Skip to content

Research notes — revenue as a multi-factor model

Working research direction (post v0.16 / post Track-A validation). Framing, literature anchors, two identified gaps with found solutions, and a concrete minimal research design. Written 2026-09-11 from a literature + web sweep; citations verified at time of writing.

中文摘要见文末。

1. The framing

Sell-side and quant research decompose returns with multi-factor models (r_i = Σ β_k·f_k + ε, factors priced cross-sectionally). The open question this project keeps circling: can the fundamental quantity itself — revenue — be modeled the same way?

g(i,t) = α + β₁·own-momentum(i,t) + β₂·growth-reversion(i,t)
       + β₃·industry-cycle(i,t) + β₄·macro(i,t) + γ_i + ε(i,t)

Three properties make this more tractable than return factor models:

  1. No pricing requirement — factors are mechanical/causal inputs (GDP elasticity, penetration, ASP), not risk premia. Betas are physical quantities ("durable-goods orders +1% → revenue +0.4%") and far more stable than return betas.
  2. Signal-to-noise is inverted — returns are near-zero-SNR daily; revenue is persistent at annual/quarterly frequency (Track A pooled sMAPE 4–7% ⇒ high R²). The hard part is not noise, it is structural breaks (the NVDA Data Center lesson: the future is not in the training data).
  3. The toolkit is panel econometrics, not asset pricing — firm fixed effects, clustered standard errors, rolling validation; all standard.

2. What the project already has (≈40% of a thesis)

Piece In repo Role in the factor framing
244-company × FY2015-25 revenue panel examples/profile_validation/data/ the estimation sample
8-method battery incl. 3 profile-implied methods backtest + profile_methods.py the "experts" (factor realizations)
Industry profiles (fit classes, checks) revenue_model/industry.py cross-sectional conditioning (β₃)
Macro revision loop (elasticity, lag, evidence) qesa_adapter / macro_revision β₄ with transmission structure
Event layer (honest null at monthly/large-cap) news_impact / form8k_adapter event factor: documented zero
Pre-registered validation discipline proposal-profile-validation.md credibility

Track-A test-year evidence that factor structure exists: the best profile-implied method varies by fit class (DecelCAGR→strong 6.6%, GrowthRevert→adapt 3.6%, Damped→weak 5.1%) — i.e., "method × industry" behaves like exposures, not noise.

3. Gap 1 — formal estimation of method loadings

The gap: v0.16 hard-assigns one method family per industry profile. The upgrade is soft weights: estimate, on the panel, how much of each expert each industry (or series) should load on.

Found solution — FFORMA. Montero-Manso, Athanasopoulos, Hyndman & Talagala (2020), FFORMA: Feature-based forecast model averaging, International Journal of Forecasting 36(1), 86–92 (M4-competition winner; 450+ citations; R package pmontman/fforma). Two phases:

  1. extract time-series features (trend strength, seasonality, curvature, volatility, …);
  2. a gradient-boosting meta-learner predicts per-method weights minimizing combined loss — trained globally across series.

Mapping to us: features = damped-momentum score, mean-reversion score, volatility, CAGR, fit-class dummies; experts = the 8-method battery.

Honest clauses (from the forecast-combination literature):

  • The combination puzzle — four decades of evidence that estimated optimal weights often lose to the simple average (survey: Genre et al. 2012, "Combining expert forecasts: Can anything beat the simple average?", IJF; theory: Claeskens, Magnus, Vasnev & Wang 2016, A simple theoretical explanation of the forecast combination puzzle; recent: Elliott 2026, On why averaging beats optimal linear weights, JASA/JABES). ⇒ an equal-weight damped-family benchmark is mandatory; the meta-learner must beat it to earn its complexity.
  • Globalization is the escape hatch — Thompson (2024), Flexible global forecast combinations, IJF: pooling weight estimation across series is one of the few designs that beats equal weights. Our panel is exactly that setting.
  • Scale warning — 244 series × 11 annual points is tiny vs M4. The meta-model must be minimal (multinomial logit or heavily-shrunk GBM); deep mixture-of-experts (LeMoLE 2412.00053, WaveMoE 2604.10544, Ziel & El Mahtout 2605.10330's expert-loss gating) would overfit — cite, do not ship.

4. Gap 2 — testing against expectations (the surprise layer)

The gap: revenue forecasting earns its keep at the moment it disagrees with consensus. The test: (model forecast − consensus) vs realized surprise vs post-announcement drift.

