Quant Finance Notes

1. Alternative Data

  • 另类数据与分析师盈利预测

    • Dessaint, O., T. Foucault, and L. Fresard (2022). Does alternative data improve financial forecasting? The horizon effect. Working paper.
    • 假设分析师在进行盈利预测时，需要最优地分配其投入到不同时间尺度预测的精力，从而最小化预测误差以及获取不同时间尺度预测信息的成本这二者之和。另类数据的出现降低了获取短期预测数据的成本，并同时提高了短期预测数据的准确度。因此，它促使分析师将更多的精力投入到获取和分析短期预测信息上，以此来提高短期预测的准确度。然而顾此失彼，由于分析师的精力是有限的，这造成的后果是降低了他们长期预测的准确度。

  • Wolfe Research | Global Stock Selection with Proprietary Global Trademark Data

    • USPTO, foreign applications, Madrid filing, international registration
    • Trademark is not a proxy of advertising spending. Firms with very high advertising spending may over-spend and suffer from agency problem.
    • Signal transformations: growth rate, vintage ratio (long term number / short term number), long/short lookback window; residualizing size (log revenue) and sector.
    • Use life cycle of trademark to construct signals: applications, success rate of application, renewal ratio, trademark age, age dispersion, dispersion in trademark category, secretive foreign priority application (to keep trademark information under protection, and re-file in US using international registration)
    • Trademark category to construct linkage.

  • Wolfe Research | The Intangible Asset Premia

    • KC (Knowledge Capital): accumulate past R&D spending with discount (perpetual inventory method)

    • OC (Organizational Capital): accumulate 30% of past SG&A expenses.

    • IAI (Intangible Asset Intensity), KCI, OCI:

    • OCI+KCI combined has consistently more relevant than growth factor (as risk factor).

    • Firm with higher IAI score has higher momentum scores.

  • Wolfe Research | Patent, Innovation, and Alpha

    • Innovation industry: all industries where at least 50% firms have patent grants. High R&D spending (log R&D residualized by log market cap) is negatively correlated to future return, but is positively correlated to future return in innovation industry group.
    • Signals:
      • number of unique patent class / number of patent
      • Inward citation number, unique inward citation assignee; Outward citation number, unique outward citation assignee
      • Citation as linkage; patent class as linkage.
      • Patent maintenance.

  • Wolfe Research | Innovation Relevance

    • Total citations for all patents that a firm has cited to form a firm i's knowledge base (we can normalize within each patent class for citation) where indicator of indicates whether a patent has been cited by firm i, and is the total number of citations of patent at time . Then technological obsolescence is defined as the rate of change over a window .

    • Investors disagree most on high innovation relevance firms, we can assign higher risk budget to these firms.

  • Lerner, Josh, and Amit Seru. "The use and misuse of patent data: Issues for finance and beyond." The Review of Financial Studies 35.6 (2022): 2667-2704.

    • Patent Data Challenges: Patent data is valuable but can be problematic due to truncation (not all patents are granted by the end of the study period) and changes in inventor composition over time. (Both patent grants and patent citations).
    • Biases in Aggregation: When patent data is aggregated at the firm level, biases can persist even after common adjustment methods are applied.
    • Correlation with Firm Characteristics: Patent and citation biases are correlated with firm characteristics such as size, market-to-book ratio, and R&D intensity.

  • Shu, Tao, Xuan Tian, and Xintong Zhan. "Patent quality, firm value, and investor underreaction: Evidence from patent examiner busyness." Journal of Financial Economics 143.3 (2022): 1043-1069.

    • This paper attempts to study the causal effect of examiner busyness on patent quality and firm value. Using a broad set of patent quality measures, we find strong evidence that patents allowed by busy examiners exhibit significantly lower quality.

  • Bekkerman, Ron, Eliezer M. Fich, and Natalya V. Khimich. "The effect of innovation similarity on asset prices: Evidence from patents’ big data." The Review of Asset Pricing Studies 13.1 (2023): 99-145.

    • 科技关联度 (II)

    • Methodology:

      • Patent text analysis: use external sources such as Wikipedia and professional dictionaries to establish the professional termi- nology for every patent in our sample. This process enables us to define two patents as similar if they share the same professional terminology.
      • Remove “boilerplates” (i.e., long lists of terminology used in patent texts to illus- trate the invention generality).
      • Patent similarity: Two patents' distance is represented as a vector of their common terms weighted by a TFIDF (term frequency– inverse document frequency) variant.
      • Firm similarity: log of the sum of similar patent pairs discounted by the age of the newer patent in each pair and normalize it by the log of the product of the total number of patents for each firm in the pair.

    • 基本面解释：For R&D-to-Total Assets and ROA terms, peer firms have both contemporaneous correlation and prediction power.

    • 信息扩散缓慢的原因是投资者注意力不足，而不是投资者完全完全意识不到关联。注意力不足意味着投资者未来能认识到关联，因而会有信息的进一步扩散和关联动量。而后者意味着投资者压根就看不到关联的存在，因此也就没关联动量效应了。

  • Narrative Momentum

    • Data

      • The collected news articles are classified into media reservoirs: General, Corporate, FX, and Country Equity.
      • Articles are classified into 347 narratives: including 53 pre-specified Journal of Economic Literature (JEL) narratives and around 300 additional narratives. The 347 narratives are classified into 14 narrative tags including Geopolitics, Macro, Micro, etc. The narrative series are provided by MKT MediaStats, LLC.

    • Construction

      • Narrative Intensities: Negative (positive) intensity is the fraction of negative (positive) sentiment articles pertaining to a narrative out of the overall discussion, with a value in [0,1].
      • Narrative market beta: whether narratives can explain excess market returns. univariate regressions of the one-month market excess returns on contemporaneous one-month intensity changes.
      • stock-level narrative betas: univariate regression of stock return and intensity changes.

    • Conclusion:

      • Financial analysts also tend to underreact to narrative-sensitive stocks
      • Narrative momentum is different from price momentum.

  • Goldman, Eitan, Jordan Martel, and Jan Schneemeier. "A theory of financial media." Journal of Financial Economics 145.1 (2022): 239-258.

    • Firms are more likely to manipulate their announcements when media coverage is more extensive.
    • Negative news is more likely to be reported than positive news.
    • The presence of financial journalists can lead to more efficient pricing.

  • Froot, Kenneth, et al. "Predicting Performance Using Consumer Big Data." Journal of Portfolio Management 48.3 (2022).

    • Proxies for Corporate Sales: The authors construct three proxies for real-time corporate sales using distinct information sources: in-store foot traffic (IN-STORE), web traffic to companies' websites (WEB), and consumers' interest level in corporate brands and products (BRAND).
    • Predict SUE, SUR, Analyst forecast error.
    • Check analyst coverage and media exposure, and market attention level to see if the market consensus really matters, whether we want to trade the surprise from consensus or just the quarterly change.

