skills/jp-stock-move-reason/references/pts-turnover-ranking.md

12 KB · Oct 2, 2026 · 00:33 UTC

# Japanese Stock Mover Estimated Trading Value Ranking Sub-skill

Use this sub-skill when the user asks for current Japanese-stock mover Top10
lists or Kabutan PTS mover lists ranked by estimated trading value, especially
phrases such as:

- `当前日股涨跌Top10`
- `现在的涨跌榜`
- `PTS夜间涨跌榜推定成交额Top10`
- `涨跌幅1%以上的推定成交额top10`
- `不是成交量，是成交额`

## Core Rule

The ranking pages provide current price and volume, but the workflow does not
have every execution price. Compute an estimate:

```text
regular session: estimated_trading_value = 当前价 * 出来高
PTS session: estimated_trading_value = PTS株価 * PTS出来高
```

Default screening is applied independently to the increase and decrease sides:

```text
abs(騰落率) >= 3%
出来高 > 2000
```

If one side has fewer than five qualified rows, re-fetch only that side using
`abs(騰落率) >= 1%` while retaining `出来高 > 2000`. Always rank by estimated
trading value. Report the effective threshold and the side-specific fallback in
the final output.

Call this `推定成交额` in Chinese and `売買代金推定` in Japanese. Do not call
it actual `成交额` / `売買代金`, because actual trading value is the sum of
each execution price times each execution quantity. Rank by the estimate, not
by page order and not by volume alone.

`出来高` is an eligibility filter, not the default ranking key. For generic
requests such as `看看当前涨跌Top10`, `再跑一下PTS`, or colloquial mentions of
`成交量` within this established workflow, preserve `推定成交额` ranking. Do
not silently sort by raw volume. Use raw-volume ranking only when the user
explicitly asks for `按出来高排序` or an equivalent unambiguous instruction.

## Sources

Use the current Yahoo Finance Japan regular rankings during cash-market hours
and the current Kabutan PTS warning pages in PTS windows. Read 50 rows per page.
Kabutan uses the `shared_perpage=50` cookie:

- Regular increase: `https://finance.yahoo.co.jp/stocks/ranking/up?market=all`
- Regular decrease: `https://finance.yahoo.co.jp/stocks/ranking/down?market=all`
- Night increase: `https://kabutan.jp/warning/pts_night_price_increase`
- Night decrease: `https://kabutan.jp/warning/pts_night_price_decrease`
- Day increase: `https://kabutan.jp/warning/pts_day_price_increase`
- Day decrease: `https://kabutan.jp/warning/pts_day_price_decrease`

## Section Selection

Select by JST clock and the verified exchange trading calendar:

- `09:00-11:30`: use Yahoo regular increase/decrease pages.
- `11:30-12:30`: use PTS day-section pages.
- `12:30-15:30`: use Yahoo regular increase/decrease pages.
- `15:30-17:00`: use PTS day-section pages.
- `08:00-09:00`: use PTS day-section pages.
- All other times, weekends, and known non-trading days: use PTS night-section
pages.

Check exchange holidays as well as weekends; on non-trading days use the night section and label its actual source date.

Equivalent routing logic, always evaluated in JST:

```text
if non_trading_day:
section = pts_night
elif 09:00 <= JST < 11:30 or 12:30 <= JST < 15:30:
section = regular
elif 08:00 <= JST < 09:00 or 11:30 <= JST < 12:30 or 15:30 <= JST < 17:00:
section = pts_day
else:
section = pts_night
```

Treat each boundary as start-inclusive and end-exclusive. For example, `11:30`
switches to PTS day, `12:30` switches back to the regular pages, and `15:30`
switches to PTS day.

For Yahoo regular pages, the percentage change is against the prior regular
close. Yahoo states that Tokyo Stock Exchange transaction prices are real time,
but all-market volume and reported trading value are delayed by at least 15
minutes. Since this workflow multiplies the displayed current price by displayed
volume, regular-session `推定成交额` can combine a real-time price with delayed
volume. Report that limitation instead of describing the estimate as tick-level
or exchange-reported trading value.
For day-session PTS, remember Kabutan compares pre-close prints against the
previous regular-session close and after-close prints against the same-day
regular-session close.