Literature anchor. Jegadeesh & Livnat (2006), Revenue surprises and stock returns, Journal of Accounting and Economics 41(1-2), 147–171 (539+ citations). Four findings that double as our test-design checklist:

  1. revenue surprises relate to contemporaneous and future returns;
  2. analysts under-react to revenue surprises (slow revision);
  3. PEAD is stronger when revenue and earnings surprises align in direction (interaction to replicate);
  4. reactions are larger for R&D-intensive firms (context split).

The novel bit (not in JL2006): an independent bottom-up revenue model as a third input. Question: does (our forecast − pre-announcement consensus) predict (a) the signed surprise and (b) the drift, incremental to consensus dispersion/revision momentum? A genuinely new angle — 2006 had no such model.

Data path. WRDS / IBES Summary History (point-in-time monthly consensus, revenue estimates included). UIUC holds a WRDS subscription (Gies faculty publish with it; library data service bis@library.illinois.edu); student account on request. Free fallback for a live/forward version only: stockanalysis.com analyst estimates via the existing sa_adapter (no history). Estimize on WRDS as robustness.

5. Minimal viable research design

  1. Panel A (done): battery × fit-class validation — ship as docs/profile-validation.md.
  2. Panel B: FFORMA-mini — features → equal-weight benchmark vs shrunk multinomial-logit weights; pre-registered; report win/loss vs benchmark honestly (the puzzle clause).
  3. Panel C: with WRDS/IBES — consensus alignment panel; surprise direction hit-rate of (model − consensus) gap; PEAD double-sort (gap × earnings-surprise alignment, replicating JL2006 finding 3 with a model-based twist).
  4. Thesis-scale narrative: industry is a factor-loading conditioner, not a label — v0.16's profiles are the coarse prior, the estimated loadings the refinement, and the weak-fit redirect the risk model.

6. Reference list

  • Montero-Manso P., Athanasopoulos G., Hyndman R.J., Talagala T.S. (2020). FFORMA: Feature-based forecast model averaging. IJF 36(1) 86–92.
  • Genre V., Kenny G., Meyler A., Timmermann A. (2012). Combining expert forecasts: Can anything beat the simple average? IJF.
  • Claeskens G., Magnus J.R., Vasnev A.L., Wang W. (2016). A simple theoretical explanation of the forecast combination puzzle.
  • Elliott G. (2026). On why averaging beats optimal linear weights.
  • Thompson R. (2024). Flexible global forecast combinations. IJF.
  • Jegadeesh N., Livnat J. (2006). Revenue surprises and stock returns. JAE 41(1-2) 147–171.
  • Ziel F., El Mahtout B. (2026). Fast training of MoE for TSF via expert loss integration. arXiv:2605.10330.

中文摘要

  • 构想：营收本身可做多因子模型——因子无需定价（是物理弹性）、信噪比远高于收益（难点在结构突变而非噪声）、工具是面板计量（公司固定效应+聚类SE）。
  • 项目已有约四成：244 家收入面板、8 方法电池、行业画像条件、宏观弹性链、事件层（诚实 null）、预注册纪律。
  • 缺口一（方法载荷估计）→ FFORMA（2020 IJF，M4 冠军，被引 450+）：序列特征→元学习器预测方法权重，全局训练。诚实条款：组合预测谜题（等权平均极难打败，Genre 2012 / Claeskens 2016 / Elliott 2026）——等权基准强制在场；全局化估计（Thompson 2024）是少数胜出路径恰为我们设定；样本小（244×11），元模型必须极小，深 MoE 只引不用。
  • 缺口二（惊喜检验）→ JL2006（JAE 被引 539，四发现=检验清单：营收惊喜预测当期+未来收益、分析师反应不足、营收盈利同向 PEAD 更强、R&D 密集反应大）。创新点：以独立自下而上模型为第三输入，"(模型−一致预期)"gap 能否增量预测惊喜与漂移。
  • 数据钥匙：UIUC 可申请 WRDS / IBES Summary History（含 point-in-time 营收预期；图书馆 bis@library.illinois.edu）。
  • 最小可行设计：A 面板（已完成）→ B 面板（FFORMA-mini vs 等权基准）→ C 面板（consensus 对齐 + PEAD 双排序）。
  • 论文叙事：行业不是标签，是因子载荷调节器——画像是粗先验，载荷估计是细化，weak 档重定向是风险模型。