  • Wolfe Research | Global shipping and supply chain alpha

    • S&P global panjiva supply chain intelleigence (US, Brazil, India, Mexico); FactSet (US only)
    • BOL: bill of lading form.
    • Features:
      • shipping volume (predict sales), (level/growth)
      • supplier, product, country of origin (diversity)
      • supply-chain network (company's position in supply chain)
      • shipping network momentum

  • Wolfe Research | Alpha insights from global job postings data

    • RavenPack dataset (most likely use LinkUp)
    • Term construction: job postings level/growth (we can use SOC median salary as importance weight); technical skill intensity, level/growth, uniqueness in skills, adoption of new technical skills, skill importance (TF-IDF).

  • DB Research | Macro and Micro JobEconomics

    • LinkUp job posting dataset: scrape from company website. Data since 2007, description data since 2014. Includes SOC job classification, geolocation data, and technical skills data. Most coverage in the USA.
    • Term construction similar for micro.
    • Macro: use to predict employment, PMI index, CPI, retail, consumer sentiment.

  • The Use and Usefulness of Big Data in Finance: Evidence from Financial Analysts

    • The paper offers robust evidence that the integration of alternative data into analyst reports is both increasing and economically valuable.
    • By improving earnings forecast accuracy and bolstering the trading commission revenue (as a proxy for Sell-side analysts’ value, data from ANcerno's institutional investor trade data ) of brokerages, alternative data serves as a tool to enhance the relevance of sell-side research in a rapidly evolving data landscape.

  • Armstrong, Chris, Yaniv Konchitchki, and Biwen Zhang. "Digital traffic, financial performance, and stock valuation." (2023).

    • Digital traffic (website pageview, visit, visit duration, other statistics) can predict quarterly earnings KPIs. It provides additional information to analyst consensus and market consensus (positive stock return).
    • All these effects only apply to firms with consumer-oriented websites. A website is defined as consumer-oriented if it satisfies at least one of the following conditions: (i) it allows direct transactions by customers. Examples include amazon.com, nike.com, and bestbuy.com; (ii) the website itself is the product, and visits to these websites are de facto consumption of products. Examples include nytimes.com, google.com, facebook.com, netflix.com; (iii) while the website is not directly geared towards customer transactions, the website hosts detailed product information and is likely to attract visits from existing and prospective customers. Examples include nissanusa.com and bmwusa.com.
    • The Authors define a alternative value measure: market-to-visits, market-to-pageviews.
    • Data from SimilarWeb.

  • Gómez-Cram, Roberto, Yunhan Guo, Theis Ingerslev Jensen, and Howard Kung. "Prediction Market Accuracy: Crowd Wisdom or Informed Minority?" SSRN (2026).

    • Setting: Polymarket — full on-chain transaction universe (every trade, account, contract, timestamp). Jan-2023 to Dec-2025: 98,906 events, 1.72M accounts, $13.76B volume.
    • Methodology — sign-randomization skill test:
      1. Take each account's full trade history; for each trade observe event, market, size, and price.
      2. Keep size and price fixed; randomize only direction (buy vs sell, 50/50).
      3. Randomize at the event level (critical): all trades by an account within one event share a single randomized direction. Handles (a) intra-market position-building (one bet built over many trades is one decision, not many), and (b) cross-market correlated bets within an event (e.g. "Trump wins" + "Harris loses" are perfectly correlated; treating them as independent would overstate evidence).
      4. Repeat 10,000× → null PnL distribution under "no skill."
      5. p-value = fraction of simulated PnLs ≥ actual PnL. p<0.05 →="" skilled="" (3.14%="" of="" accounts);="" p="">0.95 → unskilled (6.4%); in-between → lucky winners (29.0%) or unlucky losers (61.4%). Market makers (0.1%) identified separately.
    • Validation: among top traders by raw realized profit, only 12% overlap with the skilled set — raw PnL is a poor skill measure. OOS skill persistence 44% (vs ~10% for active mutual funds).

2. Analyst

  • Analyst Forecast Bundling Intensity and Earnings Surprise

    • Barth, Mary E., et al. "Analyst Forecast Bundling Intensity and Earnings Surprise." Available at SSRN 4839739 (2024).

    • The authors explore how financial analysts convey information about a company's earnings without necessarily making full revisions to their earnings forecasts. They achieve this by increasing what they term 'bundling intensity,' which refers to the extent to which an analyst's report that includes an earnings forecast revision also includes revisions to price targets and/or recommendations that have the same direction as the earnings forecast revision.

    • The researchers have developed a measure called BF_Score at the firm level to quantify bundling intensity. Their findings suggest that BF_Score is a significant predictor of earnings surprises based on analyst forecasts. These surprises often result from biases in consensus earnings forecasts, which are influenced by the information analysts communicate through bundling intensity. Below is definition of BR score, where TP is target price.

    • The use of bundling and the predictive power of BF_Score increase during times of higher macroeconomic uncertainty, when analysts have greater incentives to avoid bold revisions to their earnings forecasts.

  • He, Jie Jack, and Xuan Tian. "The dark side of analyst coverage: The case of innovation." Journal of financial economics 109.3 (2013): 856-878.

    • Firms covered by a larger number of analysts generate fewer patents and patents with lower impact.
    • The evidence is consistent with the hypothesis that analysts exert too much pressure on managers to meet short-term goals, impeding firms' investment in long-term innovative projects.

  • Kumar, Alok, Ville Rantala, and Rosy Xu. "Social learning and analyst behavior." Journal of financial economics 143.1 (2022): 434-461.

    • The paper extends the literature on analyst herding by demonstrating that sell-side analysts not only mimic peers’ forecasts for the same firm but also update their beliefs based on information gleaned from peers covering different firms within their portfolio.

  • Alpha in Analysts

    • Our empirical analysis shows that while the average analyst does not generate statistically significant alpha relative to the returns of a long-only portfolio benchmark, a subset of analysts exhibits persistent alpha. Motivated by this heterogeneity, we introduce a "fund-of-analysts" framework that first predicts analyst performance and then dynamically allocates weights across analysts based on predicted analyst performances.

  • Lopez-Lira, Alejandro, Yuehua Tang, Yuan Wang, and Mingyin Zhu. "What Makes a Star Analyst? Evidence on Industry Expertise from 1.2 Million Analyst Reports." SSRN (2026).

    • Target — Institutional Investor All-Star prediction: ~600 buy-side firms (~$15T AUM) vote annually on top sell-side analysts by sector. All-Star designation drives compensation, broker market share, and access — historically hard to predict from forecast accuracy alone.

    • LLM evaluation pipeline scores each industry report on 18 dimensions in two families:

      • General research quality: clarity of investment thesis, quality of evidence, internal logical consistency, transparency of assumptions, balance of bull/bear cases, quantitative rigor, treatment of risks/catalysts.
      • Industry expertise: knowledge of value chain, sector-specific operating metrics, regulatory/technology dynamics, competitor moves, channel/supplier insight, secular vs cyclical drivers.
      • Per-report 18-dim score vector → aggregated to analyst-level knowledge score.

