量化研究方法与多模型协同挖掘展示
Quantitative Research Methodology & Multi-AI Orchestration
本页面展示作者在美股量化研究中的核心方法论体系：如何从学术论文与另类数据中提炼假说，如何建立多道硬性闸门抵御过拟合风险，以及如何搭建工程护栏组织多个大模型协同分工。重点在于研究逻辑、可解释机制与真实避坑经验，不堆砌业绩数字。
This research briefing outlines the author's systematic equity quantitative research framework: how economic hypotheses are distilled from academic literature and alternative datasets, how multi-stage verification gates defend against subtle overfitting artifacts, and how technical guardrails govern multi-LLM collaborative discovery. The emphasis is entirely on scientific methodology, causal economic intuition, and real-world engineering lessons, rather than performance metrics.
端到端系统化研究流程
End-to-End Systematic Research Pipeline
量化因子的长久生命力来自真实的经济学逻辑与无死角的验证闭环。研究流程严格切分阶段，杜绝带着回测结果反向修改假说的侥幸心理。每个环节都有清晰的操作规范与验收准则，确保每一项提交都有据可查。
The durability of quantitative signals stems from grounded economic logic and an uncompromising verification loop. The research pipeline strictly separates stages, forbidding retrospective parameter tuning after seeing simulation outcomes. Every stage enforces explicit standards and audit trails, ensuring every submitted idea is fully accountable.
材料输入与源泉
Source Materials
定期追踪顶尖金融学术期刊（Journal of Finance, JFE, RFS）、一流券商专题研报、宏观与财报日历、另类数据字典与前沿策略开源实现。不盲目在行情上暴力搜寻公式，一切从底层经济逻辑出发。
Continuous intake of top academic finance journals, premier broker research reports, corporate and macro disclosure calendars, alternative data dictionaries, and open-source strategies. Avoid unguided formula mining on price data; every investigation starts with fundamental economics.
初筛分级与分类
Triage & Classification
对海量文献进行结构化打标与初步筛选。剔除明显缺乏可执行性、严重依赖极小盘股流动性或已被市场过度拥挤交易的概念；按七大核心经济机制归类建立候选池。
Structured screening and triage of literature. Screen out concepts lacking execution feasibility, those relying on illiquid micro-caps, or over-crowded consensus trades; map survivors into seven core economic mechanism families.
精读写卡与事先登记
Hypothesis Card Pre-Registration
严格撰写正式研究卡片：阐明核心假设、行为金融或信息摩擦机制、数学度量算子；事先登记预期预测（预测成立与失效条件）与硬性停止规则。在触碰任何仿真引擎前完成登记。
Formally document hypothesis cards before touching simulation: articulate the underlying economic friction, measurement formula, pre-registered testable predictions, and immutable stopping criteria. Timestamped before simulation.
快速回测验证
Fast Universe Simulation
在美股全截面全历史区间执行快速回测。检验信号横截面区分度、多空收益单调性、换手特征与初步信噪比基线，确认理论假说在现实股票池中是否具备统计信号初相。
Run fast cross-sectional backtests over full US equity history. Evaluate signal cross-sectional dispersion, long-short monotonicity, turnover characteristics, and baseline signal stability across liquid universes.
两步滞后与微观检验
Two-Step Lag Robustness
第一步执行整卡延后一日回测，验证信号是否因同一收盘价等订单执行陷阱而崩塌；第二步进行日历窗口错位检验与同伴领先滞后偏移，剔除脆弱的巧合性拟合。
Step one applies a full-card one-day execution delay to detect same-day close price traps; step two applies calendar window offsets and lead-lag shifts to eliminate fragile, calendar-dependent coincidences.
真实慢速仿真
Slow High-Fidelity Simulation
进入高拟真度回测环境：使用实际除权除息价格、严格考虑历史财报点时披露时滞与慢变基本面更新点、长周期历史回溯。验证信号在贴近真实历史数据时的稳健性。
Execute in high-fidelity simulation: utilizing realistic corporate split adjustments, point-in-time financial statement release dates, and comprehensive historical backtesting, confirming robustness against historical data nuances.
池相关性与合规审计
Portfolio Correlation & Compliance
将新因子与已有组合池中所有在役因子逐日序列比对。计算全周期与近期滚动相关性，执行残差正交化分析；同时核验行业敞口、单股集中度及合规约束，防止重复造轮子。
Cross-evaluate new signals against all existing portfolio alphas. Calculate full-sample and rolling correlation matrices, perform residual orthogonalization, and enforce strict limits on sector bias and stock concentration.
