gahow/quant-trading
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: agent-reviewed overnight_recovery_agent_top10; guard monitor hot-reload

Browse Source 
overnight_recovery_agent_top10 (strategies/agent_review.py) re-ranks the
overnight_recovery top 50 with an LLM agent's daily scores:
final = quant percentile + 0.25 x score (-2..+2), vetoes for five event types,
top 10 on the base strategy's rebalance rows and hysteresis. Score files are
stamped by `agent_review.py submit` and steer a session only if stamped before
its 16:00 New York close; without one the strategy holds exactly what
overnight_recovery_top10 holds (checked on the live panel 2016-2026).
- agent_review.py: candidates / submit / evaluate; skill
  .claude/skills/agent-stock-review drives the daily research.
- US only, skipped by precompute (no history before its first score file);
  runs in the original book and in selected_2026_09 (from the book's start).
- .gitignore tracks .claude/skills/ and ignores data/agent/.

Monitor fix: hot-reload re-imports trader.py but not strategies/*. After the
2026-09-29 strategy commits the stale modules made every phase's panel refresh
fail ("Unknown signal 'low_risk'") and US books skipped 09-29 until a restart.
_reload_strategies now builds every factory and keeps the old registry when
any fails; CLAUDE.md says to restart the monitor after changing strategy code.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Gahow Wang
2026-09-30 20:25:20 +08:00
parent ab3acfe4cb
commit 761725d2c8
9 changed files with 676 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
CLAUDE.md
View File

@@ -38,6 +38,11 @@ uv run python trader.py log --market us --strategy sim_recovery_mom_top10 --star
uv run python trader.py simulate --integer-shares --fixed-fee 2.0 ... # Integer shares mode
uv run python trader.py compare --market us --book selected_2026_09 # A named live book
# Agent-reviewed strategy overnight_recovery_agent_top10: daily loop (skill .claude/skills/agent-stock-review)
uv run python agent_review.py candidates # after the US close: top-50 pack -> data/agent/us/candidates_<as_of>.json
uv run python agent_review.py submit data/agent/us/draft_<as_of>.json # before 16:00 New York of the next session
uv run python agent_review.py evaluate # score IC, veto outcomes, book vs overnight_recovery_top10
# Precompute the strategy page's backtests (writes data/precomputed/)
uv run python precompute_backtest.py
@@ -67,6 +72,14 @@ PYTHONPATH=. uv run python -m research.lab.fetch_lazy_prices && PYTHONPATH=. uv
PYTHONPATH=. uv run --with numba python -m research.lab.earnings # earnings-announcement return (needs sec_8k_earnings.csv)
PYTHONPATH=. uv run --with numba python -m research.lab.industry # industry momentum on SIC groups
PYTHONPATH=. uv run python -m research.lab.fetch_us_mid holdings|prices|status # PIT S&P 400/600 from iShares holdings (research/lab/panel_mid.py)
PYTHONPATH=. uv run python -m research.lab.review_live_us # Apr-Sep 2026 replay review: attribution, AI-chain stages, missed winners
PYTHONPATH=. uv run python -m research.lab.fetch_supply_links [--universe mid] # 10-K customer/supplier/competitor mentions (sec_10k_links.csv)
PYTHONPATH=. uv run --with numba python -m research.lab.lead_lag --part all # economic-link momentum, lead-lag network, earnings spillover
PYTHONPATH=. uv run --with numba python -m research.lab.theme_propagation # does a hot industry's linked laggard catch up?
PYTHONPATH=. uv run --with numba python -m research.lab.lead_lag_small # the same links on S&P 400/600, 2019-26
PYTHONPATH=. uv run python -m research.lab.sector_cycle # business-cycle rotation, sector momentum/reversal, semis cycle
uv run python -m research.lab.fetch_wiki_attention && PYTHONPATH=. uv run --with numba python -m research.lab.attention # Wikipedia-pageview attention
PYTHONPATH=. uv run --with numba python -m research.lab.concentration # industry caps on the momentum sleeve
# Setup
uv sync # Install/sync dependencies
@@ -91,9 +104,9 @@ uv run python -m unittest discover tests
**Backtest engine** (`main.py`): Orchestrates data loading, strategy execution, and visualization. `backtest()` has two modes. Close mode (no `open_data`, used by `main.py` and old research scripts) is vectorized and trades at the close the signal saw. Live mode (`open_data` given, used by the web UI through `backtest_runner.py`) keeps the book like `trader.py`: same signal timing (`strategy_positions`), holdings that drift, the $50/0.1% order floor, `trader.RESIZE_BAND`, fees on the orders placed (`simulate_live_book`).
