This is a legitimate quant finance request about A-share market analysis, specifically identifying conditions that trigger a "主升浪" main rally phase. I want to give a thorough, practical answer covering how to structure the analysis and model design, while being honest about the limitations of predicting market moves. I'm mapping out an operational definition for "主升浪"—since it's a subjective, post-hoc concept from Elliott Wave and 缠论, I need to turn it into something quantifiable like a labeled price rise threshold with drawdown constraints. Then I'm thinking through data sources for A-shares (Tushare, AKShare, JoinQuant, Wind, etc.) and the wide range of features needed: adjusted OHLCV, limit-up/down flags, ST status, turnover, northbound flow, 龙虎榜, sector classifications, and intraday data for finer signals. I want to define the event using something like a triple-barrier method—forward returns over 20-60 days exceeding a threshold with limited drawdown—then treat this as an event study/classification problem. I also need to account for A-share-specific mechanics: 涨跌停 limits, T+1 settlement, limited shorting, and retail-driven dynamics like 连板 and sector rotation, which shape how these rallies typically start and unfold. From there I'm starting to sketch feature categories, beginning with price pattern signals like breakouts above prior highs. Now I'm expanding into more feature categories: volume-related signals like turnover changes and chip distribution, capital flow indicators like main-force net inflow and margin balance shifts, and fundamental catalysts like earnings surprises and revenue growth acceleration. Then I'm moving to theme/sector context (limit-up counts within a sector, sector leadership, concept heat), broader market environment metrics (index position, market turnover, limit-up/down counts, sentiment cycles), and valuation fundamentals like PE/PB percentile and market cap effects. Finally I'm considering modeling approaches, starting with rule-based screening for its interpretability and resistance to overfitting. Then I'm thinking through machine learning options — gradient boosting with proper time-series cross-validation, evaluating precision at K rather than accuracy given the extreme class imbalance, and noting deep learning on raw price data tends to overfit. Event study analysis comparing hit rates against base rates seems like the key method for identifying trigger causes, and I'm moving into validation pitfalls like survivorship bias and look-ahead bias. Now I'm cataloguing more traps: concept sector labels are assigned retroactively after a stock already rallied, so that data leaks the future; limit-up days can't actually be bought at signal price, so fills need realistic simulation; transaction costs including stamp tax, commission, and transfer fees must be modeled along with slippage for small caps; and T+1 settlement constrains same-day trades. I'm also estimating base rates — only a small percentage of stocks see major rallies in a given year, varying a lot by market conditions. For analyzing the trigger causes behind each identified uptrend, I want to attribute each breakout day to a category — policy/news-driven themes, earnings inflections, hot-money momentum plays, pure technical breakouts with no news, restructuring/M&A, or sector rotation — and then look at how these trigger types are distributed