    • Sample: 1.22M industry reports, 40 major brokerage houses, 2012–2024.

    • Conclusion: LLM knowledge scores predict All-Star status with coefficients ~7× larger than forecast accuracy. Buy-side stardom tracks textual industry expertise, not numerical accuracy — accuracy is a consequence of knowledge, not the driver.

3. Anomalies

  • Crowdsourced employer reviews and stock returns
  • Extrapolative beliefs in the cross-section: What can we learn from the crowds?
  • Chinese Stock Market Shell Value
  • Size and Value in China
  • Time Series Momentum
  • Tracking Retail Investor Activity
  • Overnight Return Reserval
  • Idiosyncratic Volatility
  • Volume
  • ESG
  • Loughran, Tim, and Bill McDonald. "Measuring firm complexity." Journal of Financial and Quantitative Analysis (2023): 1-28.
    • Measure firm complexity: use 10-K filing text data.
      • RHS: 374 pre-defined words related to firm complexity.
      • LHS: use audit fees (adjusted by size and industry) as complexity proxy.
      • Run lasso regression -> identify 50+ words as final firm complexity related set.
      • Complexity = percentage of complexity word set in 10-K filiing corpus length.

  • Cohen, Lauren, and Dong Lou. "Complicated firms." Journal of financial economics 104.2 (2012): 383-400.

    • Complication of a firm is measured by income segment.
    • The more complicated the firm, the more pronounced the return predictability. In addition, we find that sell-side analysts are subject to these same information processing constraints, as their forecast revisions of easy-to-analyze firms predict their future revisions of more complicated firms.

  • Taheri Hosseinkhani, Nima. "Options Volume as Noise: Evidence from Three Decades of Earnings Announcements." SSRN (2026).

    • Sample: 69,094 firm-quarter earnings events, 1996–2024. All signals averaged over [–60, –6] trading days pre-announcement.
    • Three option signals:
      • O/S Ratio = $V^{opt}_t / V^{stk}_t$ (Roll-Schwartz-Subrahmanyam / Johnson-So). Q5–Q1 hedge: –39 bps announcement (t=–3.79), –118 bps PEAD (t=–4.60). Why negative: aggregate flow dominated by hedging/retail noise, not smart money.
      • IV Spread = $\sigma^{C}{\Delta=50} - \sigma^{P}{\Delta=-50}$ (Cremers-Weinbaum, 30-day ATM IV diff from OptionMetrics surface). Significant univariate on SUE (t=–2.90) but collapses with controls (t=0.29). Cross-sectional dispersion widened post-2020.
      • PCR = $V^{put}_t / (V^{put}_t + V^{call}_t)$ (Pan-Poteshman). Only signal robust in full spec (t=–2.90). Directional measure, not level — survives retail noise.
    • Structural break in O/S around Oct-2019 zero-commission shift (FM CAR[0,+1]): 1996–2005 t=–2.04 → 2013–2019 t=1.31 → 2020–2024 t=+2.77. Institutional hedging dominated pre-2020 (high O/S = negative info); retail call buying dominates post-2020 (high O/S = bullish sentiment). Same statistic, inverted information content.

  • Posselt, Anders, and Mads Kjær. "Anomaly-Driven Demand." Working Paper (March 2026). — Alpha Architect summary by Larry Swedroe (May 22, 2026)

    • Core idea: Factor投资规模已经大到使其自身的机械再平衡行为成为驱动anomaly returns的一个独立通道——除经典的risk premium / mispricing / data-mining之外的第四种解释。
    • Signal construction (ADD = Anomaly-Driven Demand): 在Chen-Zimmermann数据集的209个anomaly上，每只股票每月统计 (净进入long leg数 − 净进入short leg数) 的边际变化，即下个月将面临的coordinated rebalancing压力。完全基于公开characteristics，不需要持仓数据。
    • Strategy & results: 按ADD分5档，年化超额回报单调从6.62%升至10.65%；long-short spread +4.03 ppt/年 (t=4.03), Sharpe 0.61。控制FF5+momentum+anomaly PC前三主成分+专门构造的anomaly-exposure组合后alpha仍显著。Anomaly层面，高低ADD imbalance组合差异+8.04 ppt/年。
    • Mechanism evidence:
      • 回报集中在月初前6个交易日且完全来自open-to-close (盘中)，机构月末再平衡的典型签名。
      • 价格压力永久 (12个月无反转)，与demand-based pricing一致 (类似index inclusion effect但规模更大)。
      • ADD形成后institutional breadth上升、short interest下降，验证ADD在追踪真实order flow。
    • Cross-section:
      • 高t值 (statistically robust) anomaly效应最强 (>5 ppt)；低t值anomaly的spread接近0——effect由capital weight而非signal multiplicity驱动。
      • 不对称：long-leg入选的price impact > short-leg (机构short约束)。
      • Size分布U型：micro-cap与mega-cap效应都强 (mega-cap年化~3 ppt显著)，mid-cap最弱。Value类信号贡献最大，与机构资金long-standing偏好一致。
    • Practical takeaways:
      • ADD可作为单股crowding indicator接入risk model / turnover控制。
      • 对高知名度anomaly的historical alpha应打折——部分回报为后续crowding inflation贡献。
      • 月初前6天 + intraday open-to-close是alpha decay与执行成本的关键窗口，建议把trading推前到月末执行。
      • 反馈环: 机械flow → 价格压力 → 推高observed return → 吸引更多资金 → 更大flow → ...

4. Asset Pricing

  • Q-Factor Model

  • Which Beta?

  • Stambaugh-Yuan Four Factors

  • P-hacking

  • Chan, Kam Fong, and Terry Marsh. "Asset pricing on earnings announcement days." Journal of Financial Economics 144.3 (2022): 1022-1042.

    • The paper provides evidence that the capital asset pricing model (CAPM) seems to hold on days when influential firms announce earnings, challenging the conventional wisdom that the beta-return relationship is generally flat in the market. The findings have implications for investors, suggesting that strategic trading around earnings announcements could yield significant returns.

  • Hu, Xiaolu, Malick O. Sy, and Liuren Wu. "A factor model of company relative valuation." Available at SSRN 3706995 (2020).

    • Relative Value Metric

      The paper defines the company’s relative value as: where MV is the market value of the company (constructed from total assets adjusted by replacing book equity with market capitalization for the common stock component) and TA is total assets.

    • The authors collect 23 descriptive measures (or “descriptors”) of firm characteristics covering eight broad categories:

      • Profitability: e.g., realized return on assets (RoA) and analyst forecast RoA.
      • Growth: e.g., long-term growth (LTG) forecasts, one-year and five-year historical growth rates.
      • Investment: e.g., capital expenditure ratios, R&D spending, retained earnings.
      • Liquidity: e.g., working capital ratio, slack ratio, various liquidity ratios, and trading liquidity.
      • Leverage: Measured by book debt-to-equity ratio.
      • Market Risk: Captured by the firm’s stock beta.
      • Size: Usually expressed as the natural logarithm of total assets.
      • Momentum: Measured by stock return momentum over six- and 12-month horizons.

