Methodology

How the engine computes, what it approximates, and what we refuse to predict. This page is the long-form version of the anti-promise discipline framing on the landing page and complements Scope (what we model) and Security (how we handle your data).

Principles

Four principles that decide every modeling choice in the engine.

1. Deterministic math, not probabilistic forecasts

The proforma engine is a deterministic calculation layer. Given the same inputs, it returns the same outputs to the cent. We do not run Monte Carlo simulations, we do not output probability distributions, and we do not produce “AI-generated forecasts” that vary between runs.

Why this matters: an IC committee can audit a deterministic model. Two analysts can re-run it and converge. The Excel export round-trips to the same numbers your IC sees on screen. Probabilistic outputs sound sophisticated but break the chain of audit a sponsor needs to defend a recommendation.

2. Compute what you assumed, don’t assume on your behalf

The wizard pre-fills with tier-appropriate defaults — Smith Travel ADRs by chain scale, USALI expense ratios by service tier, franchise fee schedules by brand. Every default is visible, editable, and labeled. None of them silently override an input you typed.

When you save underwriting defaults in Settings, the wizard pre-fills the New Deal flow with your firm’s assumptions. Same principle: everything is visible, banner-counted, and per-deal editable without touching the saved default.

The engine never invents inputs. If you leave a field blank, the engine either uses the documented default or fails loud — never silently zero.

3. Anti-promise on the variables that matter

Some assumptions are too consequential and too uncertain to predict. We say so, in the wizard, in the AI Brief, and here.

Sell-out velocity on branded residences. Three preset curves (aggressive / baseline / conservative). All three are reasonable in some markets and wrong in others. We compute the economics that follow from the curve you pick. We do not predict which is right for your deal.

Exit cap rate.Drives terminal value, drives levered IRR, drives everything. The default reflects the comparable-cap snapshot at the property type + market level. The wizard surfaces it, the sensitivity matrix lets you stress it, the AI Brief calls it out as the most important assumption to defend at IC. We don’t know what your buyer will pay in year 5.

Stabilization curves.ADR ramp, occupancy ramp, F&B revenue ramp. We provide tier-aware ramps; you override per-deal. The model doesn’t predict the leasing market.

Construction cost overrun.Not modeled. Build a contingency line into your sources & uses; the engine computes around it.

4. Hospitality-native, hospitality-only

Underwrote.AI models hospitality. Hotels — select-service, full-service, extended-stay, lifestyle, luxury. With first-class support for the OOD lines and mixed-use components that show up in luxury and resort deals (spa, golf, retail, telecom, internet, membership clubs, branded residences).

We do not model multifamily, office, retail-as-primary, industrial, or specialty CRE. We chose to do hospitality completely instead of six asset classes shallowly. The underwriting conventions — USALI categories, Smith Travel benchmarks, sub-brand franchise fees, ADR ramping mechanics — are well-defined enough that one tool can do them right.

If you underwrite hotels AND other asset classes, Underwrote.AI is the right tool for the hotel side. Use whatever you already use for the rest.

How the engine computes

The proforma is a hybrid

The engine ships with a single underwriting framework derived from a real institutional deal (Alley North Hotel — a select-service ground-up Sun Belt project). Two layers compose to produce the proforma:

1. An Excel template parsed and evaluated per-request via hyperformula. The same formulas an analyst would write in Excel run on the server, deterministically.
2. A JS calculation layerthat fills the gaps the template doesn’t express — luxury OOD, club revenue, residential sell-out, sensitivity matrices, scenario overlays, waterfall variants — and overlays sponsor-specific overrides.

The Excel export gives you live-formula round-trip. Open the workbook, edit a cell, watch every dependent metric recalculate. Hand it to a lender; they can stress your assumptions inside Excel without coming back to us.

Output coverage

• 10-year operating proformawith departmental P&L per USALI (rooms, F&B, OOD, undistributed, fixed charges, reserves).
• Capital structure— senior debt + optional mid-hold refinance, mezzanine (closed-form annuity P&I + exit fee), preferred equity (current / accrued / PIK), OZ equity with per-rank pref allocation, sponsor co-invest with promote.
• Five partnership waterfall structures— Simple Split, Pref + Split, Pref + ROC + Split, IRR Lookback, Catch-Up. Engine-vs-Excel parity within $1/year on per-bucket distributions. Three-bucket (LP + OZ + GP) supported.
• Returns metrics— levered IRR, equity multiple, DSCR by year, debt yield, exit value at chosen cap rate.
• Sensitivity matrices— IRR and DSCR over a 5×5 grid (exit cap vs. stabilized ADR by default; configurable).
• Scenario overlays— Base / Upside / Downside toggled in one click; the engine recomputes the full proforma per scenario.