over time and across market regimes. I also want to separate the "soil" (state features like consolidation duration, volume shrinkage, chip concentration, market cap) from the "spark" (event features like limit-up days, volume surges, announcements, sector kickoffs), modeling the probability of a major uptrend conditional on both, using feature importance/SHAP and conditional lift analysis to see how much each factor shifts the baseline probability. Then I'm sketching the implementation architecture: a data layer for downloading, cleaning, and storing data efficiently, a feature layer with vectorized, point-in-time-correct computations, a labeling layer, and a research layer for exploration and statistical testing. For backtesting I'm considering either vectorized or event-driven frameworks like backtrader, vnpy, or Microsoft's qlib since it has good A-share support with built-in data and models, and for live use the plan is a daily post-close scan generating a candidate pool executed at next-day open or during the call auction. Since the user wants something that actually runs, I should provide a minimal rule-based strategy plus a working code skeleton using akshare or tushare with pandas and lightgbm, kept to a reasonable length, while setting realistic expectations that predicting major uptrends ahead of time is inherently low-precision—maybe 15-25% precision at top-K, relying on payoff asymmetry to make it worthwhile. The model should really be framed as a candidate-pool screener rather than a crystal ball, since position sizing and stop-loss discipline matter more than the raw signal, and since A-share momentum runs are heavily concentrated in thematic phases, a market regime filter becomes essential. I'm also thinking the user may be relatively new, so I should structure the answer with clear headers, concrete numbers, and honest caveats, writing in Chinese and including a code skeleton without letting it balloon in length. For the label definition itself, I'm weighing a simple option—future 60-day max return ≥60% with drawdown from signal to peak capped at 20%—against a more rigorous triple-barrier method. I'm also considering a third option: identifying momentum runs after the fact via zigzag detection, then tracing back to examine conditions before the run's starting point, which fits the user's interest in understanding triggering conditions rather than pure prediction. My conclusion is that the zigzag/post-hoc approach suits attribution research while the triple-barrier method suits predictive modeling, and both deserve mention, along with distinguishing different flavors of A-share momentum runs like speculative "妖股" limit-up chains versus healthier 60-day rallies backed by bullish moving averages. A key data pitfall is that concept-sector membership (同花顺/东财) isn't point-in-time, so I'd need to snapshot it daily or fall back on 申万行业 classification since tushare's index_member data includes historical in/out dates -- I'll need to build my own snapshot database for concept tags. Another pitfall is that financial data should be aligned to announcement date (ann_date) rather than reporting period (end_date) to avoid look-ahead bias, and tushare conveniently provides ann_date for this. For return calculations I should use 后复权 (backward-adjusted) prices, but for limit-up detection I need unadjusted prices with pre_close, and