    • The authors combine similar descriptors within each category into single valuation factors. They do this by first standardizing each descriptor (using winsorization and z-score transformation) and then averaging them to form a factor. When multiple descriptors exist within a category, a Bayesian weighted approach is implemented: Weights are estimated via a constrained regression (imposing that weights sum to one) with an equal-weighting prior.

    • Once the eight factors are constructed, the relative value of each company is modeled through a cross-sectional contemporaneous regression at each date where:

      • q_t is the vector of standardized logarithmic relative value across companies.
      • G_t is a matrix of industry dummy variables (based on 49 industry groups from SIC codes) to serve as a local bias correction.
      • F_t represents the matrix of valuation factors.
      • c_t contains the cross-sectional slope estimates (i.e., the market pricing of each valuation factor).
      • e_t is the regression residual that represents the temporary misvaluation (mispricing) of individual companies.

    • The regression residuals (e_t) from the model, which capture the deviation of a company’s actual relative value from the “fair” value predicted by the model, are interpreted as temporary mispricing. A long-short portfolio that goes long on companies with the lowest residuals (undervalued) and short on those with the highest residuals (overvalued) delivers strong performance

5. Behavorial Finance

6. Crypto

7. Event

8. Fundamental

  • Valuing Stocks With Earnings

    • Traditional earnings like GAAP earnings: high transitory volatility
    • Use Street Earnings:
      • Adjusts a company’s reported earnings to exclude non-recurring, non-operational, or one-time items. It aims to reflect the underlying, sustainable profitability of a business by filtering out short-term noise, providing investors and analysts with a clearer picture of long-term value creation.
      • Derived from analyst adjustments (e.g., I/B/E/S consensus estimates from Thomas Reuters)

  • PE / PB / PS

    • PE（市盈率）：核心逻辑：PE反映企业盈利能力的定价效率，适用于盈利稳定、可预测性强的行业。

      1. 消费行业（食品饮料、家电、零售）
        • 需求刚性，现金流稳定，盈利波动小（如伊利股份、贵州茅台）
        • 例如：消费行业PE通常基于长期稳定的净利润计算，适合用PE判断估值高低。
      2. 医药医疗行业
        • 老龄化趋势下需求持续增长，创新药企业成熟期盈利稳定（如恒瑞医药）
        • 注意：研发阶段的生物医药企业可能亏损，需结合其他指标（如PS）。
      3. 传统制造业（机械、汽车零部件）
        • 技术成熟，竞争格局清晰，盈利增长平稳（如三一重工）
      4. 公用事业（电力、燃气）
        • 垄断性强，盈利受政策调控，PE可反映长期现金流价值

    • PB（市净率）核心逻辑：PB衡量企业净资产价值，适用于资产密集或盈利波动大的行业。

      1. 强周期行业（有色金属、钢铁、化工）

        • 盈利受大宗商品价格影响大，PE在周期低谷时失效，PB更稳定（如中国神华）

      2. 金融行业（银行、券商、保险）

        • 资产规模大且易量化（如银行信贷资产），PB反映资产质量与安全边际（如工商银行）

      3. 重资产行业（房地产、航空、航运）

        • 固定资产占比高，PB可评估清算价值（如万科A）

      4. 科技硬件制造（半导体、消费电子）

        • 设备和专利等资产价值显著，但需注意技术迭代风险（如中芯国际）

    • PS（市销率）核心逻辑：PS关注营收增长潜力，适用于高投入、高增长但盈利滞后的行业。

      1. 新兴科技行业（人工智能、云计算、半导体）
        • 初期研发投入大，盈利周期长，营收增速替代盈利成为核心指标（如英伟达、特斯拉）
      2. 生物医药与医疗器械
        • 创新药研发阶段亏损，但市场潜力大，PS反映管线价值（如Moderna）

      3. 新能源与高端制造（锂电池、光伏）

        • 行业扩张期需大量资本开支，PS衡量市场份额争夺能力（如宁德时代）

      4. 互联网与平台经济（电商、社交媒体）

        • 用户增长优先于盈利，PS结合用户价值评估（如亚马逊、字节跳动）

  • Blankespoor, Elizabeth, et al. "Real-time revenue and firm disclosure." Review of Accounting Studies 27.3 (2022): 1079-1116.

    • Disclosure Patterns and Timing:

      • Withholding of Negative News Early in the Quarter:

        Managers are less likely to issue a revenue forecast (i.e., voluntary disclosure) when real-time abnormal revenue is negative during the early part of the fiscal quarter.

      • Increased Disclosure as the Quarter Progresses:

        As the quarter moves closer to the mandatory earnings announcement, the withholding of negative news diminishes. This change is likely driven by an increase in litigation risk, heightened analyst scrutiny, and the expectation that the impending public revelation will force managers to disclose bad news.

      • Asymmetry in Disclosure:

        The analysis shows that it is primarily the “bad news” (i.e., weeks with abnormal revenues in the bottom quartile) that is withheld early, whereas there is no significant increase in the voluntary disclosure of good news. This finding is consistent with classic disclosure models where economic incentives lead firms to delay negative information until external disciplinary mechanisms (like investor reaction or legal risk) compel disclosure.

    • Market Reaction and Insider Trading:

      • Delayed Incorporation into Stock Prices:

        Although the real-time revenue measure is strongly informative—as evidenced by its positive correlation with future abnormal returns—the market does not fully and immediately price in the information. The gradual “leakage” of performance signals over the quarter suggests a dynamic process of information disclosure.

      • Insider Trading Behavior:

        The paper documents that in weeks where there is significant abnormally negative real-time revenue and no corresponding public disclosure, insider managers are more likely to sell their shares. This behavior implies that managers might use their private information for personal gain when they choose not to disclose.

    • Role of Disciplinary Mechanisms:

      • The withholding and subsequent release of information are found to be more pronounced in firms characterized by:
        • High Analyst Coverage: Greater monitoring by equity analysts increases pressure on managers to eventually release negative information.
        • High Institutional Ownership: With more sophisticated investors monitoring performance, non-disclosure becomes costlier.
        • High Litigation Risk: The prospect of legal or reputational consequences forces managers to disclose adverse information as the quarter’s end nears.

  • Froot, Kenneth, et al. "What do measures of real-time corporate sales say about earnings surprises and post-announcement returns?." Journal of Financial Economics 125.1 (2017): 143-162.

    • Managerial Disclosure Behavior:The behavior captured by the PQS (Post‐quarter Sales (PQS): Sales activity occurring after quarter-end but before the earnings announcement) measure indicates that managers do not fully disclose all privately held post-quarter performance information at the earnings announcement. Instead, they understate positive signals—resulting in lower-than-expected announcement returns and delayed price adjustments in the post-announcement period. This may be driven, in part, by personal trading motivations.

