TWMD Python SDK

TWMD = TW Market Data = twmarketdata.com — the official Python SDK for Taiwan market data.

pip install twmarketdata

from twmd import Client

df = Client().daily_price("2330")     # TSMC daily OHLCV, no API key needed

That returns a pandas.DataFrame. No key, no signup, no config — the free tier serves five demo symbols so the example above runs as written.

The distribution is twmarketdata; the import is twmd.

---

Why this SDK exists

The REST API is not uniform. Across the 82 sellable datasets, the security identifier is spelled seven different ways, date bounds five different ways, as_of exists on 17 routes, offset on 9, and six routes are not the kebab-case of their dataset key. Response rows arrive under rows, items, or data depending on the dataset.

This SDK absorbs all of that, so you write ticker=, start=, end= everywhere and get one response shape back.

It also does something most market-data clients don't: it tells you what it doesn't know.

---

Three things this SDK guarantees

1. Two lines to a DataFrame

Every one of the 82 sellable datasets has a named method, generated from the live API's own registry so none can be missed:

from twmd import Client

c = Client()                                  # free tier
c.monthly_revenue(ticker="2330")              # 月營收
c.security_master(limit=10)                   # 證券主檔
c.trading_calendar(start="2026-01-01")        # 交易日曆

There's a generic escape hatch too, and a capability lookup so you never have to discover limits by trial and error:

import twmd

c.dataset("monthly_revenue", ticker="2330", start="2024-01-01")

twmd.capabilities("monthly_revenue")
# {'tier': 'free', 'status': 'active', 'as_of': 'client_unsafe',
#  'point_in_time_safe': False, 'knowledge_time_field': 'as_of_date',
#  'pagination': 'limit_only', 'runnable_without_key': True, ...}

2. Point-in-time that refuses to lie

as_of= replays a dataset to what was knowable on a given date. The interesting part is what happens when it can't:

c.balance_sheet(ticker="2330", as_of="2023-06-30")   # server-side replay

c.company_peer_groups(as_of="2023-06-30")
# raises PointInTimeUnavailable:
#   this dataset declares no knowledge-time axis, so there is no honest way to
#   replay it to a past date

That refusal is the feature. Take monthly_revenue: its declared as_of_date is the revenue period, not the announcement date — June revenue is disclosed in July. Filtering on it would mark June revenue as known on June 30, which is precisely the look-ahead as_of exists to prevent. So the SDK will not filter on that column.

It does not refuse blindly, though. The registry is a snapshot and the API is not, so for these datasets the SDK makes the request and looks: if the response carries a real server-supplied knowledge_date, that outranks the classification and the query is answered. monthly_revenue returns one as of 2026-08-12, so as_of="2026-06-30" now correctly drops the June figure (knowledge date 2026-07-10) and keeps May. When no such column comes back, the call raises — and says it checked.

Five states, one per dataset, all visible in capabilities():

state	datasets	behaviour
server	16	the route filters; passed straight through
client	45	PIT-safe, knowledge column published; filtered locally
client_unsafe	8	declared column isn't a disclosure date; refused unless you pass as_of_policy="declared_field"
client_unverified	5	declared column isn't in the published schema; verified against actual rows at runtime
unsupported	8	no knowledge axis at all; refused

When the API supplies a knowledge_date column, it wins over the declared field — and if those dates are imputed, you hear about it:

import warnings
# ImputedKnowledgeDateWarning:
#   1,234 of 1,242 rows (99.4%) carry kd_imputed=true (kd_source=statutory_deadline):
#   the knowledge date was derived from a statutory filing deadline, not observed
#   from an announcement. Treat this as a rule-based approximation of what was
#   knowable, not as an observed disclosure timestamp.

An imputed knowledge date is a rule, not a fact, and this SDK will never present it as one.

3. Gaps and truncation surfaced, never filled

No zero-filling. No forward-filling. No interpolation. Missing data is reported as missing, labelled with where the knowledge came from:

df = c.twse_daily_price(ticker="2330", start="2024-01-01")
m = df.twmd                       # or c.last_meta

m.truncated       # True if the row limit was hit on a route that can't paginate
m.data_gaps       # [Gap(2024-02-08..2024-02-14, no_row_for_trading_day), ...]
m.gaps_source     # 'server' | 'client_derived' | 'unsupported' | 'unknown'
m.coverage_min, m.coverage_max
m.source_role, m.lineage, m.data_as_of

gaps_source matters. Only 22 of 82 routes report gaps themselves; for daily per-entity datasets the SDK can derive them against trading_calendar (derive_gaps=True, costs one extra request), and where it can't do either it says unknown rather than implying there are none.

Truncation is flagged because only 9 of 82 routes support offset. On those 9 the SDK paginates to completion; on the other 73, hitting the limit sets truncated=True and warns. A short result is never silently passed off as a complete one.

