diff --git a/docs/llms-full.txt b/docs/llms-full.txt index 5b2f0de..aa487e2 100644 --- a/docs/llms-full.txt +++ b/docs/llms-full.txt @@ -163,6 +163,9 @@ Async batch query is supported by datasets that require `data_id` parameter. Not - Columns: date (str), stock_id (str), Trading_Volume (int64), Trading_money (int64), open (float64), max (float64), min (float64), close (float64), spread (float64), Trading_turnover (float32) - Note: For emerging-board (興櫃) stocks, open is the previous-day average price (前日均價) published by TPEx, not an opening price, so it can fall outside [min, max] on volatile days. This is a field-definition difference, not a data error — max/min/close are the day's high/low/last and are correct. Listed (TWSE)/OTC (TPEX) stocks are unaffected. TaiwanStockInfo.type reflects the current market, so a stock that later moved to TWSE/TPEX no longer shows its earlier emerging period. - Note: When TWSE/TPEx publishes no traded price for a stock on a given day, open/max/min/close are shown as `--` in the source and are stored here as 0, so open/max/min/close/spread are all 0 for that row. The source publication carries no price — this is not a crawling gap. It can occur in any market (TWSE/TPEX/Emerging): a halted stock, a day with no trades, or a day with only sporadic trades that produce no published price. Example (no trades): 2317 on 2025-07-30 has Trading_Volume=0 and open/max/min/close=0, while 2025-07-29 (close=171.5) and 2025-07-31 (close=178) are normal. Example (volume but no price): 9929 on 2025-07-31 has Trading_Volume=451 yet open/max/min/close=0. Because of the second case, do NOT filter these rows with `Trading_Volume > 0` alone — filter on the price column instead, e.g. `df[df["close"] > 0]`, or forward-fill from the previous trading day. TaiwanStockPriceAdj does NOT show 0 on these days (see its own note). +- Note: Survivorship bias before 2005 (listed/TWSE only). For the listed market before 2005 the universe covered here consists mainly of securities that kept trading afterwards; securities that were trading at the time but have since been delisted may not have their history for that period fully covered, so fewer securities are present than were actually listed then. Some stocks also start later here than their real listing date - e.g. 2311 and 2325 begin no earlier than 2004-02-11, and 2319 is absent entirely. Building a stock universe, industry population or custom index from pre-2005 listed data therefore skews the sample toward survivors and systematically overstates returns and understates risk. Start long backtests after 2005, or state the limitation explicitly. OTC (TPEx) data is unaffected. The same applies to TaiwanStockPriceAdj, which is derived from this table. +- Note: OTC (TPEx) daily quotes are missing for the 71 trading days from 2007-01-02 to 2007-04-20. TPEx did not publish a per-stock daily quote for that window, so this is a gap at the source, not a crawling omission. 