tushare's stk_limit field gives the actual limit price. I also need to handle delisted and newly-listed stocks properly -- excluding IPOs within their first 60-120 trading days but still including historically delisted stocks in the universe -- and remember ST stocks trade under a 5% limit band rather than 10%. Finally, I should flag that northbound flow data (北向资金) had its real-time daily disclosure discontinued by HKEX starting August 19, 2024, which breaks any strategy relying on that feed. For evaluating the model, I'll want Precision@K, IC (Spearman correlation between predicted score and forward return), decile-based lift analysis, portfolio turnover, and a realistic backtest incorporating trading costs, plus a check of performance stability across different market regimes/years. Given the small effective sample size from overlapping labels across thousands of stocks, I should benchmark against a random baseline and use block bootstrap over time to gauge statistical significance properly. Since the user wants something "真实的能跑的" (real and actually runnable), I need to focus heavily on practical execution: the data pipeline, a daily scheduler, and real trading constraints like inability to buy limit-up stocks and slippage — possibly touching on live execution via QMT/PTrade or manual candidate lists. I'm now outlining the overall structure: first turning "主升浪" from a vague concept into a computable label definition, then covering data preparation with A股-specific pitfalls, then research methodology starting with post-hoc attribution before moving to predictive modeling. For the methodology, I'm laying out cutting historical main-uptrend samples via zigzag, then a three-layer factor framework (soil/pre-conditions, spark/trigger events, weather/market environment), followed by conditional probability and lift analysis as the core answer to the "underlying conditions and triggers" question. From there I'll cover a feature engineering checklist with A股 특성, modeling that moves from rules to tree models before considering deep learning, validation with time-based splits and cost/limit-up constraints, system architecture with code skeletons, and finally live trading and expectation management — aiming for a comprehensive but well-structured answer rather than an exhaustive book, roughly 2500-3500 words. For the code skeleton, I'm sketching a Python workflow using akshare or tushare for data, pandas for processing, a labeling function using triple-barrier or zigzag logic, a handful of feature examples, lift analysis, and a LightGBM model with walk-forward validation — keeping the code compact and functional while making sure the labeling logic (forward max return with drawdown constraint) is correct. For each stock's price series, my labeling function scans forward within a fixed horizon to check whether price first hits an upside threshold before hitting a downside drawdown threshold, assigning a binary label based on which barrier is touched first. It's a simplified triple-barrier method, favoring clarity over vectorized speed, though I note numba or polars could speed it up later. I'm also considering a zigzag approach for post-hoc validation of the labels. Writing zigzag pivot logic... Refining the trend-tracking logic... Still writing the zigzag function... Now I'm thinking about defining a "main upswing" segment as consecutive low-to-high