---

Command line

Installing the package also installs twmd:

pip install twmarketdata

twmd                                 # no arguments: guided menu (or the full TUI)
twmd 2330                            # ticker shortcut: prices + revenue + what to look at next
twmd datasets --free-only            # what answers with no key
twmd describe monthly_revenue        # grain, filters, and its as_of semantics
twmd coverage twse_daily_price       # the window we actually cover
twmd get monthly_revenue --ticker 2330 --as-of 2024-06-30 --format csv
twmd ask "台積電最近三個月的月營收"      # plain language (needs a Pro plan)
twmd tui                             # full-screen UI  (pip install twmarketdata[cli])
twmd auth status                     # which key is in use (never echoes it)

twmd ask is routed to the existing ask tool on the MCP server — the CLI does no question parsing of its own. A CLI that guessed which dataset you meant would carry a second copy of that routing, and the two would drift: same question, two answers, and nobody would notice. Answers come back with their sources printed to stderr; if one arrives without source ids, it says so, because an unsourced answer and an invented one read the same.

twmd tui and the guided menu only ever start on a real terminal. Piped, in CI, or called by an agent, they would take over the terminal and wait for a keypress that never comes — so those paths print data (or --help) instead.

It is a thin shell over the same Client and registry this README documents — not a second implementation. A CLI that built its own requests would carry a second copy of the point-in-time rules, and two copies drift.

Three things it does on purpose:

1. Omitting --as-of prints a warning to stderr. You get the latest revision, including values revised after the date you may be reasoning about. For a backtest that is a look-ahead leak, and the response looks completely normal without the warning.
2. Gaps and truncation go to stderr; data goes to stdout. So twmd get … --format csv > out.csv gives you a clean CSV and you still see what was missing. (Python de-duplicates repeated warnings by default; the CLI turns that off, because every gap is worth seeing.)
3. Exit codes are classified: 3 auth, 4 entitlement, 5 unknown dataset, 6 rate limited, 7 invalid parameter, 8 upstream. Returning 1 for everything forces you to grep error strings, and error strings change.

twmd does not pull in pandas — it uses the SDK's no-pandas path, so the install stays small.

---

Agents: the MCP server

If you are wiring an agent rather than writing Python, the same data is served over MCP at https://mcp.twmarketdata.com/mcp — 34 tools, plus reference resources that read with no key at all. See TW-Market-Data/tw-market-data-mcp.

---

FinMind compatibility

Existing FinMind-shaped code can run against TW Market Data with the call sites unchanged:

from twmd.compat import finmind as fm

df = fm.taiwan_stock_daily(stock_id="2330", start_date="2026-08-01")

Every recognised call is graded, and the grade decides what happens:

grade	meaning	behaviour
A (19)	same grain and meaning	returns data
B (17)	same fact, different shape	reshapes, or refuses and points at the native method
C (24)	TWMD covers it differently	returns data + CompatSubstitutionWarning explaining the difference
D (46)	no equivalent in the 82 datasets	raises NotMappedError

fm.taiwan_stock_trading_daily_report(stock_id="2330")
# NotMappedError: taiwan_stock_trading_daily_report has no TW Market Data
# equivalent: Broker-branch daily flow is not in the 82. TWMD has the branch
# ROSTER (broker_branch_reference / securities_firm_master) but not
# branch-level buy/sell.

D-grade calls raise instead of returning an empty frame, because an empty frame reads as "your query matched nothing" — a different claim from "this data doesn't exist here". Nine further mappings are candidates that haven't been verified row by row yet; they also raise, and name the candidate dataset, rather than shipping possibly-wrong.

Two honest limits. Parameter names are mirrored, but response column names are TWMD's — a stock_id alias is added where an identifier column exists. Mirroring column names would mean comparing live responses from both services, which this project does not do. And the mapping table was built by reading the public method signatures of the open-source package (FinMind v2.0.7, read 2026-08-12) — the FinMind service was never called and no credentials were used.

mapping/finmind_map.csv records all 144 rows with their grade, confidence, and verification source.

This project is not affiliated with, endorsed by, or sponsored by FinMind. The name is used nominatively to identify the interface being made compatible. No FinMind source code is included or redistributed. See NOTICE.

---

Free tier

No API key required for five demo symbols: 2330 台積電, 2317 鴻海, 2454 聯發科, 0050 元大台灣50, 2603 長榮.

Every example in this README runs against one of the 16 datasets measured to return data without a key:

security_master · trading_calendar · monthly_revenue · twse_daily_price · index_constituents · trading_rules_reference · bond_convertible_reference · broker_branch_reference · warrants_reference · company_industry_exposures · company_peer_groups · securities_firm_master · fund_etf_metadata · issuer_classification · stock_delisting_lifecycle · stock_split_par_value_events

twmd.runnable_without_key()     # the list above, measured not declared
twmd.free_tier_symbols()        # the five demo symbols

That list is measured, not read off the tier column: three datasets marked tier=free return 401 in practice, so they're excluded. Requesting any other symbol without a key raises FreeTierSymbolError listing the five, rather than returning an empty frame.

