolled up, that is an industry demand tracker;
read down the rows, it is a competitive scoreboard, monthly, free.
The hard limit: listed companies only. Taiwan's economy is overwhelmingly SME, and none of them are here. In a fragmented consumer category — restaurants, clinics, retail — the listed set may be a rounding error on the real market. Do not compute market share from this and present it as market share. State the coverage or don't use the number.
--min-revenue (default 1億) exists because a company going from NT$2m to
NT$6m posts +200% YoY and will top any unfiltered ranking with pure noise.
Segmentation
--quintile (income fifths) and --spend (by 縣市) are the two real axes free
data gives you. Both are structural and slow: 家庭收支 lags ~18 months.
Don't invent psychographic segments and pretend the survey supports them. If the segment isn't a cut the data actually makes, it is a hypothesis — label it.
Ansoff, PEST, and the rest
Use them as checklists, not analyses. Ansoff is a prompt to ask "is this new
product or new market"; ten minutes, verbally, no slide. PEST's P/S/T legs have
no data behind them here — the E leg is dgbas_macro.py and the rest is
reading. A PEST slide is almost always filler.
What this skill will not help you produce
- SWOT — the purest form of framework theater. Every box fills, nothing decides. If someone asks for one, ask what decision it informs; there usually isn't one. Ansoff or a sizing pass answers the real question underneath.
- BCG growth-share matrix — needs relative market share. Taiwan free data cannot give you market share in any fragmented category (see the SME gap above). You would be plotting guesses on two axes.
- Any framework requiring cost structure or margin by competitor — not available free. TWSE gives revenue, not margin.
Being unable to build these is a finding worth stating, not a gap to paper over with estimates.
From question to OPTIONS — the step before any framework
A framework analyses; it does not choose. Before framework selection, frame the
question as a choice between 2-4 MUTUALLY EXCLUSIVE options (the Martin/Lafley
discipline; biz-case-workflow automates this). Typical option sets by
question type — starting points, not menus:
| Question type | Typical mutually exclusive options |
|---|---|
| Market entry (brand/product) | Wholly-owned build · bolt onto an existing local operation · asset-light license/syndicate · don't enter (status quo) |
| Business-model sustainability | Ad-funded · direct paid/subscription · hybrid membership · B2B/licensing pivot |
| Timing | Enter now · enter on a named trigger (signpost) · wait one cycle |
| Competitive response | Match the mover · flank (different segment/model) · hold and harvest |
| Growth direction | Deepen current segment · adjacent category (Ansoff) · adjacent geography · acquire |
Rules that carry over from the choice discipline: the status quo is an option and faces the same tests; each option must be a coherent story of WINNING, not a hedge; if a proposed option ignores an existing local operation of the same company, the framing is wrong — check first.
Which MODEL feeds the analysis
"Model" means a driver tree with a two-ended range — never a single point. Route by what must be quantified:
| What must be quantified | Model shape | Feeds from |
|---|---|---|
| Market size / opportunity | TAM driver tree: population × penetration × frequency × price | household.py --tam/--mix for the ceiling; analog-market curve for penetration (see below) |
| Model sustainability | Unit economics: audience × conversion × ARPU − cost | Reuters DNR paid rates bound conversion; costs are first-party (declare untestable) |
| Timing | Cycle position + lead indicators | ndc_signal.py, moea_orders.py --orders |
| Industry trajectory | Tax-base trend, real (CPI-deflated) | mof_industry.py --history + dgbas_macro.py --cpi |
| Competitive intensity | Share-shift / concentration read | twse_revenue.py --industry (listed only — say so) |
Analog-market parameterization — the move that makes Taiwan modelable:
where Taiwan has no number (penetration endpoint, paid-content rate, adoption
speed), bound it with a NAMED analog market's actual curve — JP/KR/HK/SG by
default (retrieval methods in tw-data's references/external-sources.md).
State the lag logic ("Taiwan trails Korea's broadband curve by N years") and
why the analog is right. MGI sizes China e-tail scenarios exactly this way:
high case = ride Korea's curve, base = US/Taiwan, low = Taiwan's own.
Sensitivity is mandatory: name the two most contested drivers and show how the range moves as each swings. If the recommendation survives the whole swing, say so — that is a strong finding. If it flips inside the plausible range, the driver is a key debate, not an assumption.
Framework → argument
The framework is scaffolding; it comes down before delivery. What ships is a claim, a number, and what would falsify it:
- Bad: "Porter analysis shows moderate rivalry." (fills a box, decides nothing)
- Good: "The top 3 listed players hold X% and all three grew slower than the industry last year — the share is moving to companies we can't see, which means our competitor set is SMEs, not the incumbents." (a claim, a number, and a clear thing that would prove it wrong)
Then cross-check. tw-data's four independent reads — cycle light, listed revenue, export orders, production/retail — exist so a claim can be corroborated. When they agree, the finding is solid. When they diverge, the divergence is the finding.
Related: tw-data for the numbers, biz-case-workflow to run the whole case
end-to-end (options → conditions → debates → tests → model → choice).