2006-12-29 and 2007-04-23 each carry roughly 400 OTC stocks while the 71 days in between carry only a handful (mostly securities that traded on another board then and moved to TPEx later). Listed (TWSE) data over the same period is normal. Treat OTC data on those days as missing rather than "no trades that day" - feeding zeros in distorts returns, turnover and market-breadth calculations. +- Note: When querying a whole day without data_id, the response also contains about 30 index rows (currently 32 codes) whose stock_id is an English index name rather than a numeric code: TAIEX (capitalization-weighted index), TPEx (OTC weighted index), and 30 sector indices such as Cement, Food, Plastics, Semiconductor, FinancialInsurance, Optoelectronic. On these rows open/max/min/close are index points (TAIEX closed at 22553.72 on 2025-07-01) and Trading_Volume/Trading_money are the sector's traded volume and value. Downstream code that treats every stock_id as a stock code will pick them up and get nothing back. Filter them out with `df[df["stock_id"].str[0].str.isdigit()]`, or exclude ids whose TaiwanStockInfo.industry_category is 大盤 or Index. Invert either condition if the indices are what you want. ### TaiwanStockPriceAdj (台灣還原股價資料表) - Tier: Free (with data_id) / Backer/Sponsor (all stocks by start_date only) @@ -197,6 +200,8 @@ Async batch query is supported by datasets that require `data_id` parameter. Not - Params: dataset=TaiwanStockStatisticsOfOrderBookAndTrade, start_date=2021-01-07 - Columns: Time (str), TotalBuyOrder (str), TotalBuyVolume (int64), TotalSellOrder (int64), TotalSellVolume (int64), TotalDealOrder (int64), TotalDealVolume (int64), TotalDealMoney (int64), date (str) - Note: Single day per request. +- Note: The disclosure interval varies by era despite the "every 5 seconds" name - TWSE raised its frequency over the years. Rows per trading day: 2005-01-03~2011-01-14 = 60s, 271 rows; 2011-01-17~2014-02-21 = 15s, 1081 rows; 2014-02-24~2014-12-27 = 10s, 1621 rows; 2014-12-29~now = 5s, 3241 rows. This is the granularity published at the source at the time, not a gap here - TWSE's own publication for 2005-09-28, for instance, is also 271 rows. Do not assume 3241 rows on every trading day; resample on the day's actual Time column instead of slicing by a fixed row count or positional index. +- Note: Columns are running totals for the day, so 09:00:00 normally starts near zero. On 2023-08-04 the 09:00:00 timestamp carries, besides the normal opening row, a second row whose seven numeric columns are identical to the previous trading day's close (2023-08-02 13:30:00); 2023-08-03 was a typhoon closure. The TWSE source file for that day already contains this value - it is not a processing error. Differencing the table with diff() therefore produces a large negative first bar that day. It is the only such case seen in spot checks of the 5-second era (from 2014-12-29). Before differencing, keep only the smallest running total per timestamp: `df = df.loc[df.groupby(["date", "Time"])["TotalDealVolume"].idxmin()]`. ### TaiwanVariousIndicators5Seconds (台股加權指數) - Tier: Free @@ -204,6 +209,7 @@ Async batch query is supported by datasets that require `data_id` parameter. Not - Params: dataset=TaiwanVariousIndicators5Seconds, start_date=2020-07-01 - Columns: date (str), TAIEX (float64) - Note: Single day per request. +- Note: The disclosure interval varies by era despite the "every 5 seconds" name - TWSE raised its frequency over the years. Rows