With a key, everything your plan includes is available:

c = Client("your_api_key")            # or set TWMD_API_KEY

The key is read from the argument first, then TWMD_API_KEY. It is never logged and never appears in repr().

---

Errors

TwmdError
├── TwmdConfigError → FreeTierSymbolError
├── TwmdAuthError   → MissingApiKeyError, InvalidApiKeyError,
│                     TierRequiredError, InsufficientCreditsError
├── TwmdRequestError → DatasetNotFoundError, UnsupportedParameterError
├── RateLimitedError        429, and 403 temporarily_blocked
├── EndpointRetiredError    410
├── TwmdServerError         5xx and transport failures
├── PointInTimeUnavailable  as_of can't be honoured for this dataset
└── NotMappedError          compat call with no TWMD equivalent

Two of these are worth calling out:

403 temporarily_blocked is a rate limit, not a permissions problem. It reads like "forbidden" and sends people hunting for a plan bug that isn't there. The SDK classifies it as a rate limit, retries with exponential backoff and jitter, and says so in the message. Default max_concurrency is 2 — during measurement, four concurrent requests tripped a block that persisted for tens of minutes.

UnsupportedParameterError is deliberate. Passing start= to a dataset with no date parameter is an error, not something to quietly drop — silently ignoring it returns a full unfiltered history that looks like a filtered one.

---

Response type

TwmdFrame subclasses pandas.DataFrame, so isinstance checks pass and the whole pandas ecosystem works. Metadata rides along as df.twmd.

pandas preserves _metadata through slicing and copying, but some operations (merge, several groupby paths) build a fresh frame and drop it. So the same Meta is always on the client as c.last_meta, which no pandas operation can lose.

Without pandas the SDK still works and returns list[dict]. Pass raw=True for the decoded JSON envelope.

---

Coverage and limits

• 82 sellable datasets. The registry lists 125 entries and the OpenAPI spec mounts 119 dataset routes; 82 is the sellable, queryable subset this SDK ships.
• Dataset status is visible. 75 are active; 5 partial, 1 planned, 1 private_beta warn on use, because a partial series shouldn't be mistaken for a complete one.
• Registry is generated and dated. twmd.REGISTRY_MEASURED_ON tells you when the routes and semantics were last verified against the live API.

---

Upgrading from 0.1.0

0.2.0 keeps the same distribution (twmarketdata) and the same import name (twmd), so pip install -U twmarketdata is the whole upgrade. Everything 0.1.0 exported still resolves and still works, now emitting DeprecationWarning with the replacement named:

0.1.0	0.2.0
client.get_dataset("twse-daily-price", symbol="2330")	client.twse_daily_price(ticker="2330")
client.get_all(...) / iter_pages(...)	client.dataset(...) — paginates internally
client.list_datasets()	twmd.datasets() / twmd.capabilities(name)
twmd.access.is_key_free(...)	twmd.runnable_without_key()
twmd.frames.to_dataframe(payload)	returned frames already carry .twmd metadata
TwmdAPIError, TwmdPaymentRequired, …	still importable; TwmdPaymentRequired keeps payment / price / credits_url / purchase_hint

Two behaviours changed on purpose, both toward accuracy:

• The access tables were re-measured. 0.1.0 recorded 2 key-free datasets from a probe on 2026-07-21. A full 82-route sweep, plus a second sweep with a non-demo ticker to separate "open to anyone" from "demo symbols only", found 15 open and the same 3 sample-only. Names, semantics and return types are unchanged.
• Pagination stops when the server ignores offset. Measured on index-constituents: offsets 0, 3 and 6 returned identical pages. Continuing would append duplicate rows and present them as a full history, so pagination halts and the result is flagged truncated.

0.1.0 remains MIT; 0.2.0 onward is Apache-2.0.

---

Development

pip install -e ".[dev]"
pytest                        # offline suite
pytest -m network             # live free-tier checks, no key needed

Regenerating the registry after an API change:

python tools/build_mapping.py    # re-derive the mapping tables from live sources
python tools/gen_registry.py     # → twmd/_registry.json
python tools/gen_methods.py      # → twmd/_methods.py + .pyi
python tools/gen_compat_map.py   # → twmd/compat/_finmind_map.json

Design notes and the full evidence trail live in DESIGN_v1.md, mapping/, and mapping/sources/.

---

Links

• Website: https://twmarketdata.com
• Pricing: https://twmarketdata.com/en/pricing
• REST API (OpenAPI): https://api.twmarketdata.com/openapi.json
• MCP server: https://mcp.twmarketdata.com/mcp · docs & issues
• This repo: https://github.com/TW-Market-Data/twmarketdata

---

License

Apache-2.0. See NOTICE for trademark and compatibility statements.

Not investment advice.