TXF: 2011-01-03 (1d and intraday). MXF: 2020-03-22 (1d). Stock futures start per contract — CAF returned 1d bars at least as early as 2016-09-18 (the probe window began there, so it may reach further back)
Update
Near-real-time from the on-disk 1m store, refreshed on request
Source
Sinopac Shioaji (near-month continuous series)
Parameters
symbolstringpathRequired
Near-month continuous contract
AllowedTXF (台指期), MXF (小台), or a stock-futures id listed by ohlcv/symbols (case-insensitive)
schemastringpathRequired
Bar size
Allowed1m, 5m, 15m, 30m, 60m, 1d
startstringquery
First day
Defaultend − the schema's max rangeAllowedYYYY-MM-DD
endstringquery
Last day
Default today (UTC)AllowedYYYY-MM-DD
Max range per request (end − start): 1m 31 d · 5m 62 d · 15m 93 d · 30m 186 d · 60m 365 d · 1d 3650 d.
Intraday bars before 2017-05-15 are day session only (the night session 15:00–05:00 Taipei launched then), so bars per day jump from ~300 to ~1,140 at that date.
The series rolls contracts at monthly settlement without price adjustment — mask the roll in backtests.
Roll rule (all schemas; TXF, MXF, and stock futures alike): monthly settlement day — the 3rd Wednesday, or the next trading day when that Wednesday is closed — is the expiring month for the whole day. It trades through its 13:30 Taipei close, no bar is stamped 13:31–14:59 Taipei on that day, and the series is the next month from the 15:00 night session on. The settlement day's 1d bar is the expiring month's, so its close is that contract's last trade at 13:30.
For years of history use ohlcv/<symbol>/export/<year> instead of chunked JSON.
A range with no data returns 200 with [].
Do not assume the series runs to the last trading day — MXF and CAF asked for everything up to the request date both came back with a last bar months old. Read data[-1]["ts"].
1d bars carry a 16:00:00+00:00 time component (00:00 Taipei), not midnight UTC — take the date from ts rather than assuming the timestamp is date-aligned.
{"data": ["CDF", "DHF", "MXF", "TXF", ...]} — symbols accepted by ohlcv, ohlcv/export, and bid_ask_vol: TXF, MXF, plus the liquidity-picked subset of stock futures with minute-line coverage (most of the 231 stock futures are not covered).
{"error": "no_data", "message": "No 1m data on disk for <symbol> in <year>"}
No file
Notes
One request per year, zero server-side computation — resample locally.
Same near-month continuous series as ohlcv/<symbol>/1m, under the same settlement-day roll rule (no bars stamped 13:31–14:59 Taipei on monthly settlement day).
Rows are the near-month continuous contract under the same roll rule as ohlcv: on monthly settlement day there are no rows between the expiring month's 13:30 Taipei close and the 15:00 night open; from the night session on the rows are the next month.
Per contract listing — CDF asked from 1998-07-01 returned rows starting 2010-01-25
Update
As /daily/<futures_id>
Source
As /daily/<futures_id>
Parameters
futures_idsstringqueryRequired
Ids
Allowed Comma-separated stock-futures ids, ≤ 250
startstringquery
First date
Default1998-07-01AllowedYYYY-MM-DD
endstringquery
Last date
Default todayAllowedYYYY-MM-DD
Example
Python
body = requests.get(f"{BASE_URL}/studio/market/twfutures/stock_futures/batch/daily", headers=headers, params={"futures_ids": "CDF,DHF", "start": "2026-08-01"}, timeout=120).json()
Response
{"data": {"<futures_id>": [...rows as /daily...]}, "failed": [...]}. failed = ids whose fetch failed (retry); an id absent from both genuinely has no data.
Errors
Status
Body
When
400
{"error": "futures_ids parameter required"}
Missing
400
{"error": "futures_ids exceeds max 250 per request"}
Too many
400
{"error": "Invalid stock futures_id(s): [...]"}
Any id is not a stock future
503
{"error": "Stock futures symbol list temporarily unavailable. Please retry shortly."}
{"error": "institutional data not available for stock futures (<id>)"}
Stock future
404
error message
No data
503
upstream rate limit / symbol list unavailable
Retry later
Notes
*_amount is TWD thousand — inferred from magnitude, not stated upstream, but the arithmetic backs it: a TX row with long_open_interest_balance_amount 722870036 over ..._volume 78292 is 9,233 per contract, and a TX contract is index × 200, so 9,233 thousand ÷ 200 ≈ 46,160 index points, which is where the index was trading that day. Read as TWD it would be 1,000× too small.