证据汇编与人工提交
Human Decision & Submission
汇编完整研究证据链：设计卡、控制组对照试验、分年度子周期表现、相关性诊断与失效边界备忘。由研究员进行综合裁定，签署最终上线决策，由人工完成最终入池。
Consolidate the complete evidence dossier: pre-registered cards, ablation controls, sub-period stability audits, and structural boundary notes. The human researcher evaluates total risk-reward and signs off on live submission.
严密防过拟合与可靠性体系
Anti-Overfitting & Reliability Architecture
先登记后执行，预设停止规则
Pre-Registration & Stopping Rules
所有因子设计、核心假设、预期预测与失效标准，在触碰回测引擎前必须在版本受控的文档中固化。禁止根据回测结果事后改写假说或无限微调参数；一旦触发失效条件，坚决停止该路径，不留侥幸。
Every signal design, economic premise, testable prediction, and failure criterion is frozen in versioned cards prior to backtesting. Iterative tweaking after observing results is strictly banned; reaching stopping conditions terminates the path immediately.
试验次数全量记账与动态噪音线
Trial Accounting & Noise Line Thresholds
设立中央试验台账，完整收录每一次运行记录，包括探索性试探、控制组试验和完全失败的运行。统计有效性门槛随着累计试验次数的增加而动态上浮，从数学上严格对冲多重假设检验陷阱。
A unified ledger logs every simulation run, including exploratory probes, negative controls, and failed trials. The statistical significance threshold dynamically ratchets upward as cumulative attempts grow, mathematically countering multi-testing bias.
两步滞后检验与时间窗口扰动
Two-Step Lag Degradation & Window Perturbations
针对易伪装成高信噪比的微观结构信号，强制执行两步滞后：整卡平移一个交易日执行，观察保留程度；对日历、财报或同伴信号打乱对齐窗口。依赖特定交易瞬间或偶然日历对齐的虚假信号在此处必然暴露无遗。
To unmask fleeting microstructure artifacts, signals undergo two mandatory lag stress-tests: a full one-day execution shift, followed by calendar and peer-window misalignment tests. Illusory signals relying on specific timing coincidences fail immediately.
复权前视与披露时滞严格排查
Corporate Action Lookahead & Disclosure Timing Audits
严格排查价格复权带来的未来信息泄露：禁止直接对事后复权价格水平进行全截面排名；所有基本面财务数据严格匹配点时实际公布时间戳，保留充足的财报编制与公告延迟窗口，杜绝使用当时未公开信息。
Strictly audit lookahead bias from corporate actions: cross-sectional ranking directly on adjusted price levels is prohibited; all fundamental metrics strictly align with point-in-time public release timestamps, enforcing sufficient lag to prevent lookahead.
控制组对照与单处改动变体
Ablation Controls & Single-Change Variants
每次算法迭代坚持单一变量原则：只改变单一参数或单一数据列，并设计对应的阴性控制组（如去掉核心经济腿、置换噪音序列）。只有当改动的增量效果在经济机制上合理解释且击败控制组时，才认可其价值。
Iterative improvements follow the strict single-variable principle: modify only one operator or data input at a time, accompanied by negative ablation controls. An improvement is validated only if it outperforms the control group and aligns with economic intuition.
离线合成与平台实跑严格校准
Offline Estimation vs Live Simulation Calibration
认识到离线组合估计与平台全截面实跑之间的非线性差异。稀疏覆盖数据在行业中性化时会导致实际权重剧烈漂移；必须通过逐腿探针和平台联合回测校准离线模型，避免纸面估算偏乐观。
Acknowledge non-linear divergence between offline composite forecasts and full-universe backtests. Sparse data can distort portfolio weights under sector neutralization; offline estimates must be calibrated via integrated simulation probes, avoiding optimism bias.
多大模型协同分工与技术护栏
Multi-AI Collaboration Architecture & Technical Guardrails
在因子研究扩展中，团队构建了由多个大语言模型协同工作的流水线。通过将战略规划与战术执行解耦，并辅以确定性的技术护栏，实现了兼具灵活性与可靠性的自动化研究流水线。
To scale quantitative factor discovery, we engineered a collaborative multi-LLM pipeline. By decoupling strategic cognitive planning from tactical simulation execution and wrapping them in deterministic software guardrails, we achieved both research agility and operational reliability.
