Skip to content

China A-Shares: Sector Rotation & Hot Money (游资) — Which Models Detect the Effect, How to Monitor It

Research date: Aug 30, 2026. Follow-up research note on the short-horizon China A-share effect: sector/concept momentum over roughly 3–10 trading days driven by macro headline sentiment, policy releases, and hot-money (游资) capital concentration, plus the Dragon-Tiger List (龙虎榜) and order-flow-imbalance signals that precede the fast mean-reversion. Academic claims are cited inline; practitioner sources are labeled as directional experience, not peer-reviewed proof.

Companion posts: Daily Market Scan Guide, Market News Timeline — 2026-08-24, Daily Market Scan — 2026-08-01.

Short answer

Which models? No single model detects the whole effect. It decomposes into four sub-problems, each with a proven model family:

  1. "Which regime / rotation phase are we in?" — Gaussian Hidden Markov Models (HMM), Markov regime-switching, and XGBoost regime classifiers on volatility × rotation-speed states. These gate when the short-horizon momentum strategy is "on."
  2. "Which sector/concept for the next 1–3 weeks?" — cross-sectional economic-link momentum regressions (industry momentum at monthly, technology-link momentum at weekly), gradient-boosted trees (LightGBM/XGBoost), and feed-forward deep factors on sector/concept features.
  3. "Is hot money actually accumulating?" — Dragon-Tiger seat identity plus seat-network factors (co-appearance clusters, 接力 vs 一日游), and Level-2 order-flow imbalance (OFI/VPIN, iceberg/spoof detection).
  4. "What triggered it?" — news/policy NLP: news→industry/style mapping, retrieval-augmented generation (RAG) over a news corpus, and LLM multi-agent concept scanners.

How to monitor? A daily end-of-day pipeline: Dragon-Tiger List (published after the close) → seat identity & cumulative net-buy profiles → limit-up pool and the 7-dimension sentiment cycle (冰点→过热) → multi-source capital consensus (northbound × institutional × margin × block trades), plus an intraday Level-2 layer (order-flow imbalance, 封单/撤单) and a news-feed layer. The signal decays in days; anything beyond a ~2-week horizon should be treated as mean-reversion risk.

1. What the evidence actually says

1.1 Stock-level momentum fails; the horizon is short

The Chinese market shows short-run time-series momentum and long-run contrarian behavior, and the effect is highly dependent on look-back/holding windows and firm characteristics. Shi & Zhou's study of Chinese indices and all A-share stocks (1991–2015) finds significant time-series momentum at short look-back/holding horizons and contrarian effects at longer horizons; the short-horizon momentum returns are also the largest in magnitude (arXiv:1702.07374). This matches practitioner experience: whatever edge exists in A-shares lives in the first days-to-weeks after a move, then decays into reversal.

The most direct academic support for "sector/concept momentum works on short horizons" is "Economic Links and Stock Returns in Chinese A-Share Market" (Duan, Wang, Zhang, Journal of Financial Research 2022, 500(2): 171–188; PKU working copy here):

  • Industry momentum is significant at the monthly level.
  • Technology-link (概念/科技关联), geographic, and customer-chain momenta are significant at the weekly level; technology-link momentum is the most powerful when all economic links are controlled simultaneously.
  • Technology-link momentum predicts the focal firm's returns for 1–3 weeks; a long-short strategy earns 0.16% weekly (8.67% annualized).
  • The mechanism is mispricing under limited attention, amplified by gambling-prone retail investors who "buy winners and sell losers"; it is stronger in SOEs and after innovation-policy releases (the National Patent Development Strategy, 2010).

That last point is the academic echo of the practitioner claim: policy headlines and hot-money concentration accelerate and shorten the momentum window. The 3–10 day window traders actually trade is shorter than the weekly academic horizon, but the direction of the finding (short-horizon, retail-driven, policy-sensitive, concept-level momentum) is consistent.

1.3 Dragon-Tiger List: short-term predictive power, no long-term value

Two peer-reviewed studies support the Dragon-Tiger signal:

  • Hong, Yao & Zhuang (2025), Emerging Markets Finance and Trade 61(14): 4545–4559 — daily institutional trading derived from Dragon-Tiger data is positively related to short-term stock performance, especially for stocks with low analyst coverage; it is also positively correlated with retail sentiment. There is no significant relation with long-term returns or firm fundamentals (abstract).
  • Zhai, Hou & Li (2020), Journal of Business Economics and Management 21(3): 914–941 — investor attention measured via Dragon-Tiger listing negatively predicts returns when the listing-day cumulative return is negative (paper). Attention itself is a reversal machine: the list creates the crowd that later unwinds.