2007-01-02 — TXO asked from 1998-07-01 returned nothing before that date
Update
Daily; server cache 5 min
Source
FinMind
Parameters
option_idstringpathRequired
Option product
AllowedTXO (2–6 letters accepted)
startstringquery
First date
Default1998-07-01AllowedYYYY-MM-DD
endstringquery
Last date
Default todayAllowedYYYY-MM-DD
Example
Python
data = requests.get(f"{BASE_URL}/studio/market/twfutures/option/large_traders/TXO", headers=headers, params={"start": "2026-08-01"}, timeout=60).json()["data"]
Response
{"option_id", "data": [...]}, 6 rows per day (call / put × week / current month / all). Fields: date, option_id, name, put_call, contract_type, and the same top-5 / top-10 trader and specific-legal-person open-interest fields as futures large traders, plus market_open_interest.
2023-10-18 (TAIFEX publishes only about the last three years)
Update
Daily after TAIFEX publishes — Blave jobs at 15:40 and 17:40 Taipei; server cache 5 min
Source
Blave estimate from TAIFEX data only: 三大法人-區分各期貨契約 (TX / MTX / TMF) and 期貨每日交易行情 (TX close)
Blave's estimate of the average cost of one institution's net TAIEX-futures position, with its unrealized and realized PnL. An estimate from daily aggregates, not the institution's real fills.
Method:
Position = long OI − short OI, converted to 大台 (TX) contracts: TX × 1, MTX × 1/4, TMF × 1/20 (scope=all), or TX only (scope=tx).
Adding to the position uses that institution's own average trade price of the day on the side it added (TAIFEX 交易契約金額 ÷ (口數 × 200)) and updates a weighted-average cost. Reducing books realized PnL at the opposite side's average price and leaves the cost unchanged. A flip, or going flat, realizes the whole old position; the new direction starts at that day's price.
Every monthly settlement day (third Wednesday; the next trading day if closed) starts a new cycle: cost resets to the mark price and realized PnL to 0. This makes the cost independent of where the data starts.
Mark price = TAIFEX's own valuation of the open position (OI contract value ÷ (contracts × 200), i.e. at each month's settlement price). Unrealized PnL = (mark − cost) × net contracts × 200.
all = TX + MTX + TMF in TX-equivalent contracts; tx = 大台 only (the basis most Taiwanese sites quote)
DefaultallAllowedall, tx
The ledger always runs from the start of the stored history; start / end only cut the window, so the same day has the same values whatever window you ask for.
{"identity", "scope", "data": [...]}, one row per trading day, oldest → newest; "stale": true is added only when the server's copy is behind TAIFEX's last published day (the newest row may be missing).
Field
Type
Unit
Description
date
string
—
Trading day (Taipei)
net_open_interest
float
TX-equivalent contracts
Long − short OI (negative = net short)
net_open_interest_change
float
contracts
Change vs the previous trading day
cost
float | null
index points
Weighted-average cost of the net position; null when flat
txf_close
float | null
index points
TX near-month day-session close (the expiring contract is skipped on settlement day)
mark_price
float | null
index points
TAIFEX valuation of the position, at settlement prices
unrealized_pnl
int
TWD
(mark − cost) × net contracts × 200
realized_pnl
int
TWD
Realized since cycle_start; resets to 0 on every settlement day
{"error": "Invalid start date, expected YYYY-MM-DD"} (or end) / {"error": "start must not be after end"}
Bad date / reversed range
503
{"error": "Market data is not ready on this server. Please retry later."}
Store not ready — retry later, never read as "no position"
500
{"error": "Internal error"}
Unexpected server error
Notes
Futures only: option positions (including dealers' hedges) are not in the ledger, so a dealer's "loss" here may be offset elsewhere.
Right after a settlement day the cost equals the mark by construction — the first days of a cycle say little about where the position was really built.
Other sites' 外資成本 figures use different rules (TX only, near-month close as the trade price, or undisclosed) — numbers will differ; compare like with like (scope=tx).
Let Your AI Do Quanton Your Own Computer
Blave Desktop is free, open source and needs no sign-up. Connect Claude Code or Codex: it writes the strategy, runs the backtest and trades it live.
Downloading.Drag it into Applications,open it, and pick your AI.The first launch sets up its runtimeand takes a few minutes.If you like it, star it on GitHub.
Downloading. When you open the installer, Windows first shows “Windows protected your PC”:
Click “More info”.
Click “Run anyway”. Blave installs and opens on its own.
The first launch sets up its runtime and takes a few minutes.
Windows warns because the installer isn't code-signed yet (we've applied for a certificate), so it can't identify the publisher. The source is public on GitHub. See the full steps