**Daily trader** (`trader.py`): Live/forward-testing system with persistent portfolio state in `data/trader_{market}_{strategy}.json`. The `auto` subcommand runs both signal generation and execution in a single invocation — designed for cron. The `simulate` subcommand replays a date range day-by-day with realistic portfolio tracking (fractional shares, cash, commissions). Orders: whole A-share lots are allocated nearest-lot within cash (`_allocate_lots`); a held name is resized only when ≥ `RESIZE_BAND` (10%) off target. Available strategies: `recovery_mom_top10`, `recovery_mom_top20`, `momentum`, `momentum_quality`, `dual_momentum`, `inverse_vol`, `trend_following`, `buy_and_hold`, 32 factor-combo strategies (`fc_{signal}_{freq}` — see `strategies/factor_combo.py`), the robust-momentum bar strategies `overnight_recovery_top10`, `lowrisk_momentum_top15`, `lowbeta_sharpe_mom_top20`, `lowrisk_top20`, and the US sleeve portfolios `us_three_sleeve` and `us_momentum_qqq_trend`.
**Daily trader** (`trader.py`): Live/forward-testing system with persistent portfolio state in `data/trader_{market}_{strategy}.json`. The `auto` subcommand runs both signal generation and execution in a single invocation — designed for cron. The `simulate` subcommand replays a date range day-by-day with realistic portfolio tracking (fractional shares, cash, commissions). Orders: whole A-share lots are allocated nearest-lot within cash (`_allocate_lots`); a held name is resized only when ≥ `RESIZE_BAND` (10%) off target. Available strategies: `recovery_mom_top10`, `recovery_mom_top20`, `momentum`, `momentum_quality`, `dual_momentum`, `inverse_vol`, `trend_following`, `buy_and_hold`, 32 factor-combo strategies (`fc_{signal}_{freq}` — see `strategies/factor_combo.py`), the robust-momentum bar strategies `overnight_recovery_top10`, `lowrisk_momentum_top15`, `lowbeta_sharpe_mom_top20`, `lowrisk_top20`, the US sleeve portfolios `us_three_sleeve` and `us_momentum_qqq_trend`, and the agent-reviewed `overnight_recovery_agent_top10` (US only, forward only).
**Live books** (`trader.LIVE_BOOKS`): a book is a cohort of strategies, each with its own account, started on one date with one capital and a flat fee per order. The original book (every live strategy, `MONITOR_CAPITAL`, `MARKET_FEES`) keeps state in `data/trader_*.json`; a named book keeps it in `data/books/<book>/`. The monitor runs every book; `auto`/`status`/`log`/`compare`/`replay` and `bridge.py live` take `--book`, and `bridge.py books` lists them for the UI's book selector. `selected_2026_09` (29 strategies, US $10,000 / CN ¥100,000, $2 / ¥10 per order) starts on 2026-09-25 (US) and 2026-09-28 (CN).
**Live books** (`trader.LIVE_BOOKS`): a book is a cohort of strategies, each with its own account, started on one date with one capital and a flat fee per order. The original book (every live strategy, `MONITOR_CAPITAL`, `MARKET_FEES`) keeps state in `data/trader_*.json`; a named book keeps it in `data/books/<book>/`. The monitor runs every book; `auto`/`status`/`log`/`compare`/`replay` and `bridge.py live` take `--book`, and `bridge.py books` lists them for the UI's book selector. `selected_2026_09` (30 strategies, US $10,000 / CN ¥100,000, $2 / ¥10 per order) starts on 2026-09-25 (US) and 2026-09-28 (CN); `overnight_recovery_agent_top10` joined it on 2026-09-30 and starts with the book.
**Strategy protocol** (`strategies/base.py`): All strategies inherit from `Strategy` ABC and implement `generate_signals(data) → DataFrame` where the returned DataFrame contains portfolio weights (rows = dates, columns = assets, values sum to ~1.0 per row). Each strategy is responsible for applying its own 1-day lag via `.shift(1)` to avoid lookahead bias — the backtest engine does not shift. A `BarStrategy` declares `fields` (close/open/high/low/volume) and receives `{field: DataFrame}`; live it uses completed bars only (row d from bars through d-1, traded at the close of d), while close-price strategies read the session's open as their latest price.
@@ -111,6 +124,7 @@ uv run python -m unittest discover tests
- `robust_momentum.py` — Bar strategies from the 2026-09 PIT research (report: https://claude.ai/artifact/Mo9nrzs5SWj5FfyshERyWS): US `overnight_recovery` (overnight-return momentum + distance from 126d low), CN `lowrisk_momentum` (amplitude-cut momentum + low beta + price-volume corr + momentum/vol) and close-only `lowbeta_sharpe_mom`. Top-N with rank hysteresis, rebalance every 10 days, weights drift between rebalances. `low_risk` (US `lowrisk_top20`: idiosyncratic vol + low beta, top 20, monthly) is the defensive sleeve of the 2026-09-29 US research (report: https://claude.ai/artifact/VGhZvKkS9UBYZkLeqGjcNC).