1.4 Limit-up premium & the "daredevil" economy

The hot-money game is structurally built around the limit-up (涨停) premium — queue to buy a sealed limit-up board, sell to the next-day crowd. Recent market data (Sina Finance, Aug 2025) shows how the edge is shrinking under the 20%-band boards: in 2024, first-board 炸板率 (failed seal rate) reached 62% on 20%-band stocks and next-day average premium fell to 3.2%; the top-30 ordinary (non-Eastmoney, non-HQ) Dragon-Tiger seats averaged only 42% limit-up success (Sina). Academic work on A-H twins finds that attention following limit-up events creates price premiums that later reverse (Journal of Banking & Finance 2026).

1.5 Order flow imbalance: minutes, not days

Zhang, Xie & Wang (2025), Asia-Pacific Journal of Accounting & Economics — using high-frequency Chinese market data, order imbalances positively predict returns over 5–30 minutes, and the relation reverses over 60–120 minutes (DOI 10.1080/16081625.2025.2604824). This is the microstructure layer under the day-level hot-money effect: OFI tells you whether the accumulation is real during the session; the day-level Dragon-Tiger data tells you who did it after the close.

2. Which model can detect it?

The table maps signal type → model family → evidence/examples → primary data.

Sub-problem Model family Evidence / examples Primary data
Market regime / rotation phase Gaussian HMM, Markov regime-switching, XGBoost regime classifier Gaussian HMM on price-volume factors decodes market-style states and drives a sector-rotation strategy (Huang & Lu, 2024, Finance 3: 1053–1071); end-to-end industry rotation uses four interpretable states defined by volatility × rotation speed with XGBoost regime probabilities (Shen, Cheng & Liu, CDEMS 2026, Nanjing Univ., Atlantis Press) Index & sector returns, volume, turnover, breadth
Sector/concept ranking for 1–3 weeks Cross-sectional regression (Cohen–Frazzini economic-link style), GBDT (LightGBM/XGBoost), FFN deep factors Industry/tech-link momentum (Duan et al. 2022); ML short-term reversal with RF/AdaBoost/GBDT/XGBoost shows high Sharpe and positive skew in US & China (SSRN); deep-factor extraction + directional classification (Shen et al. 2026) Sector/concept returns, related-firm returns, turnover, valuation
Temporal sequence learning LSTM/GRU, attention (Transformer), multi-scale heterogeneous attention Temporal HAN / Multi-scale Temporal Fusion Network on CSI Level-1 sectors with mixed-frequency alternative data outperforms SOTA on accuracy and portfolio returns (Wang et al., Data Science in Finance, Springer 2026); Siamese deep LOB models beat baselines on A-share stocks (arXiv:2505.22678) Daily sector panel + weekly/monthly fundamentals + alt data; LOB snapshots
Hot-money cluster detection GNNs on seat co-appearance graphs; community detection Seat-network factors: 游资联动图 (hot-money linkage graph), 机构主导 (institution dominance), 接力 vs 一日游 (relay vs one-day) (practitioner system, quant-ashare); GNN community detection on dynamic stock-correlation networks captures sector-rotation rhythms (VGAER, Physica A 2026) Dragon-Tiger top-5 buy/sell seats, stock-symbol co-occurrence
Sector/concept graph learning GNNs on heterogeneous graphs Multi-layer heterogeneous graph: 申万 L1/L2/L3 industries + concept boards + institutional holdings + DTW shape similarity + Pearson correlation, with GAT and MF-IAMGCN architectures (practitioner skill, quantskills) Industry taxonomy, concept membership, holdings, returns
Intraday flow detection OFI factor, VPIN, trade classification, anomaly detection OFI positively predicts 5–30 min, reverses 60–120 min (Zhang et al. 2025); Level-2 逐笔委托 analysis — 撤单比例, 委托方向持续性 — to infer hidden institutional intent (开源金工 "聪明钱3.0", Aug 2026); iceberg/spoofing (冰山单/幌骗单) feature detection (FinClaw a-share-hft-microstructure skill) Level-2 snapshots & 逐笔委托/逐笔成交 (tick-by-tick)
News/policy trigger Sentiment scoring, RAG, LLM multi-agent CICC maps high-frequency news to industry/style using DeepSeek-R1 for growth/value style switching (Mar 2025); 华泰 builds LangGraph multi-agent concept-selection and macro-hotspot systems over a knowledge base (Apr 2026); "逐鹿" ALPHA uses a local RAG over hundreds of millions of news items for timing + industry rotation (Apr 2024); central-bank text predicts returns best for ChiNext (CSCIED index) News wires (财联社, 交易所公告), policy text, earnings releases
Sentiment cycle (short-term market state) Rule-based scoring / state machine, optionally ML-calibrated 7-dimension score (涨停数, 最高连板, 炸板率, 赚钱效应, 成交额 vs 20d, 跌停对比, 市场宽度) → 5 phases 冰点/低温/回暖/高潮/过热 with explicit turning-point rules (practitioner framework, ClawHub stock-sentimental-cycle) Daily limit-up/limit-down pool, turnover, breadth