pivots where the rise is at least 50% within 120 days, keeping in mind the low pivot is only confirmed retrospectively so this works for attribution analysis but needs triple-barrier labeling for prediction. Then I want to examine trigger conditions around the low point—limit-up days, volume spikes, announcements, or sector-wide limit-ups—to build a lift analysis. Setting up a walk-forward validation with LightGBM, splitting by year with an embargo period between train and test sets... I'm reconsidering the embargo window size—since labels look 60 trading days forward, I want enough calendar-day buffer (~90 days) between training end and test start to avoid leakage. Now I'm writing a precision@K evaluator that ranks stocks by score each date and compares the top-K's realized label rate against the baseline. I also need to make sure all features are computed with rolling windows using only past data (min_periods set appropriately) to avoid lookahead bias. For the feature set, I'm sketching a compact function covering momentum returns (20/60/120-day), moving average ratios and slope, volume ratio, turnover stats, volatility measures like ATR and Bollinger band width percentile, distance from 250-day high, price range tightness over 120 days, count of limit-up days in the past 60, relative strength vs index, market cap/PE percentile, and sector-relative ranking. I should also add market-regime features like the count of limit-up/limit-down stocks across the whole market, total turnover, and index MA status. Then thinking through common A-share trigger patterns: sector/theme catalysts driven by policy or news, earnings surprises around quarterly report windows, restructuring or M&A events, technical breakouts from long consolidation with volume, and known institutional or "hot money" activity on the dragon-tiger list. For preconditions, I should list things like prolonged low-volatility consolidation, moving average convergence, prior history of limit-up behavior, small-to-mid float market cap, declining shareholder counts signaling concentration, and either deep pullback from highs or fresh breakout to new highs, plus sector-wide activity picking up — but I need to frame all of these as testable hypotheses rather than confirmed facts, emphasizing the assumption-to-verification pipeline. I should also flag survivorship bias in my own recollection — I remember the consolidating stocks that later surged but forget the many that never did, so base rates matter, which is exactly why lift analysis over raw memory is essential. I want to be honest that a decent model might only lift the probability of a 60-day 50% gain from a ~3-5% base rate to 10-20% in the top decile, still useful but far from certain, and I should note historical spike periods like 2015, 2020, and September 2024. For execution, I'm thinking signal at close then buy next-day open, skipping if gap-up or limit-up, scaling in gradually, setting stop-losses around -8% to -10% or a break below MA20, trailing take-profits, and equal-weighting across 10-20 positions. I want to cover the practical tooling landscape too: tushare pro for data (needs points/donation), akshare as a free but less stable alternative, baostock for daily bars, qlib for Microsoft's ML pipeline, JoinQuant/RiceQuant for online backtesting with limited export, and duckdb/parquet for local storage — plus live