per trading day: 2005-01-03~2011-01-14 = 60s, 271 rows; 2011-01-17~2014-02-21 = 15s, 1081 rows; 2014-02-24~2014-12-27 = 10s, 1621 rows; 2014-12-29~now = 5s, 3241 rows. This is the granularity published at the source at the time, not a gap here - TWSE's own publication for 2005-09-28, for instance, is also 271 rows. Do not assume 3241 rows on every trading day; resample on the day's actual date column (which carries the time) instead of slicing by a fixed row count or positional index. ### TaiwanStockDayTrading (當日沖銷交易標的及成交量值) - Tier: Free (with data_id) / Backer/Sponsor (all stocks by start_date only) @@ -262,6 +268,7 @@ Async batch query is supported by datasets that require `data_id` parameter. Not - Params: dataset=TaiwanStockEvery5SecondsIndex, start_date=2025-05-09 - Columns: date (str), time (str), stock_id (str), price (float), kind (str) - Note: Single day per request. +- Note: The disclosure interval varies by era despite the "every 5 seconds" name - TWSE raised its frequency over the years. Rows per trading day per index: 2005-01-03~2011-01-14 = 60s, 271 rows; 2011-01-17~2014-02-21 = 15s, 1081 rows; 2014-02-24~2014-12-27 = 10s, 1621 rows; 2014-12-29~now = 5s, 3241 rows. This is the granularity published at the source at the time, not a gap here - TWSE's own publication for 2005-09-28, for instance, is also 271 rows. Do not assume 3241 rows per index on every trading day; resample on the day's actual time column instead of slicing by a fixed row count or positional index. ### TaiwanStockSuspended (台股暫停交易公告) - Tier: Backer/Sponsor diff --git a/docs/tutor/TaiwanMarket/Technical.en.md b/docs/tutor/TaiwanMarket/Technical.en.md index 9cdc26d..2f0f4bb 100644 --- a/docs/tutor/TaiwanMarket/Technical.en.md +++ b/docs/tutor/TaiwanMarket/Technical.en.md @@ -346,6 +346,49 @@ In Taiwan stock technical data, we have 20 datasets, as follows: **Difference from `TaiwanStockPriceAdj`**: the adjusted-price table is **not** `0` on these days — it carries forward the previous trading day's adjusted price (while `Trading_Volume` keeps the day's actual value). For instance, 2317 鴻海 on 2025-07-30 in `TaiwanStockPriceAdj` is `open=169.94`, `max=170.43`, `min=166.06`, `close=166.54`, i.e. 2025-07-29's adjusted price. Use `TaiwanStockPriceAdj` if you want a series with no `0` gaps; use `TaiwanStockPrice`'s `0` if you need to identify which days had no published price. +??? note "Listed (TWSE) data before 2005: the history of delisted securities may be incomplete — mind survivorship bias in long backtests" + For the **listed (TWSE)** market **before 2005**, the universe covered by this table consists mainly of securities that kept trading afterwards. Securities that **were trading at the time but have since been delisted** may not have their history for that period fully covered, so the number of securities present is smaller than the number actually listed back then. + + Some stocks also start **later in this table than their real listing date**. For example, 2311 and 2325 begin no earlier than `2004-02-11` here, and 2319 is absent entirely. + + **Impact on backtests**: building a stock universe, an industry population or a custom index directly from this table's pre-2005 listed data skews