Practical model guidance

  • Start with the cheap models. A Gaussian HMM for regime + a weekly LightGBM ranking of sectors/concepts + a seat-net-buy scorecard will capture most of the effect. Deep and graph models add edge mainly when you need to model relationships (which concept leads which, which seats move together).
  • The horizon decides the label. Train classification/regression targets at 5–10 trading days for concept momentum and 1–3 weeks for technology-link momentum. Longer labels are the reversal trade, not this effect.
  • Regime-gate the strategy. The same momentum factor flips sign between hot-money "回暖/高潮" regimes and "冰点/退潮" regimes. A rotation-speed × volatility regime model (XGBoost or HMM) is the risk control, not an optional extra.
  • Do not train on price alone. The academic mechanism is attention + policy + retail flow, so the features that matter are turnover, limit-up pool state, seat accumulation, news sentiment, and policy-event flags — not just sector returns.
  • LLMs are for mapping, not forecasting. Current practitioner usage maps news → industries/styles/concepts (CICC, 华泰, 逐鹿). That mapping is genuinely useful; treating the LLM as a return forecaster is not supported by evidence.

3. How to monitor

3.1 Know the raw signals and their quirks

Dragon-Tiger List (龙虎榜). Exchange-published after the close each trading day. A stock is listed if it triggers any of: daily close deviation ±7% (main board), daily amplitude 15%, daily turnover 20%, or 3-day cumulative deviation ±20% (ST variants use lower thresholds; board-specific rules apply; SSE picks top-3 per condition, SZSE top-5) (证券之星). Consequences for monitoring:

  • It is a conditional sample: only extreme-move stocks appear, so any statistics built on it are selection-biased by construction.
  • Only top-5 buy and top-5 sell seats per listed stock are shown; the signal is partial.
  • Data is after close, so a signal is actionable at the next session's open at the earliest — you are trading the second day of the move.
  • The seat name is a brokerage branch, not an individual; identity must be inferred and changes over time (马甲).

Moneyflow (主力净流入). Eastmoney-style "main-force net inflow" is a soft estimate: it mechanically buckets orders by size (e.g., large orders > threshold) and labels them "main force." Useful for direction, unreliable for precision (ashare-mcp notes).

Northbound (北向资金). Since May 13, 2024, real-time northbound turnover is no longer disclosed; only after-close daily totals/penny counts and quarterly holdings remain (HKET, 10jqka). Any monitor relying on intraday northbound flow is broken; the surviving proxies are the LHB 深股通专用 seat, daily close data, and quarterly holdings changes.

Level-2. True order-flow imbalance requires 逐笔委托/逐笔成交 (order-by-order / execution-by-execution) data plus snapshot depth, not the 3-second five-level snapshots. This is the only layer that can detect spoofing/iceberg behavior and compute OFI/VPIN.