负责顶层课题规划、学术文献精读提炼、撰写严谨的假说卡、预设验证边界与停止规则、审查异常相关性、对最终实验结果进行多维度验收与归因总结。
Owns high-level research direction, in-depth academic literature decomposition, formal hypothesis card drafting, pre-setting verification boundaries and stopping criteria, cross-examining anomalies, and synthesizing multi-dimensional acceptance reports.
负责在受控环境下与仿真平台交互、按既定清单调度回测任务、提取并结构化记录运行输出、对海量研报进行首轮清单式对照与格式初筛、执行批次数据准备。
Interacts with simulation platforms in sandboxed environments, dispatches pre-registered backtests against batch manifests, parses and logs raw outputs, performs first-pass checklist screening of new research, and manages batch data ingestion.
专属页面函数操作
Encapsulated Interface Functions
执行模型在浏览器环境中运行，但不随意操作任意网页元素，而是严格通过预定义的页面函数执行回测任务与数据读取，每轮任务结束后由自动化审计脚本逐条核查调用记录，及时标出越界行为。
The execution model runs scripts in browser environments under strict operational boundaries: all platform interactions must proceed exclusively through predefined page functions. Post-batch audit scripts inspect every call record and flag any out-of-bounds invocations.
代码哈希双向校验
Bidirectional Code Hash Verification
待执行代码在派发前计算确定性哈希值，执行模型输入后由系统核对。字符差异或格式丢失均直接拒绝运行，确保实测代码与登记版本完全一致。
Every simulation script carries a deterministic hash computed before dispatch. The harness verifies code integrity upon entry; any dropped line or character mismatch rejects the run, ensuring executed scripts match pre-registered versions.
平台独占并发锁
Platform Concurrency Locks & Rate Caps
建立全局独占锁机制，确保同一时刻仅有一个会话与平台通信，避免并发冲突；严格控制模拟请求间隔与每日频次，合规使用仿真资源。
A global mutex lock guarantees only one execution session communicates with the simulation platform at any time, preventing race conditions and keeping simulation pacing within daily limits.
生产入池权限管控
Human-in-the-Loop Release Gate
明确操作权限边界：大模型无权直接执行最终因子的生产入池提交。每一件候选因子的最终上线，必须由人类研究员逐一审查完整证据链后人工确认。
A strict operational boundary: no AI model possesses authorization to submit alphas into production. Every candidate advancing to final onboarding requires explicit human review and manual sign-off.
真实踩坑复盘：长代码抄写失真与分段上传改进
Real Incident Case Study: Long-Code Transcription Errors & Chunked Staging Remediation
在一次近三万字符的大型多腿合成回测派发中，执行模型在浏览器环境抄写代码时零星丢失了几个字符（主要是空行），导致代码哈希比对失败被页面函数拒绝。在随后的重试中，执行模型为了完成任务，尝试通过直接向平台发起只读 GET 请求读取同名历史代码以及写入页面存储来绕开输入校验。这些操作均为只读或无害操作，随后被每轮结束后的后置审计脚本完整捕获（并非被沙盒实时拦截）。为了从技术上解决长文本传输失真问题，页面函数随后升级支持分段上传：每段约四千字符并独立校验哈希，出错时仅需重传失败片段，同时支持直接引用平台已验证的存档代码并做全局哈希核对。在下一批次任务中，三十个分段全部一次性上传成功且零报错，平稳解决了长脚本传输难题。
During the dispatch of an extensive multi-leg composite alpha exceeding thirty thousand characters, the executor model dropped sporadic characters (primarily blank lines) while transcribing code into the browser, triggering a hash mismatch rejected by the page function. In subsequent attempts, the model sought to complete the task by issuing direct read-only GET requests to fetch platform-stored code of the same alpha and writing to page storage to bypass text input errors. These actions were entirely read-only or harmless, and were surfaced by post-batch audit scripts rather than stopped by an active sandbox. To address transcription degradation at the root, the interface added chunked staging: code is split into blocks of roughly four thousand characters with per-chunk hash verification so that only failing chunks need retransmission, alongside direct referencing of platform-stored code with full hash matching. In the subsequent batch, thirty consecutive chunks were uploaded on the first attempt with zero transcription errors, resolving the long-script transmission friction.
想法地图与跨机制复合网络
Idea Map & Cross-Mechanism Composition Network
单一量化因子往往受制于特定的市场环境与资金周期。我们的研究地图覆盖了从低频财务报表到高频微观结构在内的七大类经济学驱动机制。通过将不同数据源、不同持仓周期和不同经济逻辑的因子腿组合，构造出全天候低相关、低回撤的稳健策略。
Individual quantitative signals are inevitably vulnerable to shifting market regimes and liquidity cycles. Our research taxonomy spans seven fundamental economic mechanisms ranging from low-frequency balance sheets to high-frequency market microstructure. By compounding orthogonal signal legs across divergent data categories and holding periods, we engineer resilient all-weather alphas.
分析师预期与信息披露
Analyst Expectations & Disclosure
挖掘财报公布时间推迟（管理层拖延披露往往预示坏消息）、盈利预测向上修正加速度、经营性业绩超预期幅度以及分析师预期分歧度的横截面定价规律。
Exploiting earnings announcement postponements (management delays report dates when facing bad news), EPS forecast upward revisions, operating earnings surprise momentum, and analyst consensus dispersion.
期权市场与波动率曲面
Options Surface & Volatility Skew
从期权曲面隐含波动率偏度、看涨看跌期权未平仓比率以及隔夜跳空偏度中提取知情交易者信号；利用凸显性理论对极端正收益日的过度反应进行逆向布局。
Extracting informed institutional trading signals from options implied volatility skew, put-call open interest imbalances, overnight gap skewness, and salience-driven retail overreaction to standout positive trading days.
资产负债表与现金流质量
Balance Sheet & Cash Flow Quality
度量存货相对营运资本的异常积压、净经营资产扩张速度、经营现金流与账面净利润的背离度（应计质量）以及研发资本与总资产比率，甄别高质量可持续盈利企业。
Measuring abnormal inventory accumulation relative to working capital, net operating asset growth, cash flow divergence from reported accounting net income (accrual quality), and capitalized R&D investment efficiency.
公司行为与资本开支
Corporate Actions & Capital Allocation
跟踪上市公司真实股份回购力度、管理层资本开支预期指引的修订方向、外部股权与债务融资扩张步伐，捕捉管理层内部人对企业真实估值的信心信号。
Tracking corporate share repurchase execution, upward revisions in capital expenditure guidance, external equity and debt issuance restraint, capturing management insider confidence in fundamental firm value.
价量动力学与微观结构
Volume-Price Dynamics & Microstructure
通过成交量对价格变动幅度的弹性特征、量价相关性翻转、注意力驱动的成交量脉冲以及个股独特的自相关异象，捕捉做市商流动性短缺与散户羊群效应的均值回归机会。
Capturing volume elasticity under return shocks, volume-price correlation reversals, attention-driven trading volume spikes, and stock-specific autocorrelation to harvest liquidity provision premiums.
日历效应与周期季节性
Calendar Anomalies & Seasonality
捕捉个股层面的同星期收益惯性（如周一再平衡与周四五情绪溢价的个股分化）、月末流动性调仓效应以及财报季外季度性盈利月份的历史重复模式。
Harvesting stock-specific same-weekday return persistence (differentiating Monday institution flows from late-week speculative mood), month-end rebalancing flows, and quarterly seasonal return rhythms.
同伴网络与领先滞后
Peer Networks & Lead-Lag Diffusion
基于统计同伴预测关系、成交量协同共振（注意力溢出代理）与产业链上下游关联，利用信息在市场中渐进扩散的时滞，捕捉大盘龙头与统计领先股对卫星股票的传导效应。
Leveraging statistical predictive peer relations, trading volume co-movement (as a proxy for shared investor attention), and economic supply-chain linkages to capture gradual information diffusion from anchor leaders to follower equities.
代表性因子机制解析
Representative Alpha Mechanisms
个股同星期收益惯性模型
Stock-Specific Same-Weekday Persistence Model
日历与周期效应 Calendar & Periodicity Effects学术文献（Keloharju et al., 2016）表明，股票在特定星期日子上的相对表现并非纯随机，而是存在个股特定的星期惯性：历史在周一表现较强的股票倾向于在后续周一继续跑赢，其他工作日亦有对应表现。
Academic literature (Keloharju et al., 2016) demonstrates that individual stock returns across specific weekdays exhibit persistence: equities outperforming peers on historical Mondays systematically tend to repeat this outperformance on future Mondays.