the sample toward the securities that survived, which **systematically overstates returns and understates risk** (survivorship bias). We suggest: + + - starting long backtests **after 2005**; or + - if you must cover earlier periods, stating explicitly in your methodology and conclusions that the pre-2005 listed sample is not the complete universe of that time, so the bias is not mistaken for strategy performance. + + **OTC (TPEx) data is unaffected** — the pre-2005 OTC universe is normal. The same note applies to `TaiwanStockPriceAdj` (adjusted price), which is derived from this table. + +??? note "OTC (TPEx) daily quotes are missing for 2007-01-02 ~ 2007-04-20 (not published at the source)" + For the **OTC (TPEx)** market, this table has no per-stock daily trading information across the 71 trading days from **2007-01-02 to 2007-04-20**. TPEx did not publish a corresponding per-stock daily quote for that window, so this is **a gap at the source, not a crawling omission**. + + **What you can observe**: 2006-12-29 and 2007-04-23 each carry roughly 400 OTC stocks, while the 71 trading days in between carry only a handful (mostly securities that traded on another board at the time and moved to TPEx later). **Listed (TWSE) data over the same period is normal.** + + **Practical advice**: if your analysis window covers the first half of 2007, treat OTC data on those 71 trading days as **missing**, not as "no trades that day". Feeding them in as zeros distorts returns, turnover and market-breadth calculations; prefer listed data for that window, or skip the OTC sample there. + +??? note "When fetching a whole day without `data_id`, `stock_id` also contains index codes (not stocks)" + When you query with the "all data for a given date" form (no `data_id`), the response contains — besides stocks, ETFs and warrants — about **30 index rows** (currently 32 codes). Their `stock_id` is an **English index name** rather than a numeric code, for example: + + - Market indices: `TAIEX` (the capitalization-weighted index) and `TPEx` (the OTC weighted index) + - Sector indices: `Cement`, `Food`, `Plastics`, `Semiconductor`, `FinancialInsurance`, `Optoelectronic`, and 24 more + + On index rows, `open` / `max` / `min` / `close` are **index points** (e.g. `TAIEX` closed at `22553.72` on 2025-07-01), while `Trading_Volume` / `Trading_money` are that sector's traded volume and value for the day. If a downstream step treats every `stock_id` as a stock code (looking each one up for K-bars, chip data or financials, say), these index rows get picked up and return nothing. + + **Two ways to filter them out**: + + ```python + # Option 1: keep only codes starting with a digit (stocks, ETFs, warrants) + df = df[df["stock_id"].str[0].str.isdigit()] + + # Option 2: exclude them via TaiwanStockInfo's industry_category + stock_info = api.taiwan_stock_info() + index_ids = stock_info.loc[ + stock_info["industry_category"].isin(["大盤", "Index"]), "stock_id" + ] + df = df[~df["stock_id"].isin(index_ids)] + ``` + + Conversely, if the market and sector indices are exactly what you want, just invert either condition. + !!! example === "Package" ```python @@ -1360,6 +1403,34 @@ In Taiwan stock technical data, we have 