- `sleeve_portfolio.py` — Several strategies in one account with fixed capital shares. `us_three_sleeve` = 1/3 each of `overnight_recovery_top10`, `lowrisk_top20`, `trend_rider_v7_vt24` (a mixed stock+ETF strategy); sleeve shares drift with their returns and reset at a month start once any is >25% off. 2011-2026 research: Sharpe 1.21 vs 0.88-1.03 for the sleeves alone. Research finding behind it: choosing factors by their recent record (rotation, factor momentum, last year's winner) did not beat a fixed composite; construction knobs (N, weighting, dynamic N) barely matter. On synthetic 3x ETFs from 1990 (`research/lab/letf_history.py`), TrendRider V7's 2016-2026 record does not hold: 1990-99 Sharpe 0.2-0.3, 2003-07 negative, 1990-2026 about 0.7 with MaxDD -58% (vt24) to -69% (vt36).
- `qqq_trend.py` — `QQQVolTrend`: QQQ exposure = 1{QQQ > 200-day average} × min(1.5, 30% / EWMA vol), held as TQQQ (E/3) + SHY. 1990-2026 on synthetic ETFs: 16.0% vs QQQ 14.9%, MaxDD -37% vs -83%; 2011-2026 on real ETFs it trails QQQ (17.5% vs 19.1%) with a smaller drawdown. `us_momentum_qqq_trend` (in `sleeve_portfolio.py`) holds it half and half with `overnight_recovery_top10`. Research behind it (`research/lab/qqq_scan.py`, `beat_qqq.py`): measured as alpha against QQQ with a 2006-16 design / 2017-26 holdout split, momentum top 10 is the only stock selection whose alpha held (+10%/yr, t 1.4); single factors, walk-forward ML, a mega-cap universe, SEC insider trades, 10-K/10-Q text similarity, earnings-announcement returns, industry momentum and a PIT S&P 400/600 universe (`panel_mid.py`) showed none. In the web engine (2016-09 to 2026-09, $2/order) `us_momentum_qqq_trend` earns 21.0% vs QQQ 20.9%, MaxDD -32%.
- `agent_review.py` — `overnight_recovery_agent_top10`: overnight_recovery's top 50, re-ranked each day by an LLM agent's scores (−2..+2 from filings, fundamentals, demand-chain logic, event risk; `final = quant percentile + 0.25 × score`), vetoes for five event types, top 10 on the base strategy's rebalance rows and hysteresis. Score files in `data/agent/us/` are stamped by `agent_review.py submit` and steer a session only if stamped before its 16:00 New York close; without one the strategy holds exactly what `overnight_recovery_top10` holds, so the two books' difference measures the agent. `AGENT_REVIEW_STRATEGIES` in trader.py keeps it US-only and out of `precompute_backtest.py` (it has no history before its first score file). Research behind the design (2026-09-30): picking 10 of the quant top 50 at random earns 13.4%/yr vs 26.2% for the top 10 (2006-2026), and an LLM cannot be backtested before its training cutoff, so the record starts 2026-09-29.
- `factor_combo.py` — Configurable factor-combination strategies with daily/weekly/biweekly/monthly rebalancing. US champions: `rec_mfilt+deep_upvol`, `ma200+mom7m+rec126`, `rec_mfilt+ma200`, `mom7m+rec126`. CN champions: `up_cap+quality_mom`, `down_resil+qual_mom`, `rec63+mom_gap`, `up_cap+mom_gap` — these were mined on today's CSI 300 members; on PIT members they earn ~3-7% a year (2016-2026). All registered in trader.py as `fc_{signal}_{freq}` (e.g., `fc_rec_mfilt_deep_upvol_monthly`). 32 new strategies total.
**Metrics** (`metrics.py`): Standalone functions for portfolio analytics (Sharpe, Sortino, Calmar, max drawdown, etc.). `summary()` prints a formatted report and returns a dict.
@@ -140,7 +154,7 @@ systemctl --user enable --now quant-trading-monitor quant-trading-api quant-trad
journalctl --user -u quant-trading-monitor -f
```
The monitor catches up on startup, so restarting it between phases is safe. If it starts mid-day (after open, before close), it falls back to `auto` mode for the evening phase.
The monitor catches up on startup, so restarting it between phases is safe. Its hot-reload re-imports trader.py only, not `strategies/`: after changing strategy code, restart the monitor (a registry whose strategies fail to build is rejected and logged with "restart the monitor"). If it starts mid-day (after open, before close), it falls back to `auto` mode for the evening phase.
Compare live books after a while: