revenue-model-builder

Forecast revenue the way sell-side analysts do — as a driver tree the engine can defend, grade, and stress-test.

中文文档：README-zh.md

Zero dependencies. Pure stdlib Monte Carlo. Every number carries a source and a credibility grade.

segment_revenue = market_base × penetration × share × price total_revenue = Σ(segments) + residual

Pre-registered, out-of-sample validated

The v0.16 industry-fit claim was tested on 244 S&P 500 constituents (anti-survivorship anchor 2023-12-31) + six hand-built driver trees, per a spec frozen before any test data was pulled:

• Driver layer: industry-default forecasts beat naive per-driver trending on 3 of 4 testable trees (SBUX 0.7% vs 3.6%, META 4.6% vs 6.7%, NVDA Gaming 3.0% vs 10.9% sMAPE)
• Weak-fit redirect works: warnings fired 2/2 with 0/4 false alarms, and Monte Carlo P10–P90 bands framed all three weak-fit test years (JPM NII landed at P46; NVDA Data Center at P75/P66)
• Honest nulls published: growth-revert falsified at home; totals layer belongs to statistical baselines

Full scorecard: docs/profile-validation.md

Why this exists

Open-source finance tooling covers trading and backtesting (zipline, QuantLib) and DCF valuation — but driver-based revenue forecasting, the base × penetration × share × price decomposition that sell-side analysts and PE associates actually build in Excel, has no runnable engine. Prompt skills describe the method; this library is the method — the math enforced in code, not left to a spreadsheet comment.

A revenue model lives or dies on whether you can defend every number. Here every driver carries a credibility grade (A/B/C) and a source, the residual is a first-class line (not a fudge), and checks catch the classic traps — back-solved penetration, residual dominance — before they poison the forecast.

One image: where driver trees work — and where they break

Same company, same formula, same engine — Gaming (mature market) tracks at 1.0% sMAPE while Data Center (AI regime shift) misses 6× and the engine flags it before the hold-out opens, redirecting to Monte Carlo scenarios whose Bull tail frames the actual $115B. Accuracy is a property of the industry's growth mechanism — v0.16 encodes that as 10 mechanism profiles with strong/adapt/weak fit classes (validated out-of-sample).

60-second start

pip install revenue-model-builder

from revenue_model import (
    Driver, Segment, RevenueModel, BASE, PENETRATION, SHARE, PRICE,
    forecast_segment, segment_warnings, simulate_segment,
)

seg = Segment(
    "cockpit-domestic",
    base=Driver("China passenger car sales", BASE, {2022: 22.0, 2023: 23.0},
                level="A", unit="M units", source="CAAM"),
    penetration=Driver("DMS penetration", PENETRATION, {2022: 0.04, 2023: 0.06},
                       level="B", unit="fraction", source="research institute"),
    share=Driver("market share", SHARE, {2022: 0.10, 2023: 0.12},
                 level="C", unit="fraction", source="estimate"),
    price=Driver("ASP", PRICE, {2022: 600, 2023: 620},
                 level="C", unit="yuan", source="benchmark"),
    industry="consumer_electronics",   # <- the one tag that changes everything
)
model = RevenueModel("DemoCo", [seg], total_revenue={2022: 78.0, 2023: 163.0})

print(model.validate_all())               # Σ segments + residual == reported
fc = forecast_segment(seg, [2024, 2025])  # industry-default extrapolations
for w in segment_warnings(fc):            # fit verdict, before you forecast
    print(w)
mc = simulate_segment(fc, 2024, {"market share": (0.10, 0.18)}, n=20000)
print(mc.median, mc.percentiles["p5"], mc.percentiles["p95"])

Or watch the GIF above. CLI: python -m revenue_model {build, simulate, excel, docx, extract, sec, akshare, tushare}.

What's inside

Deep dives (each links to docs + runnable example)

• Industry fit — the matrix, NVDA/Luxun natural experiments → docs · examples/industry_demo

• Profile validation — the pre-registered 244-company test → docs · examples/profile_validation

• News-impact validation — do 8-K events predict revenue? (spoiler: no, and that's the finding) → docs

• Backtest — adaptive methods vs driver structure on real data → examples/backtest_demo

• Real-data demos — NVDA (SEC, quarterly granularity), Luxun/Desay SV (A-share 20-year), zhipu ARR ladder →

  examples/

Install

Core is dependency-free; extras opt in:

pip install revenue-model-builder                 # core, zero deps
pip install revenue-model-builder[excel]          # openpyxl
pip install revenue-model-builder[docx]           # python-docx + matplotlib
pip install revenue-model-builder[backtest,data]  # statsmodels + akshare

Python 3.9–3.13 · MIT license · Documentation · Changelog

Design principles

Five hard-won rules, enforced structurally: structural residual · A/B/C data grading · incremental (not growth-rate) penetration · certainty pyramid · history-first workflow → docs/design-principles.md

Research/education tool, not investment advice — see DISCLAIMER.

Roadmap

• v0.18 (data anchor): Damodaran industry benchmarks wired into profile checks — thresholds become citations
• v0.19 (experience): profile auto-recommendation from backtest fingerprints; profile catalog page

Shipped — v0.17 (math kernel): churn-survival dynamics for subscription bases (base × (1+gross) − base × churn as the saas/telecom default, with a net-churn check that no ARPU escalator offsets a shrinking base); DampedTrend + DeceleratingCAGR graduated from the pre-registered validation into the standard backtest battery (7 methods); hand-coverage fix in forecast_segment; G1 code gate in CI (ruff, pinned ruleset + mypy on the kernel); FIG teaching-consensus citations for the weak-fit financials class.

Full history: CHANGELOG.md · Releases

Who is this for

Sell-side/PE associates who want their revenue model to be code; quants researching forecastability by industry mechanism; students learning driver-based valuation; FP&A teams escaping spreadsheet sprawl. Contributions welcome — the docs double as the design record.

from revenue_model import (
    Driver, Segment, RevenueModel,            # core tree
    BASE, PENETRATION, SHARE, PRICE,          # driver kinds
    implied_driver,                           # calibrate one driver to a known
                                              # revenue (prefer PRICE/BASE over
                                              # PENETRATION: avoids back-solve)
    forecast_segment, segment_warnings,       # industry defaults + checks
    check_segment, list_profiles, resolve_industry,
    simulate_model, simulate_segment, scenarios,   # Monte Carlo
    tornado,                                  # per-driver sensitivity ranking
)
# Driver extrapolations: trend (fit_trend), mean_reversion, erosion, growth,
# hold, logistic, incremental — and net_growth (v0.17): subscriber bases as
# base x (1+gross) - base x churn, the two knobs kept separate.

# Driver(name, kind, {year: value}, level="A"|"B"|"C", unit=..., source=...)
# Segment(name, base=..., penetration=..., share=..., price=...,
#         reported_revenue={...}, industry="saas_subscription" | "40" | "银行")
# Segment.revenue(year) -> float
# RevenueModel.validate_all() -> [YearResult(segment_sum, residual, warnings)]
# CLI: python -m revenue_model {build, simulate, excel, docx, extract,
#                               sec, akshare, tushare}