20 datasets, as follows: - Data range: 2005-01-01 ~ now +??? note "The disclosure interval varies by era — early data is not one row every 5 seconds" + Despite the "every 5 seconds" name, **TWSE raised its disclosure frequency over the years**, so earlier data comes at a coarser interval. Rows per trading day: + + | Period | Disclosure interval | Rows per trading day | + |---|---|---| + | 2005-01-03 ~ 2011-01-14 | 60 seconds | 271 | + | 2011-01-17 ~ 2014-02-21 | 15 seconds | 1,081 | + | 2014-02-24 ~ 2014-12-27 | 10 seconds | 1,621 | + | 2014-12-29 ~ now | 5 seconds | 3,241 | + + This is **the granularity published at the source at the time, not a gap in this table**: TWSE's own publication for those dates carries the same number of rows (2005-09-28, for instance, is 271 rows officially). + + **Practical advice**: do not assume 3,241 rows on every trading day. When aligning time series across eras or computing fixed 5-second increments, resample on the day's actual `Time` column rather than slicing by a fixed row count or positional index. + +??? note "On 2023-08-04 the `09:00:00` open carries an extra row with the previous session's closing cumulative values" + Every column in this table is a **running total for the day** (cumulative order count/volume, cumulative deal count/volume/value), so `09:00:00` normally starts near zero. On **2023-08-04**, however, `09:00:00` carries — in addition to the normal opening row — a second row whose seven numeric columns are identical to the **previous trading day's close, 2023-08-02 `13:30:00`** (08-03 that week was a typhoon closure). **The TWSE source file for that day already contains this value**; it is not a processing error on our side. + + **Practical impact**: if you difference this table (`diff()`) to recover the per-5-second increments, the first bar of that day comes out as a large negative number. + + In spot checks across the 5-second era (from 2014-12-29) this is the only such case; no other trading day shows the same pattern. Before differencing, you can recover the true opening row by keeping only the smallest cumulative value per timestamp: + + ```python + df = df.sort_values(["date", "Time"]) + # where a timestamp has several rows, keep the smallest running total (the real opening row) + df = df.loc[df.groupby(["date", "Time"])["TotalDealVolume"].idxmin()] + inc = df.groupby("date")[["TotalDealVolume", "TotalDealMoney"]].diff() + ``` + !!! example === "Package" ```python @@ -1439,6 +1510,20 @@ In Taiwan stock technical data, we have 20 datasets, as follows: - Data range: 2005-01-01 ~ now +??? note "The disclosure interval varies by era — early data is not one row every 5 seconds" + Despite the "every 5 seconds" name, **TWSE raised its disclosure frequency over the years**, so earlier data comes at a coarser interval. Rows per trading day: + + | Period | Disclosure interval | Rows per trading day | + |---|---|---| + | 2005-01-03 ~ 2011-01-14 | 60 seconds | 271 | + | 2011-01-17 ~ 2014-02-21 | 15 seconds | 1,081 | + | 2014-02-24 ~ 2014-12-27 | 10 seconds | 1,621 | + | 2014-12-29 ~ now | 5 seconds | 3,241 | + + This is **the granularity published at the source at the time, not a gap in this table**: TWSE's own publication