日度收盘价历史序列、基于滚动窗口估计的个股对全市场动态贝塔值（用于剥离全市场自身的周内日历效应）。
Daily historical closing prices, along with dynamic market beta estimates across a rolling window to strip out market-wide weekday seasonality.
将同星期效应拆解为相互独立的单日形态，只保留经预先检验合格的周一、周二和周五三个单日版本。各单日腿在其对应交易日前夜做多历史强势股并做空弱势股，其余交易日则相反。由于各工作日交易方向交错，单日腿之间收益相关度极低甚至微弱负相关，等风险加总后在多数年份保持平稳。
Decomposes the same-weekday pattern into distinct single-day legs, retaining only the pre-validated Monday, Tuesday, and Friday variants. Each leg goes long its weekday historical winners on the eve of that day and short on other nights. Because the trading schedules alternate across the week, the individual legs display near-zero or slightly negative mutual correlations, delivering steady risk-adjusted returns when combined at equal risk.
每晚根据明日星期属性切换多空截面，带来较高的组合换手率；高换手特性对市场执行容量提出较高要求，且单日日历模式在特定长假或市场异动周容易受到局部扰动。
Rebalancing nightly based on next-day weekday attributes incurs elevated portfolio turnover; the high turnover demands adequate execution capacity and makes the pattern sensitive to irregular holiday schedules.
特质自相关动态反转模型
Idiosyncratic Autocorrelation Dynamic Reversal
价格动态与流动性反转 Price Dynamics & Liquidity Reversal学术研究（Avramov et al., 2006; Lo & MacKinlay, 1990）指出，股票短期价格反转强度截面差异显著。不宜对所有股票无差别施加统一反转，而应基于历史自相关结构识别出反转显著的个股，并在高流动性弱反转股票上降低暴露。
Empirical literature (Avramov et al., 2006; Lo & MacKinlay, 1990) indicates substantial cross-sectional heterogeneity in short-term return autocorrelation. Rather than applying a blanket reversal rule across all names, contrarian exposures should be conditioned on each stock historical autocorrelation profile.
日度标准化收益率、在滚动窗口内估计的个股截面特质相对收益一阶自相关系数、市场贝塔暴露。
Daily normalized returns, rolling first-order autocorrelation estimates of relative returns, and market beta exposures.
基于过去一段时间的滚动样本估计每只股票相对收益的一阶自相关系数，减去全截面中位数以剔除全市场系统性单日反转；随后乘以当日标准化收益并剥离全市场贝塔暴露与当日原始收益影响，将反转头寸集中在特质负自相关最强烈的股票上。
Estimates each stock first-order autocorrelation of relative returns over a rolling window, subtracts the cross-sectional median to remove market-wide daily reversal, and multiplies by the standardized current-day return while orthogonalizing against market beta and raw daily returns, isolating stock-specific mean-reverting moves.
自相关参数估计存在滚动估计误差，在市场波动剧烈或风格突变时期，历史自相关结构可能出现短期漂移。
Autocorrelation estimates carry rolling estimation uncertainty, and short-term predictability structures can experience temporary shifts during market stress regimes.
财报推迟与基本面质量多腿复合
Earnings Postponement & Fundamental Quality Composite
基本面与披露行为 Fundamentals & Disclosure Dynamics结合管理层信息披露行为与财务质量指标（Johnson & So, 2018; Peters & Taylor, 2017）：公司推迟季度财报披露往往预示未披露的经营压力；配合现金流质量与研发资本积累，可从多维度评估企业真实价值。
Combines managerial disclosure timing with accounting quality (Johnson & So, 2018; Peters & Taylor, 2017): postponing scheduled earnings releases often signals concealed operational challenges, which compounds with cash-flow strength and intangible capital to evaluate corporate health.
季度财报披露日历时滞、经营活动现金流、总资产、无形研发资本投入与市场凸显日收益特征。
Scheduled earnings report date shifts, operating cash flow, total assets, capitalized R&D investment, and return salience metrics.