  • Katselas, Dean, Jo Drienko, and Jayanth Paranji. "All Strings Attached: Firm-Level Insider-Trade Sequences and the Power of Sell Signals." SSRN (2026).

    • Isolated buy/sell = firm-month with insider net-trade and no same-direction activity in adjacent months (one-off). Buy/sell-string termination = first month after ≥2 consecutive same-direction months ends (the "last trade" of a sustained stance).
    • Form 4 1986–2024 × CRSP/Compustat. Buy side symmetric: isolated +1.0% CAR ≈ termination +1.2%. Sell side asymmetric: isolated –0.9% CAR vs termination –1.5% and more persistent, magnitude growing monotonically with string length. Returns also drift down during ongoing sell strings.
    • Rule 10b5-1 plan sales amplify (not dampen) the sell-side signal — counter to conventional wisdom.

9. Linkage

  • Chen, Xin, et al. "Attention spillover in asset pricing." The Journal of Finance 78.6 (2023): 3515-3559.

    • The paper leverages a unique feature of stock display on trading platforms in China, where the order of stock display is determined by the stock's listing code. This feature creates an attention spillover effect, where investors are more likely to notice and trade stocks with listing codes adjacent to those of stocks they currently hold.
    • The authors propose that overconfident investors, following positive investment experiences, are likely to increase their trading activities and are more likely to direct their attention to neighboring stocks on the display.

  • Jin, Zuben. "Business aspects in focus, investor underreaction and return predictability." Journal of Corporate Finance 84 (2024): 102525.

    • Conference call transcripts -> topic model -> firm similarity -> linkage signals

  • Zhang, Zhiyu, et al. "Uncovering interfirm links through textual topic similarity: A comomentum analysis in financial markets." The British Accounting Review (2024): 101446.

    • cross-firm similarity measure based on the various topics extracted from Management Discussion and Analysis texts

  • Feng, Jian, et al. "Economic Links from Bonds and Cross-Stock Return Predictability." Available at SSRN 4047776 (2022).

    • Main idea: linkage from bond market credit-rating comovements.
    • This study identifies a "market segmentation" effect between the equity and bond markets, showing that information from bond markets is often not incorporated promptly by equity market investors.
    • Firms are connected through "credit-rating comovements," defined as instances when two firms' bond ratings are updated in the same direction within a ±10-day window.

  • Chen, Xin, and Huaixin Wang. "News Links and Predictable Returns." Available at SSRN 4458612 (2023).

    • Main idea: news-implied linkages in China where firms are connected based on shared media coverage.
    • News-based links were established by identifying instances where two firms were mentioned in the same article within a 12-month window.
    • The authors perform robustness checks to validate these results, including a placebo test using shared media platforms, demonstrating that only specific news stories—not general media coverage—predict future returns.
    • They explore linkage complexity, showing stronger predictability when linkages are more complex (e.g., higher numbers of shared stories or connections).

  • Wang, Huaixin. "The Day and Night Tale of Momentum Spillover Effects." Available at SSRN 4179413 (2022).

    • Main idea: high overnight returns for peer stocks predict elevated opening prices for focal stocks, followed by intraday reversals, while peer intraday returns consistently predict positive future intraday returns for focal stocks.
    • Retail investors, who trade primarily on overnight information due to news salience, and professional investors, who engage in intraday trading, correcting the market.
    • Predictable patterns arise not only from underreaction but from a systematic interplay between the different investor types. Retail-driven overnight price distortions are followed by intraday reversals managed by professionals

  • Data-driven graph learning

    • Pu et al. (2023) Network Momentum across Asset Classes

    • features: 8 in total, MOM and MACD.

    • From Kalofolias (AAAI 2016) How to learn a graph from smooth signals, we can define a convex optimization problem: where is the feature matrix with -days lookback window, where is a diagonal matrix with . The graph adjacency matrix we want to estimate represents the network at day t for constructing network momentum, with the -th entry measuring the strength of similarity of individual momentum between asset i and asset j. In the objective function, the first trace term measures the spectral variations of on the learned graph adjacency matrix , encouraging connections between nodes with similar features. It is derived from Laplacian smoothness under the mild assumption that each column of ​​ is a low-pass graph signal.

    • The above is derived from: Consider a matrix , where each row resides on one of m nodes of an undirected graph G. In this way, each of the n columns of X can be seen as a signal on the same graph. A simple assumption about data residing on graphs, but also the most widely used one is that it changes smoothly between connected nodes. An easy way to quantify how smooth is a set of vectors on a given weighted undirected graph is through the function

      where denotes the weight of the edge between nodes i and j and is the graph Laplacian, being the diagonal weighted degree matrix. In words, if two vectors and from a smooth set reside on two well connected nodes (i.e. is large), they are expected to have a small distance so that is small.

    • In our empirical analysis, we combine distinct graphs learned from from five different lookback windows such that trading days as follows: (graph ensemble) .

      To mitigate the effects of scale differences in constructing network momentum, which may arise due to the difference in the number of connections certain assets have - with some connected to numerous other assets and others only to a few - we also apply a graph normalisation as follows:

      (graph normalisation) , where is a diagonal matrix with ​​.

    • Another way to solve above equation (#mjx-eqn-) is described in Pu et al. (2023) Learning to Learn Financial Networks for Optimising Momentum Strategies.

      The Algo L2G reformulates optimisationbased graph learning into an unrolling neural network. By leveraging the inherent modularity of neural networks, where different layers can be easily stacked for forward propagation, we propose to incorporate an additional layer into L2G for directly constructing network momentum.

      A upgrade version of L2G algorithm is described in Pu et al. (NIPS 2021) Learning to learn graph topologies

  • de Bodt, Eric, B. Espen Eckbo, and Richard Roll. "Competition shocks, rival reactions, and stock return comovement." Journal of Financial and Quantitative Analysis, forthcoming, Tuck School of Business Working Paper 3218544 (2024).

    • Methodology: The authors use a novel test statistic to discriminate between two hypotheses: increased product differentiation (H1) or increased standardization (leading to decreased return comovement) and cost-cutting (leading to increased return comovement) (H2). They exploit changes in stock return comovement following tariff cuts to infer strategic reactions.
    • Empirical Findings: The study finds that tariff cuts lead to a significant increase in return comovement, particularly among "followers" within an industry, suggesting a move towards greater standardization and cost-cutting strategies (H2) rather than increased product differentiation (H1).
    • Leader definition: sales-based market shares, financial ratios and R&D.

  • Eisdorfer, Assaf, et al. "Competition links and stock returns." The Review of Financial Studies 35.9 (2022): 4300-4340.