trading options like QMT, PTrade, or semi-manual order execution. I'm ready to write the full answer now in Chinese with clear structured sections. I'll design the code skeleton to be runnable with a real tushare token, using the daily, adj_factor, and daily_basic endpoints to pull price, adjustment factor, and fundamental data for one or more stocks, keeping it around 100-150 lines with notes on where users need to fill in their own parameters. There are more endpoints worth noting too — index classification, concept indices, HSGT flows, announcements, and the trading calendar, though some need higher point tiers. Given rate limits, looping by trade_date with pro.daily is most efficient — about one call per day, roughly 250 calls a year — and since each day's row count stays near the 6000-row cap but under it, this approach should work reliably. For adjusted prices, I'll compute 后复权 as close times adj_factor, and 前复权 as that divided by the latest adj_factor. For limit-up detection, I'll approximate using pct_chg or the rounded 10% pre_close comparison, flagging it as an approximation since precise limit thresholds require stk_limit data. Given the scale — roughly 12.5 million rows across a decade of full-market data — I'm leaning toward polars over pandas for performance, and I still want to sketch out the attribution table structure. I want to build a table for each 主升浪 sample recording start date, stock, gain percentage, duration, and trigger tags, summarized by year, plus an illustrative feature-to-conditional-probability example clearly marked as 示意 rather than real empirical numbers. I also plan to include a "常见坑" pitfalls list, and a note cautioning against over-relying on CNN-based candlestick pattern recognition since it tends to overfit compared to tabular OHLCV features — though candle patterns can still be encoded as numeric features if the user wants a visual approach. The core methodology should be event-based statistics, treating "主升浪" as an after-the-fact concept, so I need to first define detection rules, and different types of 主升浪 likely need separate models based on clustering by 涨速/持续时间/涨停占比. Now I'm drafting the full Markdown answer in Chinese with headers and bullet lists, starting from the conclusion that quantifying this requires clear event definition first. I'm structuring the pipeline as: data → labeling (defining 主升浪) → extracting historical samples → attribution analysis (soil/spark/weather) → feature engineering → rules/models → time-series validation → simulated trading → live trading, then defining the identification criteria using zigzag segmentation with thresholds for gain, duration, and max drawdown. For the predictive labeling side, I'm proposing a triple-barrier method looking 60 trading days ahead, and splitting stocks into two distinct archetypes—speculative "hot money" plays versus trend-following moves—since mixing them dilutes both soil and trigger analysis. I'm outlining the data section and then the core three-layer attribution framework: soil (pre-existing conditions), spark (trigger events), and weather (market environment), using lift analysis and conditional probability to dissect trigger causes. Now I need to structure the remaining sections—feature list, modeling, validation pitfalls, code skeleton, and live trading expectations—and write everything out comprehensively while keeping each point concise. 