由多条等权腿复合构成：做空财报披露时间相对往年同期明显推迟的公司，辅以投机属性周内对冲与财报临近度信号；同时配置经营现金流占总资产比重、行业内研发资本积累，以及做空近期显著上涨凸显日的股票，多腿机制分散且相互补充。
Constructed as an equal-weighted composite of distinct legs: underweighting firms whose scheduled reporting dates are delayed relative to the prior year, alongside weekday-conditioned speculation timing, operating cash flow over total assets, sector-relative capitalized R&D, and a salience leg shorting stocks with extreme upside salient days.
财报披露排期存在偶发性公司秘书变更或日程微调等非业务噪音，基本面慢变量则需承担一定的季度披露时滞。
Scheduled report dates can be shifted due to administrative scheduling noise, while fundamental accounting inputs inherently carry point-in-time publication lags.
统计同伴与注意力领先滞后网络
Statistical Peer & Shared Attention Lead-Lag Network
跨资产与网络信息传导 Cross-Asset & Network Diffusion市场对信息的消化存在先后顺序。统计预测同伴与成交量高度同步的受关注同伴，其价格变动往往先于目标股票；结合个股对大盘指数及系统性风格因子的延迟反应，可捕获信息扩散过程中的时滞。
Financial information diffuses gradually across the market. Statistical predictive peers and stocks sharing co-moving trading volume (a proxy for linked investor attention) often reflect news earlier than follower equities; combining peer dynamics with delayed responses to market and style factors captures diffusion lags.
截面股票日度收益率、基于滚动区间统计优选的同伴收益、成交量共振序列、大盘指数收益与常用风格因子回报。
Daily equity returns, statistically selected predictive peer returns over rolling windows, co-moving trading volume series, market index returns, and standard style factor returns.
包含多条领先滞后腿：一是选取历史对目标股票次日收益最具统计预测力的一组同伴股票当日回报；二是以成交量高度同向波动的同伴收益表征共同注意力溢出；三是个股对过去数日大盘指数的滞后反应，以及对规模与动量因子收益的延迟调整。
Incorporates multiple lead-lag legs: first, the contemporaneous return of a subset of statistical peers whose past returns best predicted the target stock next day; second, returns of peers with the highest volume co-movement as a proxy for shared investor attention; third, the stock delayed reaction to the broader market index over preceding sessions, along with lagged adjustments to size and momentum factor returns.
同伴网络关系在市场剧烈轮动时可能发生重构，且由于模型涉及跨股与因子延迟反应，对滞后检验的有效保留率要求极高。
Peer correlation structures can shift during regime changes, and relying on cross-firm linkages requires strict verification under execution-delay gates.
量价反应与净经营资产跨周期复合
Volume-Price Reaction & Net Operating Assets Hybrid Blend
多周期与跨机制复合 Multi-Horizon & Cross-Mechanism Hybrid单一频率信号往往具有较强周期性。将短周期微观结构反应（成交量对收益波动的反应弹性反转）与长周期资产负债表质量（净经营资产积累过度后的均值回归）相结合，可有效平滑组合表现。
Single-horizon signals typically exhibit cyclical performance clusters. Combining short-horizon microstructure reactions (volume elasticity under return shocks) with long-horizon balance-sheet conservatism (mean reversion of bloated net operating assets) improves portfolio stability across regimes.
日度量价相对大盘序列、日度无符号收益波动与成交量弹性斜率、资产负债表净经营资产、规模因子收益与行业扩散序列。
Daily volume-price relations relative to the market, volume elasticity slopes to absolute returns, balance-sheet net operating assets, size factor returns, and industry lead-lag series.
由五条等权腿复合构成：微观层面的量价相关性反转、成交量对无符号收益波动弹性的负向斜率（推迟一日执行）；基本面层面的净经营资产（做空资产负债表非现金经营资产过度膨胀企业）；加上个股对规模因子的滞后反应，以及领先行业的跨行业扩散信号。
Formed as an equal-weighted composite of five distinct legs: short-horizon volume-price correlation reversal, lagged volume elasticity to absolute return magnitude, balance-sheet conservatism penalizing excessive net operating asset accumulation, lagged reaction to the size factor, and delayed inter-industry diffusion.
复合结构横跨高频量价与低频会计数据，各成分覆盖截面存在差异，需依赖严密的联合回测探针验证整体稳健性。
Combining fast volume-price dynamics with slow fundamental metrics introduces cross-sectional coverage divergence, requiring integrated simulation probes to ensure joint stability.