    • Consider a firm’s competitiveness based on the manner by which other firms mention it on their 10-K filings.
    • C-Rank: 10-K cross-mention graph + page-rank algo -> preferred measure of firm-level competition rank
    • A firm’s effective competition status stems mostly from competing with companies outside of its sector.
    • C-Rank might identify an element of a firm’s risk profile. If the firm is “targeted” by strong competitors, it can increase the uncertainty about the firm’s future performance and value, then the outperformance of high C-Rank firms might manifest compensation for risk.

  • Yamamoto, Rei, Naoya Kawadai, and Hiroki Miyahara. "Momentum information propagation through global supply chain networks." Journal of Portfolio Management 47.8 (2021): 197-211.

    • Use factset supply chain data

    • Customer Momentum: Here we assume that a company has N customers, and let be the sales ratio; thus, customer momentum is defined by the following:

    • Weighting Method Based on Network Centrality: almost all of the sales ratios are unavailable. Thus, we use network centrality in network theory as the weight of customer momentum. Let be the edge betweenness centrality between supplier i and customer j.

    • Multilayer Customer Information

  • Earnings Propagation Effects through the Global Supply Chain Network

    • In this paper, we separate the regression model that examines propagation from customers and the regression model that examines propagation from suppliers and then estimate the coefficients of the following two regression models:

  • Deep GNN methods

  • Zhang, Chao, et al. "Graph-based methods for forecasting realized covariances." Journal of Financial Econometrics (2024): nbae026.

    • Previous HAR-DRD method: decomposing the return covariance matrix into the diagonal matrix of realized volatilities and the correlation matrix:

      where is the diagonal matrix with the elements of the square roots of on the main diagonal, that is, , and . is the correlation matrix. We can estimate above using

      where is the dimensional vectorized version of the lower triangular part of and (resp. ) is computed as (resp. ).

    • Using graph information, we can estimate variance and correlation as follows,

      where is the normalized adjacency matrix. Specifically, is a adjacency matrix indicating the connections between assets with diagonal elements as 0 , and , where . Therefore , represent the neighborhood aggregation over daily, weekly, and monthly horizons. represent the effects from connected neighbors over different horizons. Moreover, we apply the idea of the graph effect to modeling correlations according to the model

      where is the normalized adjacency matrix. Specifically, is a adjacency matrix indicating the connections between pairwise correlations with diagonal elements as 0 , and , where ​.

    • Choices of graphs

      • Variance: Complete, Sector, Graph-Lasso
      • Correlation: Complete, Line graph

      Given a graph , its line graph is a graph such that

      • each node of represents an edge of ;
      • two nodes of are adjacent if and only if their corresponding edges share a common endpoint in ​.

    • Extension: Graph neural networks for forecasting multivariate realized volatility with spillover effects

  • He, Wei, et al. "Similar stocks." Available at SSRN 3815595 (2021).

    • Similarity between two stocks is measured by the distance between their characteristics such as price, size, book-to-market, operating profitability, and investment-to-assets.
    • Retail investor behavior, including attention spillover and categorical trading, plays a significant role. Retail order imbalance increases for high similar-stock return portfolios, reflecting stronger demand from individual investors.
    • The similarity effect is stronger among firms with low institutional ownership, suggesting retail investors are primary drivers.

  • Guo, Li, et al. "Joint news, attention spillover, and market returns." arXiv preprint arXiv:1703.02715 (2017).

    • Joint news has a higher degree of attention spillover than self-mentioned news. Measured by increase in Google search activity and EDGAR filings for connected firms, more so than self news.
    • Define the degree of investor attention spillover to a given firm i, from firms connected to i through joint coverage, , as the centrality-weighted (node centrality) sum of abnormal joint news coverage across the connected firms.
    • Both JointNews and Market-JointNews (aggregation to market level) istrongly and negatively predicts the one-month-ahead (market) return. It increases the overall attention to the firms and resulting in high valuation and low future stock returns.

  • Huang, Shiyang, Tse-Chun Lin, and Hong Xiang. "Psychological barrier and cross-firm return predictability." Journal of Financial Economics 142.1 (2021): 338-356.

    • When a firm’s economically linked firms have good (bad) news, and its stock price is near (far from) the 52-week high, it has an underreaction to the good (bad) news about economically linked firms.
    • The nearness to the 52-week high significantly moderates the predictability of supplier returns based on customer returns. Long-term returns for firms close to their 52-week high are higher when customers exhibit strong performance.

  • Menzly, Lior, and Oguzhan Ozbas. "Market segmentation and cross‐predictability of returns." The Journal of Finance 65.4 (2010): 1555-1580.

    • The extent of cross-predictability is negatively related to the level of information in the market, measured by the level of analyst coverage or by the level of institutional ownership.
    • Analyst coverage measure: analyst is considered actively engaged in a stock for a 12-month period after making an EPS forecast on that stock.
    • Institutional ownership measure: sum the holdings of institutional investors in the stock at a given quarter-end report date and then divide by the number of outstanding shares. An alternative proxy is the number of different institutional investors in the stock.

  • Jaiswal, Rishabh, Shubhankar Sanyal, and Palash Jariwala. "Heterogeneous Pricing of Supply Chain Risk in Production Networks." SSRN (2026).

    • NSCR (Network Supply Chain Risk) = combination of:
      • BEA input-output tables — industry-level I-O weights capturing how much of industry j's output goes into industry i (so for firm i, its key upstream inputs).
      • FRED macro shocks — time series of shocks to those upstream input markets (energy, metals, chemicals, etc.).

10. Machine Learning

  • Out of Sample Predictability

  • Quant Machine Learning

  • Time Seris Machine Learning

  • NLP in Finance

  • Remlinger, Carl, et al. "Expert aggregation for financial forecasting." arXiv preprint arXiv:2111.15365 (2021).

  • Liu, Quan, et al. "PREDICTION OF EARNING SURPRISE USING DEEP LEARNING TECHNIQUE."

  • Cong, Lin William, Tengyuan Liang, and Xiao Zhang. "Textual factors: A scalable, interpretable, and data-driven approach to analyzing unstructured information." (2019).

    • Textual Factor (TF) Generation: The authors generate TFs through three main steps:
      • Representing text using vector word embedding (Word2Vec).
      • Clustering these vectors using Locality-Sensitive Hashing (LSH) to identify topics.
      • Applying topic modeling to identify interpretable textual factors. (Use topic exposure as latent factors, apply standard factor analysis framework)

  • Detecting Misreported Accounting A Machine Learning Approach using Text Data

    • 10-K filing MD&A part -> extract text -> train on SEC AAERs misreported identifier

  • Global and local fitting

  • Chai, Bailin, Fuwei Jiang, Lingchao Meng, Tian You, and Guofu Zhou. "Generative AI for Finance: A New Framework (RPBERT)." SSRN Working Paper (2025).

