geometrybase
943eab27cb
Request: - Add historical data around T0.5 margin heat and rebuild the model. Changes: - Add external_ipo_history to store third-party historical IPO records separately from true T0.5 market-heat snapshots. - Add scripts/archive_ipohk_history.py to archive ipohk structured listed IPO history. - Archive 807 ipohk rows, including final oversubscription, one-lot win rate, grey-market return, and first-day return where available. - Extend the v0 analysis dataset with true T0.5 market-heat columns and separate external final-heat columns. - Rebuild reports/2026-06-15_analysis_model_v0.md with T0.5 coverage and external final-heat calibration. - Add a Chinese report explaining why historical final oversubscription cannot be treated as T0.5 margin snapshots. - Update analyst and archivist skills to keep T0.5 and external final history separate. Verification: - .venv/bin/python -m py_compile scripts/build_analysis_dataset.py scripts/archive_ipohk_history.py scripts/archive_t0_5_market_heat.py - .venv/bin/python scripts/build_analysis_dataset.py --as-of 2026-06-15T19:20:00Z - Python sqlite3 PRAGMA integrity_check returned ok and foreign_key_check returned zero rows. - Confirmed 807 external_ipo_history rows, 792 rows with external final oversubscription, 5 true T0.5 market-heat rows, and 297 analysis dataset rows. - git diff --cached --check Next useful context: - True T0.5 historical backtesting still requires ongoing frozen margin-heat snapshots during each IPO subscription window.
98 lines
5.3 KiB
Markdown
98 lines
5.3 KiB
Markdown
# HK IPO Analysis Model v0
|
|
|
|
- Model version: `ipo_score_v0`
|
|
- Analysis as of: `2026-06-15T19:20:00Z`
|
|
- Rule file: `rules/ipo_score_v0.yaml`
|
|
- Dataset: `data/snapshots/analysis_model_v0_dataset.csv`
|
|
|
|
## What This Model Does
|
|
|
|
This is the first analyst model built from the downloaded archive. It creates a repeatable feature table, scores each IPO using stage-safe rules, and calibrates the score buckets against archived D1 sell outcomes. It is intentionally transparent: the output includes every score component and the archived source paths used for each ticker.
|
|
|
|
The model is built for a short IPO allocation trade: sell in T2 grey market when reliable executable data exists, or sell on D1 otherwise. It does not use grey-market data in v0 because T2 currently has no approved reproducible source. It also does not use post-listing returns as inputs; D1 is the primary sell label, while D5/D20/D60 are review labels only.
|
|
|
|
## Data Inventory
|
|
|
|
- IPO rows scored: 297
|
|
- Rows with D1 labels: 273
|
|
- Rows with structured T1 demand fields: 291
|
|
- Rows with prospectus source path: 297
|
|
- Rows with allotment source path: 291
|
|
- Rows with offer size: 297
|
|
- Rows with public oversubscription: 281
|
|
- Rows with international oversubscription: 277
|
|
- Rows with T0.5 margin heat snapshots: 5
|
|
- Rows with T0.5 margin heat and D1 labels: 0
|
|
- Rows matched to external ipohk history: 102
|
|
- Rows with external final oversubscription: 95
|
|
- Rows with external final oversubscription and D1 labels: 85
|
|
- Rows pending T1 structure: 6 (01392, 02335, 06067, 06106, 06132, 06675)
|
|
- T1 field-level blanks: public oversubscription 10, international oversubscription 14, valid applications 6, successful applications 18
|
|
|
|
## T0 Calibration
|
|
|
|
T0 uses only prospectus-stage structure: offer size, initial public offer percentage, minimum subscription amount, offer price band, and over-allotment availability.
|
|
|
|
| Bucket | N | D1 positive | D1 >= 10% | Avg D1 return | Median D1 return |
|
|
| --- | ---: | ---: | ---: | ---: | ---: |
|
|
| t0_1_to_4 | 60 | 63.3% | 40.0% | 9.6 | 3.1 |
|
|
| t0_5_to_7 | 105 | 73.3% | 51.4% | 40.1 | 13.2 |
|
|
| t0_gte_8 | 72 | 76.4% | 47.2% | 28.6 | 9.6 |
|
|
| t0_lt_1 | 36 | 58.3% | 33.3% | 12.8 | 2.3 |
|
|
|
|
## T1 Calibration
|
|
|
|
T1 adds allotment-stage demand: public subscription, international placing demand, valid application count, application success rate, and HK public offer reallocation.