Excel export — what ships

Eleven visible sheets:

1. Cover — branded with your workspace logo, EIN, IC signature.
2. Summary — headline metrics, returns waterfall summary.
3. Inputs — every wizard input, audit-traceable.
4. Operating — 10-year P&L, USALI-aligned.
5. Sources & Uses — capital stack, draw schedule.
6. Debt — senior + mezz + pref equity schedules, refinance event.
7. Returns & Exit — cashflow stream, IRR build, exit value.
8. Waterfall — live partnership distribution math, structure-specific.
9. Sensitivity — IRR + DSCR 5×5 matrices, formula-driven.
10. Scenarios — Base / Upside / Downside side-by-side.
11. Underwrote_AI_Brief— LLM-ready summary + three paste-into-Claude prompts (IC memo / risk / diligence). The brief is your data, prepared for your AI.

Every cell is a live formula. None of the sheets are flattened to values.

What we approximate

The engine ships with two material approximations. Both are documented; both can be overridden.

F&B operating cost uses a 3× food-cost multiplier

F&B operating cost = F&B revenue × food cost % × 3. The 3× multiplier is industry shorthand for COGS + labor + other operating expense. For deals with material F&B revenue or atypical F&B economics (a restaurant-driven boutique hotel, a resort with multiple F&B outlets), override the F&B expense line directly. A granular F&B department schedule is a future engine refactor.

Property tax + insurance grow at a flat rate

Both lines escalate at a fixed annual rate (default 3%, user-overridable). The engine does not model jurisdiction-specific reassessment rules. This matters for:

• Texas— annual reassessment, no statutory cap on commercial.
• California— Prop 13 caps annual increase at 2% until ownership change triggers a full market-rate reset (consequential at acquisition).
• New York— market-driven with assessment lag.

Override the property tax line per deal to reflect your jurisdiction. Jurisdiction-aware modeling is on the roadmap when buyer signal justifies the complexity.

What we don’t model

Honest accounting of what the engine doesn’t do today. The full list lives on the Scope page; the headline gaps:

• Multi-template engine— one framework today (Alley North-derived). Generalizes well across hotel deal shapes; power users with strongly opinionated workflows may want overrides.
• Second refinance event— engine models one refinance event per deal. Architecture supports N events; we ship it when a real deal needs it.
• C-PACE / NMTC / HTC layers— wizard accepts the layers; engine doesn’t yet thread credits or PACE assessments through the cashflow stream. Real for adaptive-reuse and OZ deals; held for a future phase.
• Asset management variance tracking— Underwrote.AI is an underwriting tool, not an AM tool. Quarterly reforecasts and budget-vs-actual reporting are intentionally out of scope.
• Sub-1% exit cap rates— the sensitivity matrix clamps at a 0.5% floor to avoid degenerate terminal values. The proforma itself accepts any cap rate; only the stress-testing surface is clamped.

How we validated the math

Three discipline mechanisms:

Engine-vs-Excel parity tests. Every waterfall structure is tested against a hand-built Excel reference at the $1-per-year level per distribution bucket. Tests gate the release.

Real-deal calibration. The default template + numerics are derived from a real institutional hospitality deal, not a textbook example. Headline metrics on a calibrated deal land within institutional review tolerance on first wizard pass.

Audit-every-cell discipline in the export. Every output cell in the Excel workbook traces to either a wizard input or a formula chain you can follow. No magic constants. No pasted values.

When we identify a known modeling decision (e.g., the adverse-cash corner case in the 5-structure waterfall where exit proceeds are insufficient to satisfy ROC + accrued pref), we document it on the Scope page rather than ship a speculative fix. The right behavior depends on what real deals need.

When we ship engine changes

We ship engine changes when:

1. A real deal calibration surfaces a gap.
2. A modeling decision in the codebase is provably wrong (Excel parity test failures, audit findings).
3. A customer with a specific shipped deal asks for a feature with a concrete worked example.

We do not ship features speculatively. We do not pre-announce capabilities.

Questions or pushback

If a number looks wrong, an output behaves unexpectedly, or you have a modeling convention you’d like the engine to support — email support@underwrote.ai with the deal ID (visible in the URL) and a brief description. We respond within one business day.

The fastest way to ship a feature is to have a real deal we can calibrate against.