Skip to content

Industry-Fit Analysis — Where Driver Trees Work, and Where They Break

The flagship methodology document. The NVIDIA demo proves the thesis on real U.S. equity data; this document generalizes it into a practitioner's framework for when a driver tree is the right tool, when it isn't, and what to do instead.

中文要点见文末 中文摘要 小节。

Most revenue-forecasting tools sell you on accuracy. This one sells you on where it is accurate, and just as importantly, where it isn't. The honest answer turns out to be more useful than a confident wrong one.

The central claim, proven two ways below:

A driver tree's accuracy is a property of the industry's growth mechanism, not of the formula. It is excellent where growth continues its history, and structurally unreliable where a regime shift breaks that history — and no amount of trend-fitting closes the second gap, because the future is simply not in the training data.

1. The core claim

Driver-based (bottom-up) forecasting decomposes revenue into a product of measurable factors:

segment_revenue = market_base × penetration × share × price

The method is internally sound — if each factor is independently forecastable from its own history. That "if" does all the work. It holds cleanly for mature, trend-driven industries and breaks for event-driven ones. The mistake is to treat the method as universally right or universally wrong; it is neither. It is conditional, and the condition is whether the industry's growth mechanism is trend-extrapolatable.

Two natural experiments, one per market, make this unmistakable.

2. Industry-fit matrix

Fit Growth mechanism Examples Driver tree verdict
Strong Secular trend — factors continue their history Consumer electronics, semiconductors, auto parts, branded pharma Use it; back-test will be tight
Adapt Structured but different math SaaS (MAU × ARPU), Ads (traffic × eCPM), Subscriptions (churn × ARPU) Use the form, swap the factors
Weak / avoid Event- or cycle-driven Commodities (steel, shipping), Financials (balance-sheet driven), Resources (reserve-constrained), regime-shift tech Point-forecast will fail; jump to scenarios

The bottom row is not a condemnation — it is a redirect. For those industries the honest move is not a more clever trend, it is the scenario framework in §4.

3. NVIDIA: one company, two verdicts

The NVIDIA demo (examples/nvda_demo/) is the strongest possible test, because it removes every confounder: the same company, the same fiscal years, the same driver tree, the same engine, the same extrapolation rule. The only thing that differs between the two segments is the industry's growth mechanism.

Trained on FY2019-FY2023, hold-out FY2024-FY2025 (the AI breakout):

Segment Industry type Hold-out sMAPE FY2025 error
Gaming mature, trend 1.0% +4%
Data Center AI regime shift 60.0% −84% (actual $115B vs forecast $18B)

A 60× error gap on the same formula. Gaming is a trend market — GPU shipments, attach rates, and ASPs all continue their history, so the tree lands within 4%. Data Center is a paradigm shift — the post-ChatGPT acceleration is not in any pre-2024 trend, so extrapolation under-predicts by 6×.

This mirrors the earlier Luxun (立讯精密) result from v0.5.0: consumer electronics driver tree at −0.1% sMAPE vs automotive at −51% (Leoni acquisition). One A-share industrial, one U.S. mega-cap — the pattern holds across markets.

Insight: when the same method gives 1% on one segment and 60% on another of the same company, the variable is unambiguous. It is the segment, not the model. Blaming the model here would be like blaming a thermometer for a fever.

4. What to do when growth is event-driven

Point forecasting from history is the wrong tool under a regime shift. Replace it with uncertainty management — five techniques, ordered from cheapest to richest. revenue-model-builder supports the first directly and the third by API.

1. Scenario analysis. Stop asking "what is the number?" and ask "what range could it take, and what triggers each branch?" Build Bear / Base / Bull with explicit trigger conditions. The demo's close-out does exactly this: a wide Monte Carlo over honestly-uncertain Data Center drivers produces a distribution whose Bull tail (P90) captures the actual FY2024-FY2025 breakout — the point forecast collapsed, the scenario band framed the truth.

from revenue_model import simulate_segment, scenarios
mc = simulate_segment(seg, year, driver_ranges, n=5000, seed=0)
bear, base, bull = scenarios(mc)   # P10, median, P90 — straight from the distribution

2. Leading-indicator monitoring. Track the drivers of the drivers — the upstream signals (capex commitments, order books, capacity announcements) that telegraph a regime shift before it hits revenue. Pair with change-point detection or Bayesian structural time series.

3. S-curves / Bass diffusion. When adoption follows a saturation path, a logistic or Bass model beats a blind trend. The library ships Driver(...).fit_trend(...).extrapolate(...) with a logistic option precisely for S-curve adoption; use a prior industry's curve (e.g. cloud build-out) as a template for the next one (AI accelerators).