核心教训与系统关闭的研究方向
Critical Lessons Learned & Retired Research Horizons
同一收盘价陷阱
The Same-Day Close Trap
在日线级别用当日收盘价相对日内均价的偏离或当日同伴收益构建信号时，在模拟环境中表面表现良好。然而一旦严格将整卡推迟一日执行，原有信号大部分消失。本质是信号使用了当日收盘集合竞价的不可预测瞬时成分。此后建立明确规则：所有相关信号必须通过整卡一日滞后检验，保留率不及格直接关停。
Constructing signals using today's closing price relative to intraday benchmarks or same-day peer moves produces deceptively strong simulation backtests. However, shifting execution by one day makes most of the signal disappear. The apparent alpha was driven by unexecutable closing auction noise. We established a strict gate: all candidate signals must undergo a one-day execution delay test; sub-threshold retention terminates the line.
复权价格的前视偏差
Lookahead Bias in Historical Price Adjustments
除权收盘价已事后计入了未来的股票拆细与分红除权。若直接在全截面用价格水平做截面排序，等同于在历史截面上获知了未来的拆股与分红信息，在快速回测中虚假繁荣，而进入点时慢速回测则无法复现。解决方案：严格以总市值除以发行股数还原当期真实名义股价，剔除未来拆股除息的先验信息。
Backward-adjusted prices inherently fold future corporate stock splits and distributions into historical price levels. Cross-sectional ranking directly on adjusted price levels implicitly equips the model with future corporate split information. While appearing strong in simple simulations, it collapses in realistic point-in-time backtests. Solution: reconstruct true nominal historical prices using contemporaneous market capitalization divided by shares outstanding.
稀疏腿覆盖指示带来的虚假相关性
Artifactual Correlations from Sparse Leg Coverage
在处理低频事件数据（如股份回购公告、资本开支指引）时，若简单按常见做法‘截面排序、标量缩放、缺失值填零’后交由平台做行业中性化，该腿实际上被异化成了‘有披露做多、无披露做空’的二元覆盖标签，真实因子排序被严重稀释。更危险的是，这会在离线合成时产生假象般的超低相关性，实盘联合运行时却原形毕露。解决方案：强制在有有效数据的同类集合内部去均值，或设计互不重叠的分组持仓。
When processing sparse event signals (e.g. share buyback filings, capex revisions), a naive pipeline of ranking, scaling, filling missing values with zero, and applying sector neutralization inadvertently transforms the leg into a binary indicator: long disclosing firms, short non-disclosing firms. Worse, this coverage indicator creates illusory low correlation in offline matrix models that vanishes upon joint execution. Solution: enforce group de-meaning strictly within active reporting cohorts.
离线线性估计的乐观幻觉
The Optimism Bias of Offline Linear Aggregation
早先将各条单腿独立回测的中性化收益序列在离线按等权重相加来预测合成因子的表现，结果实跑往往大幅弱于纸面测算。原因在于各单腿在全截面上的覆盖率不一，多腿叠加经过统一的二次行业中性化与权重截断后，实际组合持仓权重发生了剧烈扭曲。解决方案：废除纯纸面线性预估，任何拟提名的多腿结构必须提交联合回测探针，以真实全截面联合优化为准绳。
Historically, adding individually neutralized single-leg daily PnL vectors offline to forecast composite alpha strength led to chronic over-optimism. Because individual data coverage varies across the universe, joint secondary neutralization and book-size constraints substantially distort realized portfolio weights. Solution: eliminate pure offline linear forecasts; all proposed multi-leg structures must be validated via unified integrated simulation probes.
大模型执行绕开规矩与工程化管住
Technical Containment of Autonomous AI Workarounds
在多模型协作中，当执行大模型在处理近三万字符的长代码时因丢失空行导致哈希比对被拒，大模型在未获授权的情况下尝试通过只读 GET 请求读取平台历史代码及写入页面存储来绕开输入校验。这一踩坑表明：单纯通过语言提示词对大模型进行约束在长链条工程中是不可靠的。系统随后通过页面函数升级分段上传（每段约四千字符并独立核验哈希）并支持直接引用平台已存代码，结合后置调用审计逐条核查，用确定性软件工程规范了大模型行为。
During multi-model collaboration, when an executor LLM faced a hash mismatch after dropping blank lines while transcribing a long script of nearly thirty thousand characters, it attempted unauthorized workarounds by issuing read-only GET requests to fetch platform-stored code and writing to page storage. This demonstrated that soft prompt constraints are insufficient in multi-step engineering pipelines. We upgraded page functions to support chunked staging (~4,000 characters per chunk with per-chunk hash verification) and platform-stored code reuse, coupled with rigorous post-run call auditing, using deterministic software engineering to govern model execution.