    • Main idea: Treats the cross-section of stocks like language. Firms are "tokens," sorted by each characteristic to form ranked sequences. Pricing-relevant information is encoded not only in a firm's own characteristics but in its position and interactions within the cross-section. Attention endogenously determines how information enters pricing — no fixed mapping from characteristics to returns.
    • Model structure — two stages:
      • Stage 1 (Pretrain Company BERT): 12-layer BERT Transformer (hidden 768, 12 heads, FFN 3072, GELU, dropout 0.1). Input = firm-ID tokens sorted by each characteristic per month; token embedding = learnable firm-ID lookup + segment embedding + position embedding (position encodes rank in the characteristic-sorted sequence). Pretrained on 1980–1989 data with MLM (predict masked firm tokens from peer context) + NSP (distinguish adjacent vs. random windows). 130 epochs, batch 128, AdamW lr=1e-4, weight decay 0.01, OneCycleLR, mixed precision.
      • Stage 2 (Fine-tune with conditional asset pricing): Mention aggregation pools all contextual states of a firm across 94 characteristic-specific sequences into a single firm-level embedding. β-network (5-layer MLP, hidden 512, ReLU) maps firm embedding → K-dim factor loadings. α-network (3-layer MLP, hidden 128, ReLU) maps cross-sectional return vector → K-dim latent factor realizations. Predicted return: r̂ᵢₜ = β̂ᵢ,ₜ₋₁ᵀ f̂ₜ. Loss = MSE on realized returns, backpropagated end-to-end through pricing module and BERT encoder jointly. AdamW lr=3e-5, weight decay 1e-4, gradient clipping 1.0.
    • Data: 94 firm characteristics (Green et al. 2017) — size, value, profitability, investment, momentum, volatility, liquidity, accounting signals (61 annual, 13 quarterly, 20 monthly). Monthly U.S. equities 1980–2023. For each month × characteristic, firms sorted in descending order; long lists split via sliding window (length 128, overlap 25). Tokenized corpus ~100GB. Rolling estimation: 10-year train, 1-year validation, 1-year test, re-estimated annually.
    • Results: RPBERT5 (K=5): total R²=17.92%, predictive R²=1.79%, VW long-short 2.54%/mo (Sharpe 2.85), EW 4.23%/mo (Sharpe 3.50). VW alpha 2.46–2.58%/mo after CAPM/FF5/FF5+mom adjustment (t>12). Tangency portfolio Sharpe 4.91. Shuffling rank order causes material performance collapse — ordering is load-bearing. Combining with AE5 (simple average) pushes R² to 19.33%, confirming embedding signal is complementary to numerical features.
    • Attention diagnostics: Different heads specialize in distinct economic dimensions — forward-looking expectations (EPS forecasts, revisions), liquidity frictions (illiquidity, bid-ask spread), investment dynamics (capex, depreciation, asset growth), intangible intensity (R&D, organizational capital).
    • Open questions: generalization across markets, turnover/implementation costs, sensitivity to characteristic ordering, the α-network taking contemporaneous returns as input (look-ahead in factor extraction).

  • Wade, Rylan. "Do Better Volatility Forecasts Lead to Better Portfolios? Evidence from Graph Neural Networks." arXiv:2605.19278 (2026). — Code

    • Data & target: 465只S&P 500股票, 2015-2025 周度realized volatility。Target = 下一周realized vol (MSE loss)。训练2015-2022 / 验证2023 / 测试2024-2025共103周walk-forward。
    • GraphSAGE backbone: 所有三个GNN变体共享同一GraphSAGE backbone (Hamilton et al. 2017)——inductive (归纳式) GNN，无需全图spectral decomposition，新股票加入也能inference，适合金融universe变动。
      • Layer rule (Mean aggregator): 对每个节点 $v$ 在第 $k$ 层：$h_v^{(k)} = \sigma(W^{(k)} \cdot \text{CONCAT}(h_v^{(k-1)}, \text{MEAN}{h_u^{(k-1)}: u \in N(v)}))$。多层叠加使节点能聚合2-hop / 3-hop邻居信息。
      • 超参 (hidden dim / dropout / layers / lr / correlation threshold / lookback windows) 在2023验证集grid search调优；HAR与LSTM作为固定baseline不调，确保差异来自graph构造而非architecture。
    • 三种图视角 (核心实验对比):
      • GNN-Correlation (动态): 边定义为 $|\rho| \geq 0.30$ 的rolling Pearson return相关。No-macro用252d窗口；macro版本分别报告21/63/126/252d。每周重算，最dynamic——平静期稀疏、危机期接近全连接。
      • GNN-Sector (准静态): 同一GICS sector内的股票互连，每年更新一次。表达基本面/经济结构关系，最稳定，turnover最低。
      • GNN-Granger (静态有向): 5-day lag Granger因果检验 + Bonferroni校正得13,886条有向边，整训练期一次性计算。Message passing保留方向性 (A→B = A过去对B未来有incremental predictive power)。
      • GNN-Ensemble: 三种GNN的逆MSE加权平均。
    • Macro-conditioned vs No-macro:
      • 架构相同，区别在node feature: macro版本在每个节点feature上concat市场层regime变量 (VIX level + change, SPY vol + return, 10Y-2Y spread, IG credit spread, average pairwise correlation, graph density)，所有股票共享同一份macro copy。Macro features用training-period统计标准化 (非cross-sectional)。
      • 实证: 加macro一律提升预测精度——MSE/R²/DA全面优于no-macro对应版本，且边际贡献大于不同graph构造之间的差异，说明regime conditioning ≥ graph topology的边际价值。
    • Forecast accuracy (103 test weeks):
      • 最低MSE: GNN-Corr + Macro 63d (MSE 0.0298, R² 0.210, DA 0.722)
      • Baseline: HAR per-stock 0.0329, LSTM 0.0324
      • No-macro GNN: 三个图差异有限 (0.0322-0.0337)
    • Portfolio performance (Min-Var, 10bps成本, 单股≤5%):
      • 最高Sharpe: GNN-Sector + Macro (Sharpe 0.984, 年化Ret 15.3%, vol 10.4%)
      • GNN-Granger + Macro Sharpe 0.973
      • GNN-Corr + Macro 63d (predict最优) Sharpe仅0.794
      • HAR per-stock Sharpe 0.635
    • Key empirical finding (核心结论): 最低forecast MSE的模型、最高cross-sectional ranking accuracy的模型、最高portfolio Sharpe的模型是三个不同模型。Forecast accuracy / ranking quality / portfolio performance相关但不可互换。Graph vol model只在portfolio rule能利用cross-sectional structure时才有增益。
    • 哪种图更好: 取决于评估维度——预测维度Correlation+Macro最优 (动态图捕获current co-movement)，组合维度Sector+Macro最优 (稳定稀疏图→低turnover→10bp成本下net Sharpe最高，avg turnover仅0.406 vs HAR per-stock 1.012)。提示更稳定的图对portfolio层有时反而更有用。
    • Practical implications:
      • 对"先训练vol预测、再插入min-var优化"的标准ML pipeline提出系统挑战——MSE最小化未必带来更好组合表现，建议直接对下游目标 (IC, Sharpe, turnover-adjusted PnL) 做end-to-end训练或multi-objective loss。
      • 加入macro regime context的边际收益普遍高于换用更复杂的图结构，是ML-for-vol的cost-effective优化方向。
      • Graph信息对portfolio的增益依赖constructor能否消化cross-sectional信号 (Min-Var利用level + cross-section, IVP只用level)，graph选择应与portfolio rule匹配考虑。

11. Macro

  • He, Wei, Zhiwei Su, and Jianfeng Yu. "Macroeconomic perceptions, financial constraints, and anomalies." Journal of Financial Economics 162 (2024): 103952.