4. Causal / counterfactual reasoning. Model the mechanism — the chain from trigger to revenue — rather than fitting the historical pattern. "If hyperscaler capex rises 40%, accelerator shipments rise X" is a causal claim you can stress test, not a coefficient you extrapolate.

5. Bayesian updating. Under genuine uncertainty, start with an expert prior and update it as the event unfolds. The first two quarters of a breakout are worth more than five years of pre-breakout history; Bayesian updating lets the model learn the shift instead of denying it.

Rule: under a regime shift, point prediction is a category error. Switch to scenarios (bound it), leading indicators (anticipate it), S-curves (shape it), causal models (explain it), or Bayesian updating (learn it). The honest framework picks the right one instead of hiding behind a trend line.

5. Why revenue-model-builder chooses honesty

Most forecasting codebases report a number and a confidence interval and stop. This one treats uncertainty as a first-class citizen, in four concrete ways:

  • ABC data grading — every driver value carries a credibility tag (A = official filing, B = third-party industry data, C = estimate). A reviewer sees at a glance which columns are facts and which are guesses. The NVIDIA demo's Data Center drivers are almost all C-grade — and that is reflected openly in data/sources.md, not hidden.
  • A structural residual line — penetration is never back-solved to force the segments to tie to the total. The gap stays visible, so the model never lies about its own completeness.
  • A certainty pyramid — forecast inputs are ranked by how knowable they are, and the most certain inputs drive the forecast, not the most convenient ones.
  • Back-testing that exposes limits — the library's own hold-out tests publish where it fails (1% on Gaming, 60% on Data Center), instead of only reporting where it succeeds.

The payoff is not pessimism. It is that a user of the model never gets ambushed: they know exactly where to trust it and where to override it.

6. A practitioner's decision guide

Is the industry's growth a continuation of history?
│
├─ YES (mature, trend-driven) ─────► driver tree, point-forecast, back-test to confirm
│                                     (consumer electronics, semis, auto parts, pharma)
│
├─ STRUCTURED but different math ──► driver tree with swapped factors
│                                     (SaaS: MAU×ARPU, Ads: traffic×eCPM)
│
└─ NO (event / cycle / regime) ────► scenarios + leading indicators
                                      (commodities, financials, AI-inflection tech)

Three practical rules of thumb:

  1. Back-test before you trust. If a one-year hold-out gives >25% sMAPE across methods, the industry is telling you it is not trend-extrapolatable. Listen.
  2. Grade your inputs. A model is only as honest as its weakest-tagged column. If everything is C-grade, the wide scenario band is the deliverable, not the point forecast.
  3. Report the boundary, not just the center. A forecast that says "Base $18B, Bull $115B, and here is why they differ" is more useful than one that says "$18B" with a false sense of precision.

The differentiator

The market is saturated with projects that advertise accuracy. Almost none use a real, recognizable company to honestly demonstrate the boundary of their own method — where it is sharp (Gaming, 1%) and where it goes dull (Data Center, 60%), and what to do in the dull region. That honesty is the product. A model that admits "I don't know, but here is the range and the trigger" beats one that prints a confident number and is wrong by 6×.

中文摘要

  • 核心论点：driver tree（自下而上收入分解）的准确性取决于行业增长机制是否可趋势外推，而非公式本身。趋势延续型行业（消费电子/半导体）极准，事件驱动型（AI 爆发/收购/周期）结构性失效——而任何趋势拟合都无法弥补后者，因为未来不在训练数据里。
  • NVDA 实证：同一公司、同一公式、同一年份——Gaming sMAPE 1.0%（趋势准），Data Center sMAPE 60%（AI 爆发崩，真实 $115B vs 预测 $18B）。60 倍差距，变量只有行业。
  • 行业适配性：契合（趋势型）/ 改造（SaaS·MAU×ARPU、广告·流量×eCPM）/ 不契合（周期品、金融、资源、范式跳变科技）。
  • 事件驱动怎么办：点预测注定错，改用 5 招——情景分析（Bear/Base/Bull + 触发条件）、领先指标监测、S 曲线/Bass 扩散、因果推理、贝叶斯更新。
  • 为什么 rmb 选择诚实：ABC 数据分级 + 结构性 residual + certainty pyramid + 回测主动暴露局限——把不确定性当一等公民，让用户永远不被模型"突袭"。
  • 差异化：市面上几乎没有用真实美股案例诚实论证 driver tree 边界的方法论。多数项目只吹"多准"，rmb 敢说"哪里准、哪里崩、崩了怎么办"——这才是有用的。