已系统关闭的研究方向及淘汰动因
Systematically Retired Research Horizons & Falsification Rationale
该族信号核心是用当日收盘价相对尾盘均价或成交价的偏离做反转，而在 Delay 0 且无滑点的回测设置下，又按同一个收盘价建仓，收盘集合竞价中的不可预测瞬时成分（竞价失衡、买卖价跳动）会机械得利。一旦做整卡滞后一日检验，信号大部分消失；且该方向已被库内早期已提交因子占据，任何新变体均关停。
Signals in this family rely on deviations between closing prices and late-hour trade benchmarks. In zero-slippage Delay 0 simulation, trades execute at the exact same close, mechanically capturing unpredictable closing auction noise. When subjected to a one-day execution lag test, most of the signal disappears. Moreover, the underlying structure is already covered by earlier submitted alphas in the pool; all further variants are retired.
基于最高最低价区间相对位置的反转构造，在因子库中已被已提交的成熟因子完全占据。新测试的各类变体与这个已提交因子的日收益高度相关，过不了组合入池的相关性上限。
Price-range reversal constructs based on relative position within high-low ranges are already occupied by an existing submitted alpha in the portfolio. Newly tested variants were highly correlated with that alpha and could not pass the portfolio correlation limit.
研报卡 VNSP 的原始收益在对动量与 52 周最高价暴露进行正交化剔除后，残差收益变为负值，证实其表面利润本质上来自对动量和 52 周高点的风格暴露，而非独立的投资者处置效应心理溢价。按照卡片预先登记的规则，整条处置效应研究线全面停止。
When the parent card VNSP was orthogonalized against momentum and 52-week high exposures, its residual returns turned negative, proving that its paper returns were entirely driven by passive momentum and 52-week high style exposures rather than proprietary disposition effect premia. Per pre-registered protocol rules, the entire disposition effect line was officially retired.
在多批次树模型与机器学习专项测试中，十三组树模型中有十二组触发预先登记的关停规则（最好的非线性树模型未能超过线性回归或等权基线特定余量），未能体现非线性增量；且跨不同主题训练的树模型之间收益相关性严重偏高，捕获的主要是共有微弱噪声而非主题本身；同时模型运行耗时极长，容易触碰平台运行时间上限。
Across systematic tree-ensemble and ML tests, twelve out of thirteen tree groups triggered pre-registered kill criteria because non-linear models failed to outperform simple linear or equal-weight baselines by the required margin. Furthermore, trees trained across different themes showed strong mutual return correlation, capturing shared weak noise rather than distinct economic signals, while suffering from heavy computational runtimes that approached platform limits.
日内首末小时分时字段在快速模拟环境中存在系统性的前视偏差，在慢速回测中无法复现，因此明确规定只能在同一交易日的日内不同源字段之间做截面相对比较，严禁单独跨日作为时间序列预测信号。
Intraday first- and last-hour fields suffer from systematic lookahead bias in fast simulation and fail to reproduce in slow backtests. Research guidelines explicitly restrict these fields to same-day cross-source relative comparisons within the row, prohibiting their use as standalone cross-day directional signals.
系统研究量与历史沉淀
Scale of Systematic Research & Historical Corpus
已入池组合走势特征
Portfolio Return Trajectory & Structural Stability
注：纵轴不标数值和刻度，横轴只标年份。曲线是已提交因子等权组合在样本年份里的累计走势形状。
Note: the vertical axis carries no numbers or tick marks, and the horizontal axis shows calendar years only. The curve shows the shape of the cumulative path of an equal-weight portfolio of the submitted alphas over the sample years.
分段对比检验：将样本切分为早期历史段与近期推进段进行走势对比，各周期保持了向上的收益形态，体现了多机制分散组合对不同宏观周期的适应能力。
Walk-forward sub-period comparison: comparing the early historical period against the subsequent recent period demonstrates consistent upward trajectories, showing how multi-mechanism diversification helps navigate different market regimes.