3.2 The monitoring stack

Layer Cadence What to compute Sources
1. Dragon-Tiger + seat identity EOD (evening, T+0 after close) Seat identity tags (机构专用 / 深股通专用 / 知名游资 / 量化); cumulative net buy; post-listing 5/10/20-day win rates; avg holding period; sector preference; seat co-appearance graph; 接力 vs 一日游 classification Tushare top_list/top_inst (2005–), AkShare get_lhb_daily/get_lhb_stock_detail/get_lhb_institution_daily/get_lhb_active_branches, 散股通-style LHB platforms
2. Limit-up pool & sentiment cycle EOD + intraday 首板/连板/炸板数/炸板率, highest consecutive board (最高连板), 赚钱效应 (yesterday's limit-ups today's average move), 跌停/涨停 ratio, breadth, turnover vs 20d → 7-dim score → 冰点/低温/回暖/高潮/过热 + turning signals 涨停池 APIs (AkShare/东财), ClawHub-style scoring rules
3. Multi-source capital consensus EOD Northbound (close totals, quarterly holdings), margin financing (融资余额 changes), block trades (大宗交易), institutional seats; flag agreement vs divergence (北向×机构×融资×大宗) Tushare margin/northbound/block APIs, quantskills smart-money-profiler & capital-flow-crowding-monitor patterns
4. Intraday microstructure Real-time / tick OFI per 1–5 min bar, VPIN, 封单 size & 撤单率, iceberg/spoof anomalies, 大单挂撤单 patterns Level-2 逐笔委托/逐笔成交 + snapshot depth (vendors: 通达信/同花顺/东财 L2)
5. News & policy feed Continuous Policy-event flags per sector/concept, news→industry/style sentiment mapping, concept-formation detection 财联社 telegraph, 交易所公告, RSS/web search; CICC/华泰-style mapping pipelines

3.3 Seat watchlist: a concrete example

The 2025 H1 brokerage-branch Dragon-Tiger ranking (财联社, Jul 2025) shows the seat archetypes a monitor must label (CLS):

  • Retail/quant aggregation: 东方财富证券 "拉萨系" branches (金融城南环路 #1 by turnover, 团结路第一/第二, 东环路第一/第二) — high frequency, huge turnover, widely believed to be retail + programmatic/HFT.
  • Quant/institutional channels: 华宝证券上海东大名路 (#2, 804.39亿 turnover, ETF-heavy, medium-short trend following), 华鑫证券上海分公司 (#8, 344.18亿).
  • Regional hot-money buyers: 开源证券西安太华路 (net buy 63.23亿 in H1), 国泰海通宜昌沿江大道 (net buy 70.73亿, #2 in net buys), 银河大连黄河路.
  • Distribution seats: 中信证券北京建外大街 (net sell −41.83亿), 国泰君安总部 (−27.02亿), 国信深圳红岭中路.
  • Foreign: 瑞银证券上海花园石桥路第二营业部 (net seller in H1, turned net buyer in June), often appearing with 沪股通.

A practical scorecard per tracked seat: hm_net_5d/20d/60d, active-seat count over 30d, top-3 concentration (feature set used in a Tushare-based RL research project, aurumq-rl), plus win rate and holding-period stats from the smart-money-profiler pattern.

3.4 Alert rules worth having

  • Cycle turning points: 冰点→回暖 requires ≥3 of (跌停<10, 最高连板≥4, 溢价>0, 涨停≥30, top hot-money seat appears on LHB); 高潮/过热→退潮 fires on any 2 of (龙头 deep drop, 炸板率≥40%, 赚钱效应<0, 涨停<50, 连板断层 ≤3) (ClawHub rules).
  • Seat divergence: hot-money net buy ↑ while institutional seats net sell on the same stock, or margin financing rising while northbound retreats — classic short-term top configuration.
  • Regime flip: HMM/XGBoost regime probability shifts from 回暖 to 退潮; stop opening new concept positions.
  • OFI reversal: intraday OFI has already flipped negative 60–120 min after a positive surge — the daily signal is stale by then.
  • Board-structure change: 炸板率 spiking and 20cm first-board premium collapsing (2024: 62% 炸板率, 3.2% next-day premium) means the limit-up-chasing edge is regime-dependent, not constant.

4. Caveats

  • Selection bias: every LHB statistic conditions on extreme-move stocks. Win rates "after listing" are not tradable without considering you can rarely buy at the listing-day close (limit-up queues, T+1 settlement, stamp duty, slippage).
  • Identity inference: a branch name is not a person; hot-money operators change seats, and quant/HFT share the same branches. Seat labels decay and need re-estimation.
  • Data latency asymmetry: LHB is next-day-actionable; the best intraday signal (true OFI) needs Level-2 data that most retail setups do not have; moneyflow is a mechanical estimate.
  • The effect is short by construction: academic and practitioner evidence agree the edge is days-to-weeks; holding beyond that converts a momentum trade into a mean-reversion loss, especially when attention peaks (limit-up frenzy, 过热 phase).
  • Regime dependence: 2024–2025 20%-band boards structurally weakened the limit-up premium. Model calibration must be re-estimated across board-structure regimes.
  • This is research, not investment advice: the models and monitors above describe the effect and how practitioners measure it; none of it guarantees tradable returns net of costs.