这是一个很好的问题，但要先把一个关键认知摆在前面：**“主升浪”是一个事后概念**。任何人回头看K线都能画出主升浪，难的是在起点附近识别它。所以量化设计的核心不是“找主升浪”，而是三件事： 1. 把“主升浪”变成可计算的**标签** 2. 把K线经验拆成可统计的**特征** 3. 用**事后归因 → 事前预测 → 严格验证**的顺序做，而不是一上来就套模型下面按这个顺序给你一套完整的设计方案。 --- ## 一、先把“主升浪”定义成标签 ### 1. 事后识别（用于研究“条件和触发原因”）用 zigzag 算法把每只股票的历史走势切成上下段，筛选出： - 涨幅 ≥ 50%（可分档：50%/100%/200%） - 持续 15~120 个交易日 - 途中最大回撤 ≤ 20% 这些段的**起点**就是历史主升浪样本。A股约 5000 只股票 × 10 年，能切出几千到上万个样本，够做统计。 ### 2. 事前预测标签（用于建模）对每个 (股票, 日期) 用三重障碍法打标： - 未来 60 个交易日内，**先**触及 +50% → 标签 1 - 先触及 −15%，或到期未触及 → 标签 0 这样标签是“站在当天能否知道未来会启动”，没有后视偏差。 ### 3. 建议分型 A股的主升浪至少两类，土壤和触发原因完全不同，混在一起会互相稀释： | 类型 | 特点 | 典型触发 | |---|---|---| | 游资题材型 | 10~20 日翻倍、涨停占比高、换手极高 | 板块事件、龙虎榜席位、连板 | | 机构趋势型 | 60~120 日 +50%~100%、涨停少、沿均线走 | 业绩拐点、行业景气、北向/机构增持 | 先按“涨停天数占比”“涨速”做个简单聚类分型，分别研究。 --- ## 二、数据准备 ### 数据源 - **Tushare Pro**：日线、复权因子、daily_basic（换手率/市值/PE）、涨跌停价、龙虎榜、资金流向、融资融券、股东户数、业绩预告、申万行业（含历史成分变更）。性价比最高，需要一定积分。 - **AKShare / Baostock**：免费，接口稍不稳定，适合起步。 - **Qlib（微软）**：自带 A 股数据处理和 ML 流水线，适合后期。 - 实盘接口：QMT（xtquant）、PTrade，需券商开通。 ### 必备字段 - 日线 OHLCV + 前收盘 + 复权因子（**收益率用后复权算，涨停判断用未复权价**） - 换手率（自由流通）、流通市值、量比 - 涨跌停价（主板 10%、创业板/科创板 20%、ST 5%、北交所 30%） - 行业/概念分类、指数日线 - 事件类：业绩预告、公告、龙虎榜、股东户数、大宗交易、限售解禁 - 市场情绪：每日涨停/跌停家数、炸板率、两市成交额 ### A股特有的坑（非常重要） 1. **概念板块是事后贴的标签**。同花顺/东财的“AI概念”成分是股票涨起来之后才加进去的，直接用历史成分会严重泄露未来。必须从现在开始**每日快照**存下来，或者只用申万行业（Tushare 有 in_date/out_date）。 2. **财务数据用公告日，不用报告期**。一季报是 4 月底才公布的，不能在 3 月 31 日用。 3. **保留退市股**。只用现存股票 = 生存者偏差，会高估底部票的胜率。 4. **涨停买不到**。信号日收盘涨停，回测里不能按收盘价成交。 5. **T+1、印花税 0.05%（卖出）、佣金、小票冲击成本**都要进回测。 6. 北向资金逐日净买入 2024 年 8 月起已停止披露，历史可研究，不能作为未来实盘特征。 --- ## 三、核心研究方法：三层归因这是回答你“潜在条件、触发原因”的关键部分。把影响因素分三层： **土壤（前置状态）**：启动前 1~3 个月股票处于什么状态 **火种（触发事件）**：启动日前后 3 天发生了什么 **天气（市场环境）**：当时大盘和情绪周期在什么位置 ### 分析方法：条件概率 / Lift 对每个特征 X，算： ``` Lift(X) = P(主升浪 | X 满足) / P(主升浪) ``` 举例（**以下数字是示意，需要你自己算**）： | 条件 | 样本占比 | 60日内+50%概率 | Lift | |---|---|---|---| | 全市场基准 | 100% | 4% | 1.0 | | 流通市值 20~100 亿 | 40% | 5.5% | 1.4 | | 过去 120 日振幅 < 30%（横盘） | 25% | 6% | 1.5 | | 20 日换手率处于 2 年内最低 20% 分位 | 20% | 5% | 1.25 | | 当日放量涨停且前 60 日无涨停 | 1% | 12% | 3.0 | | 所在行业当日涨停数 ≥ 5 | 3% | 10% | 2.5 | | 横盘 + 放量突破 + 行业活跃 | 0.3% | 20% | 5.0 | 这张表就是你要的“潜在条件”。**你脑子里的K线经验（横盘、缩量、突破、均线粘合）全部拆成这样的行去检验**，Lift 不显著的直接扔掉。这一步会帮你破除很多“幸存者记忆”——你记住了横盘后启动的票，忘了大量横盘后继续横盘的票。 ### 触发原因归因流程对每个事后识别的主升浪起点，在 [T-3, T+3] 窗口自动打标签： - 是否有公告（业绩预告、重组、定增、回购、中标） - 是否涨停、量比多少 - 所在行业/概念当日涨停家数、是否板块内涨幅第一（龙头） - 是否上龙虎榜、是否有知名席位 - 是否创 120/250 日新高（突破型）还是从深跌反弹（超跌型） - 大盘当日和前 5 日表现、全市场涨停家数变化然后统计：每年触发原因的分布、不同触发对应的后续涨幅和持续时间。你会得到类似“2023 年 60% 的主升浪由题材事件触发，业绩驱动只占 15%”这样的结论——这才是数据说话的“触发原因”。 **先做归因，再做预测。** 归因告诉你哪些特征值得做进模型，能少走很多弯路。 --- ## 四、特征工程清单按上面三层组织，全部只用当天及之前的数据： **土壤（状态特征）** - 位置：距 250 日高点跌幅、距 250 日低点涨幅、收盘/MA20、MA20/MA60/MA120 排列与斜率、均线粘合度（各均线最大相对偏差） - 压缩：20 日 ATR/价格、布林带宽度在 1 年内的分位、120 日振幅（横盘紧密度）、横盘持续天数 - 量能：20/60 日均换手率及其 1 年分位、换手率趾势（缩量程度）、量价相关性 - 筹码：股东户数环比变化、成交量加权平均成本与现价距离（近似获利盘比例） - 基因：过去 60/250 日涨停次数、历史最大 20 日涨幅（曾被资金关注） - 基本面：流通市值（对数）、PE/PB 分位、营收/利润增速拐点、机构调研次数 - 相对强度：个股 20/60 日收益 − 行业/指数收益 **火种（事件特征）** - 当日涨停、涨停类型（一字/T字/放量）、连板数、炸板 - 量比、成交量/60日均量、跳空缺口幅度 - 是否突破 60/120/250 日高点、突破平台 - 龙虎榜净买入、席位类型 - 公告类型 one-hot（业绩预告超预期、重组、定增等） - 行业内涨停数、行业当日涨幅排名、个股在行业内涨幅排名 **天气（市场特征）** - 涨停家数/跌停家数、炸板率、连板高度 - 两市成交额及其 60 日分位 - 指数与 MA20/MA60 关系、指数 20 日收益 - 情绪周期标签（冰点 / 回暖 / 高潮 / 退潮，可用涨停家数 + 炸板率规则定义） K线形态本身（锤子线、十字星、缺口等）可以用 TA-Lib 的 CDL 系列函数批量生成，成本很低，但根据经验单独的形态信号 Lift 普遍不高，更多作为组合条件。 --- ## 五、建模：从规则到模型 **第一版：规则筛选。** 用 Lift 分析里最强的 4~6 个条件做 AND/评分组合，每日输出候选池。优点是可解释、不易过拟合、出问题能定位。很多能实盘跑的策略就停在这一步。 **第二版：树模型。** LightGBM/XGBoost 做二分类，输入上面的表格特征。注意： - 正样本极少（3%~5%），**不要看准确率**，看 Precision@K（每天打分最高的 20 只里有多少真启动）和分位收益 - 用 SHAP 看特征贡献，反过来验证归因结论 - 样本按时间切，不能随机切 **第三版（可选）：分型建模。** 游资型和趋势型分别建模，或加“类型”作为多分类标签。 **不建议一开始上深度学习 / K线图 CNN。** 图像里的信息 OHLCV 表格特征已经全部包含，CNN 只是换了一种更容易过拟合的表达方式；等你有了稳定的特征体系和验证框架再试。 --- ## 六、验证：这是决定“能不能真跑”的环节 1. **Walk-forward 时间切分**：用 2013–2018 训练、2019 测试；再用 2013–2019 训练、2020 测试……逐年滚动。训练集末尾要**留出 60 个交易日的空档**（embargo），因为标签本身看未来 60 天。 2. **按年看表现**。A 股是强 regime 市场：2015、2019 上半年、2020、2024 年 9 月主升浪遍地，2018、2022 极少。模型在牛市年精度 20%，熊市年可能只有 3%。**市场环境特征或环境过滤器是必须的**，否则实盘会在熊市里不停被止损。 3. **成交模拟**：信号日收盘 → 次日开盘买；次日一字涨停或高开 >5% 跳过；卖出同理。 4. **成本**：双边至少按 0.3% 算，小市值票再加冲击成本。 5. **随机基准**：同样的持仓数量、同样的市值分布，随机选股的表现是多少？模型必须显著跑赢它。 6. **多重检验**：你测了 100 个特征，总有几个“看起来有效”。用分年稳定性、Lift 的置信区间、以及最后 2~3 年完全不碰的留出集来防这一点。 --- ## 七、代码骨架以下是一个可以直接跑起来的最小框架（Tushare Pro + pandas + LightGBM），覆盖标签 → zigzag 切段 → 特征 → Lift → 滚动训练 → Precision@K。你可以在这上面逐步加特征。 ```python import numpy as np, pandas as pd, tushare as ts, lightgbm as lgb from pathlib import Path pro = ts.pro_api('你的token') DATA = Path('data'); DATA.mkdir(exist_ok=True) # ---------- 1. 按交易日拉全市场日线（每天一次请求，效率最高） ---------- def fetch_all(start='20140101', end='20241231'): cal = pro.trade_cal(exchange='SSE', start_date=start, end_date=end, is_open='1') for d in cal.cal_date: f = DATA / f'{d}.parquet' if f.exists(): continue px = pro.daily(trade_date=d) adj = pro.adj_factor(trade_date=d) bas = pro.daily_basic(trade_date=d, fields='ts_code,trade_date,turnover_rate_f,volume_ratio,pe_ttm,pb,circ_mv') df = px.merge(adj, on=['ts_code','trade_date']).merge(bas, on=['ts_code','trade_date']) df.to_parquet(f) def load(): df = pd.concat(pd.read_parquet(f) for f in sorted(DATA.glob('*.parquet'))) df['date'] = pd.to_datetime(df.trade_date) df = df.sort_values(['ts_code','date']).reset_index(drop=True) df['close_adj'] = df.close * df.adj_factor # 后复权 # 近似涨停判断（严格版用 pro.stk_limit） lim = np.where(df.ts_code.str.startswith(('300','688')), 0.20, 0.10) df['limit_up'] = (df.close >= (df.pre_close * (1 + lim)).round(2) - 0.005).astype(int) return df # ---------- 2. 标签：三重障碍 ---------- def triple_barrier(close, horizon=60, up=0.5, dn=0.15): n = len(close); y = np.zeros(n, dtype=int) for i in range(n - horizon): path = close[i+1:i+1+horizon] / close[i] - 1 hu = np.argmax(path >= up) if (path >= up).any() else horizon + 1 hd = np.argmax(path <= -dn) if (path <= -dn).any() else horizon + 1 y[i] = int(hu < hd) return y # ---------- 3. 事后识别主升浪：zigzag 切段 ---------- def zigzag(close, th=0.15): piv, trend, ei, ep = [], 1, 0, close[0] for i in range(1, len(close)): p = close[i] if trend == 1: if p >= ep: ei, ep = i, p elif p <= ep * (1 - th): piv.append((ei, ep, 'H')); trend, ei, ep = -1, i, p else: if p <= ep: ei, ep = i, p elif p >= ep * (1 + th): piv.append((ei, ep, 'L')); trend, ei, ep = 1, i, p return piv[1:] # 首个转折点可能失真，丢弃 def find_rallies(g, min_gain=0.5, max_days=120): c = g.close_adj.values; piv = zigzag(c); out = [] for (i0, p0, t0), (i1, p1, t1) in zip(piv, piv[1:]): if t0 == 'L' and t1 == 'H' and p1/p0 - 1 >= min_gain and i1 - i0 <= max_days: seg = c[i0:i1+1]; dd = (seg / np.maximum.accumulate(seg) - 1).min() out.append(dict(ts_code=g.ts_code.iloc[0], start=g.date.iloc[i0], end=g.date.iloc[i1], gain=p1/p0-1, days=i1-i0, max_dd=dd, limit_up_ratio=g.limit_up.values[i0:i1+1].mean())) return pd.DataFrame(out) # ---------- 4. 