|
|
|
|
| Bucket | N | D1 positive | D1 >= 10% | Avg D1 return | Median D1 return |
|
|
| --- | ---: | ---: | ---: | ---: | ---: |
|
|
| total_0_to_9 | 68 | 58.8% | 30.9% | 3.3 | 0.2 |
|
|
| total_10_to_17 | 29 | 55.2% | 34.5% | 13.9 | 1.5 |
|
|
| total_18_to_25 | 49 | 75.5% | 51.0% | 31.3 | 13.4 |
|
|
| total_gte_26 | 59 | 94.9% | 88.1% | 86.7 | 80.0 |
|
|
| total_lt_0 | 68 | 61.8% | 23.5% | 0.4 | 1.0 |
|
|
|
|
## T0.5 Market Heat
|
|
|
|
T0.5 uses archived subscription-period margin heat snapshots. These are non-official live signals and are kept separate from T1 allotment demand. The current archive is not yet a historical training set: it has too few rows and no D1 labels for calibration.
|
|
|
|
- T0.5 margin rows: 5
|
|
- T0.5 rows with D1 labels: 0
|
|
|
|
## External Final Heat Proxy
|
|
|
|
The ipohk history archive adds final public oversubscription, one-lot win rate, grey-market return, and first-day return where available. These fields are useful for coverage checks and post-hoc calibration, but they are not T0.5 inputs because they are final or near-final history.
|
|
|
|
- External history rows matched into this dataset: 102
|
|
- Matched rows with final oversubscription: 95
|
|
- Matched rows with final oversubscription and D1 labels: 85
|
|
|
|
| Bucket | N | D1 positive | D1 >= 10% | Avg D1 return | Median D1 return |
|
|
| --- | ---: | ---: | ---: | ---: | ---: |
|
|
| external_os_1000x_to_5000x | 33 | 93.9% | 78.8% | 60.4 | 44.2 |
|
|
| external_os_100x_to_1000x | 21 | 61.9% | 38.1% | 8.8 | 4.2 |
|
|
| external_os_10x_to_100x | 7 | 28.6% | 14.3% | -23.0 | -21.9 |
|
|
| external_os_gte_5000x | 18 | 83.3% | 72.2% | 101.7 | 89.7 |
|
|
| external_os_lt_10x | 6 | 50.0% | 16.7% | 4.7 | -4.1 |
|
|
|
|
## Current Read
|
|
|
|
After the T1 demand text backfill, the strongest v0 T1 bucket is `total_gte_26` with 59 historical D1 observations and a 94.9% D1 positive rate. The model is most useful after allotment results are available; T0 is a watchlist filter rather than a final subscription call.
|
|
|
|
The high-conviction bucket remains clearly differentiated, but the middle and low score buckets are still not monotonic. This refresh keeps the v0 score formula unchanged and updates empirical calibration only; future rule changes should come from reviewed prediction cards rather than overfitting this historical sample.
|
|
|
|
## Usage
|
|
|
|
1. Run `scripts/build_analysis_dataset.py` after archivist updates the database.
|
|
2. Use `t0_score` for prospectus-stage watchlisting.
|
|
3. Use `total_score`, `decision_band`, and `calibrated_d1_positive_rate` for T1-stage subscription cards.
|
|
4. Frame live decisions around a T2 or D1 sell, not long-term holding.
|
|
5. Treat D5/D20/D60 columns as review labels only, never as prediction inputs or holding targets.
|
|
|
|
## Known Gaps
|
|
|
|
- T1 is structurally complete for listed rows; residual field-level NULLs remain when the archived source does not explicitly state a demand field.
|
|
- Industry and issuer fundamentals are not sufficiently structured for model input.
|
|
- T2 grey-market signal is blocked pending an approved source.
|
|
- Extreme D1 returns should be audited before they drive rule changes.
|