5. Key sources

Academic

  1. Duan Binglei, Wang Rongfei, Zhang Ran (2022). Economic Links and Stock Returns in Chinese A-Share Market (南橘北枳：A股市场的经济关联与股票回报), Journal of Financial Research, 500(2): 171–188. — journal abstract / PKU copy
  2. Hong, Xin; Yao, Juan; Zhuang, Zhuang (2025). Institutional Trading and Short-Term Stock Returns – Evidence from Dragon and Tiger List Data in China, Emerging Markets Finance and Trade, 61(14): 4545–4559. — EconPapers
  3. Zhai, X.-Y.; Hou, Y.-Y.; Li, Y.-S. (2020). Investor attention and stock returns under negative shocks: an empirical analysis based on "Dragon and Tiger" list in China, Journal of Business Economics and Management, 21(3): 914–941. — open access
  4. Zhang, Ting; Xie, Chi; Wang, Gang-Jin (2025). Do order imbalances predict intraday returns? New evidence from the Chinese stock market, Asia-Pacific Journal of Accounting & Economics. DOI 10.1080/16081625.2025.2604824 — T&F
  5. Shi, Huai-Long; Zhou, Wei-Xing. Time series momentum and contrarian effects in the Chinese stock market (arXiv:1702.07374). — full text
  6. Shen, Qiuyu; Cheng, Hongsen; Liu, Haifei (2026). An End-to-End Regime-Dependent Industry Rotation Strategy in China's A-Share Market, CDEMS 2026, Atlantis Press, 473–480. — paper
  7. Wang et al. (2026). Multi-scale temporal fusion network: A heterogeneous temporal attention network with cross-frequency alternative data for sector rotation in China, Data Science in Finance (Springer). DOI 10.1007/s44443-025-00364-0 — article
  8. Huang, Jiacheng; Lu, Xianggang (2024). 基于高斯分布隐马氏模型识别市场风格构建量化策略 (Gaussian HMM for market style identification and sector rotation), Finance (Hans), 3: 1053–1071. — Hans
  9. Dynamic market structure & stock correlation networks with GNN community detection (VGAER), Physica A, 2026. — ScienceDirect
  10. An Efficient Deep Learning Model to Predict Stock Price Movement Based on Limit Order Book (A-share application), arXiv:2505.22678. — arXiv
  11. A Machine Learning Approach for the Short-term Reversal Strategy (RF/AdaBoost/GBDT/XGBoost, US & China), SSRN. — via R Discovery summary

Practitioner / data sources

  1. 财联社 (2025-07-02). 最新券商营业部龙虎榜出炉 — 2025 H1 seat rankings, 拉萨系/华宝/华鑫/开源/瑞银 examples. — CLS
  2. Sina Finance (2025-08-19). 剖析新牛市的涨停板"敢死队"：游资打板收益与风险双向扩张 — 20cm 炸板率 62%, 次日溢价 3.2%, top-30 seats 42% 打板成功率. — Sina
  3. HKEX / media (2024-05-13). Northbound real-time disclosure change. — HKET / 10jqka
  4. Dragon-Tiger listing criteria explainer. — 证券之星
  5. A-share data access: Tushare top_list/top_inst (via LobeHub Tushare MCP); AkShare LHB functions & Eastmoney moneyflow caveats (via ashare-mcp).
  6. Seat identity & profiling pattern: smart-money-profiler (identity tags, 5/10/20-day win rates, holding period, sector preference, 北向×机构×融资×大宗 consensus); hot-money cumulative features (hm_net_5d/20d/60d etc.) from aurumq-rl; seat-network 13 factors (游资联动图/机构主导/接力 vs 一日游) from quant-ashare.
  7. Sentiment-cycle scoring framework (7 dims → 5 phases, turning rules): A-Stock Sentimental Cycle; limit-up pool management (首板/连板/炸板/回封/题材分组/晋级率): skill-b6-limitup-pool.
  8. News/policy mapping systems: CICC 如何结合高频新闻和传统风格轮动框架 (DeepSeek-R1, Mar 2025) — summary; 华泰 大模型概念与宏观热点选股 (LangGraph multi-agent, Apr 2026) — summary; "逐鹿" ALPHA RAG timing & industry rotation (Apr 2024) — report.
  9. Level-2 microstructure patterns: FinClaw a-share-hft-microstructure; 开源金工 "聪明钱3.0" (逐笔委托撤单比例/委托方向持续性, Aug 2026) — Sina.