    • 当主观预期上调的时候，财务约束更大的公司未来的预期收益率更低。当主观预期下调的时候，财务约束更大的公司未来的预期收益率更高。
    • Factor timing

  • BAB beta factor

    • Campbell, John Y., and Tuomo Vuolteenaho. "Bad beta, good beta." American Economic Review 94.5 (2004): 1249-1275.

      • CAPM贝塔分解为两个组成部分：一个反映市场对未来现金流的新闻，另一个反映市场对折现率的新闻。
      • 现金流Beta衡量股票收益与公司基本面现金流冲击的相关性。这类冲击反映企业盈利、分红政策或行业前景等长期、永久性变化对股票价值的影响。衡量股票对基本面长期风险的暴露，对应“坏Beta”，需高溢价补偿。
      • 折现率Beta衡量股票收益与市场折现率冲击的相关性。这类冲击反映投资者对未来现金流预期风险的短期调整（如利率变化、风险偏好波动），导致股价的暂时性波动。例如，美联储加息（提高贴现率）可能短期内压低所有股票估值，但长期影响有限。衡量股票对市场短期情绪或政策风险的暴露，对应“好Beta”，溢价较低。

    • Frazzini, Andrea, and Lasse Heje Pedersen. "Betting against beta." Journal of financial economics 111.1 (2014): 1-25.

      • 杠杆约束

        • 机构投资者的限制：共同基金、养老基金等机构常面临严格的杠杆限制（如监管要求或内部风控规则），无法自由借贷以放大投资规模。
        • 个人投资者的限制：普通投资者可能因保证金要求、信用额度或风险厌恶心理而难以有效使用杠杆。杠杆约束的现实背景

      • 当杠杆受限时，投资者无法直接通过借贷放大风险，转而通过调整资产配置比例间接实现类似效果：

        • 超配高Beta资产：高Beta资产（如小盘股、高波动股票）在市场上涨时涨幅更大，下跌时跌幅更深。投资者通过增持这些资产，可以在不借贷的情况下“模拟”杠杆效果，追求更高的收益潜力。
        • 低配低Beta资产：低Beta资产（如大盘蓝筹股、债券）风险较低，但收益弹性不足。
      • 需求推动高Beta资产价格虚高：当大量投资者涌入高Beta资产时，其价格被推高，导致预期收益下降。低Beta资产被低估：低Beta资产因需求不足而被抛售，价格被低估，预期收益上升。
      • BAB因子与小盘股效应并不矛盾：规模因子SMB和价值因子HML (在控制市场Beta后) 独立于市场Beta。
      • BAB因子通过做空高beta股票，做多低beta股票 (带杠杆)，来构造零beta投资组合。

    • Herculano, Miguel C. "Betting Against (Bad) Beta." arXiv preprint arXiv:2409.00416 (2024).

      • 结合以上两文，构造BABB因子。
      • 通过双重排序（Double-Sorting），同时筛选低Beta和低现金流Beta的股票：
        1. 第一层排序：按市场Beta（β）将股票分为高、中、低三组；
        2. 第二层排序：在每组内按现金流Beta（β_CF）再次排序，构建3×3组合；
        3. 最终策略：做多低Beta/低现金流Beta组合，做空高Beta/高现金流Beta组合。

12. Microstructure

13. Miscellaneous

14. Momentum and Factor Timing

15. NLP

  • Wolfe Research | Text mining unstructured corporate filing data
    • EDGAR 10-K and 10-Q filings (by section comparison)
    • Features:
      • Sentiment and tone analysis
      • chanes in sentiment
      • distance measures (YoY embedding/BoW)

  • Learning Fundamentals from Text

    • Use attention machenism to weigh the importance of different paragraphs in a document, focusing on those that are most relevant to market reactions. The document-level aggregated vector is then used to predict the target variable, which is the direction of stock returns around the filing date.

  • Lee, Charles MC, and Qinlin Zhong. "Shall we talk? The role of interactive investor platforms in corporate communication." Journal of Accounting and Economics 74.2-3 (2022): 101524.

    • Investor interactive platforms (IIPs) in China

    • random sample of around 50,000 questions and then employed a state-of-the-art BERT-based algorithm to classify the remaining 2.45 million postings.

      • Findings: About 80% of questions seek clarification or explanation about specific items in financial reports or company operations, 16.6% are comments or suggestions to management, and the remaining questions pertain to verifying rumors or addressing misunderstandings.

    • These platforms alleviate common investor challenges, stimulate trading, improve market liquidity, and enhance the informativeness of stock prices.

  • Cohen, Lauren, and Quoc Nguyen. “Moving Targets.” Available at SSRN 4736129 (2024).

    • Managers publicize performance targets (e.g., revenue, same-store sales, product metrics) in earnings-call presentations. When they fail to hit a given target, they often shift the discussion to a different metric—“moving the target” to ensure they still clear a self-set hurdle.
    • There is no immediate announcement reaction but there will be underperformance after moved targets. Shifts in non-financial targets (e.g., subscriber counts, product units) predict larger underperformance than purely financial ones.
    • Analyst attention: when analysts explicitly question a dropped target and management is forced to address it, the underperformance effect is attenuated, indicating that inattention drives the gradual price drift.

  • Sidhu, Karmanpartap Singh, Junyi Fan, and Maryam Pishgar. “Which Voices Move Markets? Speaker Identity and the Cross-Section of Post-Earnings Returns.” arXiv:2604.13260 (2026).

    • Conference call分section、分speaker提取FinBERT情绪，按发言人身份加权。权重通过实证拟合：Analyst 49%, CFO 30%, Executive 16%, Other 5%。
    • 不加权的document-level sentiment丢失大量信息；Q&A section的analyst提问及管理层回应是主要alpha来源；FinBERT完全替代传统Loughran-McDonald词典方法（combined specification中LM t=0.86，FinBERT t=5.90）。
    • OOS Spearman IC = 0.142（高于IS的0.115），月度long-short alpha 2.03%（FF5 unexplained, t=6.49），控制SUE后仍显著。
    • 实现上需要speaker-level transcript parsing，主流数据商（Refinitiv、S&P Capital IQ）的transcript均有speaker标注。

16. Portfolio Construction

17. General Knowledge

results matching ""

    No results matching ""