For night-session PTS, compare against the same-day regular-session close.

## Pagination Stop Rule

Yahoo regular rankings and Kabutan PTS rankings are sorted by percentage
change. Fetch 50-row pages until the last row crosses the requested threshold:

- Increase pages: stop after the last row is below `+min_abs_pct`.
- Decrease pages: stop after the last row is above `-min_abs_pct`.
- Stop earlier if a page is empty.

Do not assume page 1 is enough unless its final row already crosses the
threshold. Day-session decrease lists can be many pages during broad selloffs.
The volume filter is applied after parsing rows; it does not change the
percentage-based pagination stop rule.

## Fetch Discipline

- During regular cash-market hours, request Yahoo's current ranking pages and
parse the structured `window.__PRELOADED_STATE__` ranking data. Do not depend
on brittle visual table positions. Report the displayed
`YYYY/MM/DD HH:MM` update time.
- For PTS, request current Kabutan warning pages with the
`shared_perpage=50` cookie. Add a cache-busting timestamp, parse the displayed
`YYYY年MM月DD日 HH:MM現在` stamp, and report it. Kabutan PTS data is normally
about 15 minutes delayed, so do not describe it as tick-level real time.
- Yahoo says Tokyo Stock Exchange transaction prices are real time, while
all-market volume is delayed by at least 15 minutes. State this regular-session
mixed-timestamp limitation wherever the estimate is presented.
- Fetch the complete percentage-qualified range before applying the volume
filter and `推定成交额` sort. Never take the first page or page order as the
Top10 unless the threshold stop condition proves that it is enough.
- Yahoo's live ranking can reorder while consecutive pages are being fetched,
so deduplicate across pages by ticker before ranking. If the same ticker
appears more than once, keep the row with the larger cumulative volume as the
newer observation.
- Stop on an empty page and keep a finite page cap. If the host returns rate
limits, DNS errors, timeouts, resets, or repeated empty responses, report the
observed failure and stop increasing request frequency.
- Add moderate randomized waits after more than three consecutive requests to
the same host. Do not parallel-burst Kabutan, Yahoo, or message-board pages.
- Refresh the ranking once before the final answer. If reason collection took
long enough for membership to change, collect only newly entered codes and
reuse already verified same-turn reasons for unchanged names.

## Cause Analysis Workflow

After ranking:

1. Deduplicate the selected codes across increase/decrease lists.
2. Unless the user explicitly asks for a raw list only or says reasons are not
needed, final mover answers must include reasons. Do not stop at a bare
ranking table.
3. Collect reasons from Yahoo 掲示板 first for the selected `推定成交额` Top
names in one sequential loop. Fetch one page per code only and cache at most
the latest 100 comments per code. Count raw comments within 24 hours before
applying the likes filter. If fewer than 100 are within 24 hours, expand the
candidate window to 72 hours using only that same cache; never fetch comment
101 or later. Discard posts with fewer than five likes, score by recency,
likes, full body length, and company-material keywords, deduplicate exact
normalized-prefix signatures to a maximum 20-comment full-text shortlist,
reorder it by time and likes, and pass only `recent_comments[:5]` to Codex for
reason judgment. Do not
use Kabutan/Traders as the first-pass substitute for the board discussion.
The score is capped at 18: recency contributes 5/4/2/1 for `<=6h`, `<=24h`,
`<=48h`, and `>48h`; length contributes 3/2/1 for 30-300, over 300, and
10-29 characters; likes contribute 4/3/2/1 for 100+, 50-99, 20-49, and
5-19; company-material keyword hits contribute at most six. AI,
semiconductors, defense, drones, and other generic sector words do not score.
Sort by score, timestamp, and likes before 60-character normalized-signature
deduplication. After taking 20, re-sort by timestamp and likes before `[:5]`.
The signature is lowercase text with all whitespace and
`、。！？ ! ? , . ・ … 「」 『』 （） () [] 【】` removed, truncated to its first
60 characters. This is exact matching, not semantic similarity. Keep the
earlier item in score/timestamp/likes order. Equal prefixes collapse even if
later text differs; small prefix differences survive. Do not Unicode-normalize
widths or explicitly remove emoji, URLs, or usernames. Scope deduplication to
one stock's current collection.