for those dates carries the same number of rows (2005-09-28, for instance, is 271 rows officially). + + **Practical advice**: do not assume 3,241 rows on every trading day. When computing returns or volatility across eras, or aligning time series, resample on the day's actual `date` column (which carries the time) rather than slicing by a fixed row count or positional index. + !!! example === "Package" ```python @@ -2178,6 +2263,20 @@ In Taiwan stock technical data, we have 20 datasets, as follows: - Data range: 2005-01-03 ~ now +??? note "The disclosure interval varies by era — early data is not one row every 5 seconds" + Despite the "every 5 seconds" name, **TWSE raised its disclosure frequency over the years**, so earlier data comes at a coarser interval. Rows per trading day, per index: + + | Period | Disclosure interval | Rows per trading day (per index) | + |---|---|---| + | 2005-01-03 ~ 2011-01-14 | 60 seconds | 271 | + | 2011-01-17 ~ 2014-02-21 | 15 seconds | 1,081 | + | 2014-02-24 ~ 2014-12-27 | 10 seconds | 1,621 | + | 2014-12-29 ~ now | 5 seconds | 3,241 | + + This is **the granularity published at the source at the time, not a gap in this table**: TWSE's own publication for those dates carries the same number of rows (2005-09-28, for instance, is 271 rows officially). + + **Practical advice**: do not assume 3,241 rows per index on every trading day. When aligning time series across eras or computing index momentum, resample on the day's actual `time` column rather than slicing by a fixed row count or positional index. + !!! example === "Package" ```python diff --git a/docs/tutor/TaiwanMarket/Technical.md b/docs/tutor/TaiwanMarket/Technical.md index 7fc3e4d..0b0aa7f 100644 --- a/docs/tutor/TaiwanMarket/Technical.md +++ b/docs/tutor/TaiwanMarket/Technical.md @@ -346,6 +346,49 @@ **與 `TaiwanStockPriceAdj` 的差異**：還原股價表在這類日期**不會**是 `0`，而是沿用前一交易日的還原價（`Trading_Volume` 仍為當日實際值）。例如 2317 鴻海 2025-07-30 在 `TaiwanStockPriceAdj` 為 `open=169.94`、`max=170.43`、`min=166.06`、`close=166.54`，即 2025-07-29 的還原價。若您需要「無價日不留 `0`」的序列，可改用 `TaiwanStockPriceAdj`；若需要辨識哪些日子原本無公告價格，則以 `TaiwanStockPrice` 的 `0` 為準。 +??? note "2005 年以前的上市（TWSE）資料：已下市證券的歷史可能不完整，長期回測請留意倖存者偏誤" + 本表**上市（TWSE）**部分在 **2005 年以前**所涵蓋的個股母體，以「之後仍持續交易的證券」為主。**在當時掛牌交易、但後來已下市／終止上市的證券**，其該段期間的歷史可能未被完整收錄，因此實際涵蓋的檔數會少於當時的實際掛牌檔數。 + + 此外，部分個股在本表的歷史起始日會**晚於實際掛牌日**。例如 2311、2325 在本表最早只到 `2004-02-11`；2319 則整段查無資料。 + + **對回測的影響**：若直接以本表 2005 年以前的上市資料建構選股池、產業母體或自建指數，樣本會偏向「存活下來的」證券，因而**系統性高估報酬、低估風險**（survivorship bias）。建議： + + - 將長期回測的起點設在 **2005 年以後**；或 + - 若必須涵蓋更早的期間，請在方法與結論中明確標註「2005 年以前的上市樣本並非當時的完整母體」這項限制，避免把偏誤誤讀成策略績效。 + + **上櫃（TPEx）不受此影響**，2005 年以前的上櫃母體正常。由本表衍生的 `TaiwanStockPriceAdj`（還原股價）亦適用同一項說明。 + +??? note "上櫃（TPEx）2007-01-02 ~ 2007-04-20 缺少個股日成交資訊（來源端即無揭示）" + 本表**上櫃（TPEx）**部分，在 **2007-01-02 ~ 2007-04-20** 這 71 個交易日沒有個股日成交資訊。該期間櫃買中心並未提供對應的個股日行情揭示，因此**不是抓取遺漏，而是來源端的資料斷層**。 + + **可觀察到的現象**：2006-12-29 與 2007-04-23 兩天各約有 400 檔上櫃個股有資料，中間這 71 個交易日則僅剩零星幾檔（多為當時屬其他板別、之後才轉上櫃的證券）。**同期間的上市（TWSE）資料正常。** + + **實務建議**：若您的分析區間涵蓋 2007 年上半年，請把這 71 個交易日的上櫃資料視為**缺值**而非「當日無成交」。在計算報酬率、換手率或市場寬度時直接當成 0 會使結果失真；建議改以上市資料為主，或在該區間跳過上櫃樣本。 + +??? note "不帶 `data_id` 取整日資料時，`stock_id` 會包含指數代號（非個股）" + 以「一次拿特定日期，所有資料」的方式查詢（不帶 `data_id`）時，回傳結果除了個股、ETF、權證之外，還會包含 **30 餘筆指數列**（目前為 32 個代號）。