特征（全部只用当天及之前数据） ---------- def make_features(g): c, v, tr = g.close_adj, g.vol, g.turnover_rate_f f = pd.DataFrame(index=g.index) f['ret20'] = c.pct_change(20); f['ret60'] = c.pct_change(60) f['c_ma20'] = c / c.rolling(20).mean() - 1 f['ma20_ma60'] = c.rolling(20).mean() / c.rolling(60).mean() - 1 f['ma_glue'] = (pd.concat([c.rolling(w).mean() for w in (5,10,20,60)], axis=1) .pipe(lambda m: (m.max(1) - m.min(1)) / m.mean(1))) # 均线粘合度 f['off_high250'] = c / c.rolling(250).max() - 1 # 距年内高点 f['range120'] = c.rolling(120).max() / c.rolling(120).min() - 1 # 横盘紧密度 f['atr_pct'] = (g.high - g.low).rolling(20).mean() / c # 波动率 f['vol_ratio'] = v.rolling(5).mean() / v.rolling(60).mean() # 量能放大 f['tr20'] = tr.rolling(20).mean() f['tr20_pctile'] = f.tr20.rolling(250).rank(pct=True) # 缩量程度 f['n_lu60'] = g.limit_up.rolling(60).sum() # 涨停基因 f['lu_today'] = g.limit_up f['log_mv'] = np.log(g.circ_mv) f['pe_ttm'] = g.pe_ttm.clip(-200, 500) f['y'] = triple_barrier(c.values) return f # ---------- 5. Lift 分析 ---------- def lift_table(df, feat, bins=10): q = pd.qcut(df[feat], bins, duplicates='drop'); base = df.y.mean() t = df.groupby(q, observed=True).y.agg(['mean', 'count']) t['lift'] = t['mean'] / base return t # ---------- 6. 滚动训练 + Precision@K ---------- def walk_forward(df, feats, first_test_year=2018, embargo_td=60): out = [] for y in range(first_test_year, df.date.dt.year.max() + 1): test_start = pd.Timestamp(f'{y}-01-01') cutoff = df.date[df.date < test_start].drop_duplicates().nlargest(embargo_td).min() tr = df[df.date < cutoff]; te = df[df.date.dt.year == y] m = lgb.LGBMClassifier(n_estimators=400, learning_rate=0.03, num_leaves=31, min_child_samples=300, subsample=0.8, colsample_bytree=0.8, subsample_freq=1, verbose=-1) m.fit(tr[feats], tr.y) out.append(te.assign(score=m.predict_proba(te[feats])[:, 1])) return pd.concat(out) def precision_at_k(res, k=20): r = res.groupby('date').apply(lambda g: pd.Series( {'p_at_k': g.nlargest(k, 'score').y.mean(), 'base': g.y.mean()})) return r.groupby(r.index.year).mean() # 按年看，观察 regime 差异 # ---------- 主流程 ---------- if __name__ == '__main__': fetch_all() raw = load() raw = raw[raw.groupby('ts_code').date.transform('count') > 300] # 剔除上市不久的票 rallies = pd.concat(find_rallies(g) for _, g in raw.groupby('ts_code')) print(rallies.groupby(rallies.start.dt.year).size()) # 每年主升浪数量 print(rallies.limit_up_ratio.describe()) # 用于分型 feats_df = pd.concat(make_features(g) for _, g in raw.groupby('ts_code')) df = raw[['ts_code', 'date']].join(feats_df).dropna() for f in ['range120', 'tr20_pctile', 'vol_ratio', 'log_mv', 'n_lu60']: print(f, '\n', lift_table(df, f)) feats = [c for c in df.columns if c not in ('ts_code', 'date', 'y')] res = walk_forward(df, feats) print(precision_at_k(res)) ``` 跑通后的迭代路径：加行业/市场情绪特征 → 接入公告/龙虎榜做触发归因 → 按类型分模型 → 接入回测引擎（Backtrader / 自写事件驱动）算净值 → 模拟盘。 --- ## 八、实盘设计与预期管理 **日常流程**：收盘后 16:00 跑数据更新和打分 → 输出候选池（10~30 只）+ 每只的触发原因标签 → 次日集合竞价或开盘执行 → 涨停/大幅高开的跳过。 **风控比信号更重要**： - 单票仓位 5%~10%，分散在 ≥3 个板块 - 止损 −8%~−10% 或跌破 MA20；止盈用移动止盈（回撤超过前高 15% 出） - 市场情绪退潮期（涨停家数骤降、炸板率飙升）降仓或空仓 **合理预期**：主升浪预测是典型的低精度、高盈亏比问题。一个不错的模型能把 60 日内 +50% 的概率从全市场 3%~5% 提升到 Top 池的 15%~25%，靠的是少数大涨覆盖多数小亏。如果你回测出 50% 以上的胜率，先怀疑数据泄露（尤其是概念板块和财务数据日期），而不是庆祝。最后一点：这个系统最有价值的产出其实不是模型分数，而是**第三步那张归因表**——它会把你对A股主升浪的经验从“感觉”变成“有多少样本、多大概率、在什么环境下成立”，这本身就是巨大的进步。

我想设计一个真实的能跑的量化模型 运行在大a股上 旨在通过分析k线图以及相关数据 找到主升浪的潜在条件 触发原因 我应该...

我想设计一个真实的能跑的量化模型运行在大a股上旨在通过分析k线图以及相关数据找到主升浪的潜在条件触发原因我应该...