5. Never fetch more than one forum page for the same code or repeat a Yahoo
forum fetch for a code in the same turn. The `--forum-only` mode avoids
quote, news, Kabutan, and Traders requests.
6. Treat the board as the primary explanation layer for ranking requests.
Use concrete news, disclosures, earnings, guidance, ratings, orders,
buybacks, lawsuits, capital actions, or shareholder benefits only to verify
an event asserted in board discussion; do not replace the board with them.
Wait a randomized 1-3 seconds between Yahoo requests.
On HTTP 403/429, access-denied content, reset, or abnormal empty output, stop
all Yahoo requests for the rest of the turn; do not retry immediately.
The collector additionally enforces a shared cross-process randomized 1-3 second Yahoo
host gap and a 30-minute cooldown after 403/429 or access-control content.
Never delete or bypass the cooldown for a ranking request.
7. For ETF/ETN rows, explain them by the underlying index or strategy instead
of forcing single-stock news. Examples:
- Nikkei inverse ETFs rise when Nikkei falls.
- Nikkei leveraged ETFs fall when Nikkei falls.
- S&P 500 income/covered-call ETFs may move on the underlying index, FX, and
thin PTS prints.
8. For rows with very small `推定成交额`, explicitly mark the reason as
low-confidence if no concrete news/disclosure exists. Thin prints can jump
several percent with little actual capital committed.

## Final Answer Pattern

Report the timestamp, source, effective threshold, and method before the tables. During the regular
cash session use:

```text
口径：Yahoo即时涨跌幅榜，YYYY-MM-DD HH:MM JST；
先筛 abs(涨跌幅) >= 3% 且 出来高 > 2,000；单侧不足5只时仅该侧回退到1%，
再按 当前价 × 出来高 算推定成交额取Top10。
东证取引值实时，但Yahoo出来高最低延迟15分钟。
```

During a PTS session use:

```text
口径：Kabutan 夜间PTS，YYYY-MM-DD HH:MM JST；
先筛 abs(PTS涨跌幅) >= 3% 且 出来高 > 2,000；单侧不足5只时仅该侧回退到1%，
再按 PTS价格 × PTS出来高 算推定成交额取Top10。
```

Use compact tables. During the regular cash session:

```text
上涨 Top10（涨幅大于3%/成交量大于2000/推定成交额排序）
| 排名 | 代码 | 名称 | 涨跌幅 | 出来高 | 推定成交额 | 原因 |

下跌 Top10（跌幅大于3%/成交量大于2000/推定成交额排序）
| 排名 | 代码 | 名称 | 涨跌幅 | 出来高 | 推定成交额 | 原因 |
```

During PTS sessions:

```text
PTS上涨 Top10（涨幅大于3%/成交量大于2000/推定成交额排序）
| 排名 | 代码 | 名称 | PTS涨跌幅 | 出来高 | 推定成交额 | 原因 |

PTS下跌 Top10（跌幅大于3%/成交量大于2000/推定成交额排序）
| 排名 | 代码 | 名称 | PTS涨跌幅 | 出来高 | 推定成交额 | 原因 |
```

Use these headings exactly. Report `日中` or `夜间` in the timestamp/method line
below the heading rather than changing the heading text.

Never omit the `原因` column in the final answer unless the user explicitly asks
for numbers only.

End with a quality note:

- Which names have hard catalysts such as earnings, guidance, buyback, large
order, lawsuit, capital action, or shareholder benefits.
- Which names are likely thin PTS jumps because `推定成交额` is tiny.
- Whether ETF/ETN rows were included or excluded.

SHA-256: aa8d5d414f9784e58300fb937a62ecab3545bfdc2bcc95e882539fae0103d8f6