這些列的 `stock_id` 是**英文指數名稱**而非數字代號，例如： + + - 大盤指數：`TAIEX`（發行量加權股價指數）、`TPEx`（櫃買加權股價指數） + - 產業類指數：`Cement`、`Food`、`Plastics`、`Semiconductor`、`FinancialInsurance`、`Optoelectronic` 等 30 個 + + 指數列的 `open` / `max` / `min` / `close` 是**指數點數**（例如 `TAIEX` 2025-07-01 的 `close=22553.72`），`Trading_Volume` / `Trading_money` 則是該類股當日的成交量值。若下游流程把 `stock_id` 一律當成個股代號使用（例如再逐檔去查分 K、籌碼或財報），這些指數列就會被誤收而查不到資料。 + + **兩種過濾方式**： + + ```python + # 方式一：只保留以數字開頭的代號（個股、ETF、權證） + df = df[df["stock_id"].str[0].str.isdigit()] + + # 方式二：以 TaiwanStockInfo 的 industry_category 排除 + stock_info = api.taiwan_stock_info() + index_ids = stock_info.loc[ + stock_info["industry_category"].isin(["大盤", "Index"]), "stock_id" + ] + df = df[~df["stock_id"].isin(index_ids)] + ``` + + 反之，若您**只想要**大盤與類股指數，把上述條件反過來即可。 + !!! example === "Package" ```python @@ -1360,6 +1403,34 @@ - 資料區間：2005-01-01 ~ now +??? note "揭示頻率隨年代不同，早期並非每 5 秒一筆" + 本表雖名為「每 5 秒」，但**證交所的揭示頻率是逐年提高的**，早期資料的時間間隔較長。以單一交易日的筆數來看： + + | 期間 | 揭示間隔 | 每個交易日筆數 | + |---|---|---| + | 2005-01-03 ~ 2011-01-14 | 60 秒 | 271 | + | 2011-01-17 ~ 2014-02-21 | 15 秒 | 1,081 | + | 2014-02-24 ~ 2014-12-27 | 10 秒 | 1,621 | + | 2014-12-29 ~ now | 5 秒 | 3,241 | + + 這是**來源端當時的揭示粒度，不是本表的缺漏**：證交所對應日期的原始揭示同樣只有這些筆數（例如 2005-09-28 官方即為 271 筆）。 + + **實務建議**：請勿假設每個交易日都有 3,241 筆。跨年代做時間序列對齊、或要計算固定 5 秒區間的增量時，請依當日實際的 `Time` 欄位重新取樣（resample），不要用固定筆數或位置索引來切資料。 + +??? note "2023-08-04 開盤 `09:00:00` 另有一列沿用前一交易日收盤的累計值" + 本表各欄位是**當日累計值**（累計委託筆數／數量、累計成交筆數／數量／金額），正常情況下每日 `09:00:00` 由接近 0 起算。但 **2023-08-04** 的 `09:00:00` 除了正常的開盤列之外，另有一列的七個數值欄與**前一交易日 2023-08-02 收盤 `13:30:00`** 完全相同（當週 08-03 因颱風停止交易）。**證交所該日的原始檔案本身即為此值**，並非本表處理錯誤。 + + **實務影響**：若您直接對本表做差分（`diff()`）來還原每 5 秒的增量，當日第一根 bar 會出現極大的負值。 + + 這在 5 秒揭示年代（2014-12-29 起）的抽查中僅見此一例，其餘交易日未見同樣情形。差分前可先以「同一時間戳只保留累計值最小的那一列」還原真正的開盤列： + + ```python + df = df.sort_values(["date", "Time"]) + # 同一時間戳有多列時，只留累計值最小者（真正的開盤列） + df = df.loc[df.groupby(["date", "Time"])["TotalDealVolume"].idxmin()] + inc = df.groupby("date")[["TotalDealVolume", "TotalDealMoney"]].diff() + ``` + !!! example === "Package" ```python @@ -1439,6 +1510,20 @@ - 資料區間：2005-01-01 ~ now +??? note "揭示頻率隨年代不同，早期並非每 5 秒一筆" + 本表雖名為「每 5 秒」，但**證交所的揭示頻率是逐年提高的**，早期資料的時間間隔較長。以單一交易日的筆數來看： + + | 期間 | 揭示間隔 | 每個交易日筆數 | + |---|---|---| + | 2005-01-03 ~ 2011-01-14 | 60 秒 | 271 | + | 2011-01-17 ~ 2014-02-21 | 15 秒 | 1,081 | + | 2014-02-24 ~ 2014-12-27 | 10 秒 | 1,621 | + | 2014-12-29 ~ now | 5 秒 | 3,241 | + + 這是**來源端當時的揭示粒度，不是本表的缺漏**：證交所對應日期的原始揭示同樣只有這些筆數（例如 2005-09-28 官方即為 271 筆）。 + + **實務建議**：請勿假設每個交易日都有 3,241 筆。跨年代計算報酬率、波動度或做時間序列對齊時，請依當日實際的 `date`（含時間）欄位重新取樣（resample），不要用固定筆數或位置索引來切資料。 + !!! example === "Package" ```python @@ -2178,6 +2263,20 @@ - 資料區間：2005-01-03 ~ now +??? note "揭示頻率隨年代不同，早期並非每 5 秒一筆" + 本表雖名為「每 5 秒」，但**證交所的揭示頻率是逐年提高的**，早期資料的時間間隔較長。以單一交易日、單一指數的筆數來看： + + | 期間 | 揭示間隔 | 每個交易日筆數（每個指數） | + |---|---|---| + | 2005-01-03 ~ 2011-01-14 | 60 秒 | 271 | + | 2011-01-17 ~ 2014-02-21 | 15 秒 | 1,081 | + | 2014-02-24 ~ 2014-12-27 | 10 秒 | 1,621 | + | 2014-12-29 ~ now | 5 秒 | 3,241 | + + 這是**來源端當時的揭示粒度，不是本表的缺漏**：證交所對應日期的原始揭示同樣只有這些筆數（例如 2005-09-28 官方即為 271 筆）。 + + **實務建議**：請勿假設每個交易日、每個指數都有 3,241 筆。跨年代做時間序列對齊或計算指數動能時，請依當日實際的 `time` 欄位重新取樣（resample），不要用固定筆數或位置索引來切資料。 + !!! example === "Package" ```python