Methodology

How a deal actually gets done.

The whole process in order, from finding out who buys in a market to the offer that goes to a seller. The thresholds are the real ones. Where a step exists because something went wrong in testing, the failure is named.

Counterparties first

Before any property is looked at, the platform builds a list of who is actually buying investment residential in a market. It reads recent purchases from the county deed record and a national property feed, then scores each counterparty against six signals: entity ownership, repeat buying, cash close, absentee mailing address, a purchase below the property's own estimated value, and buying across more than one market. Three of six is the bar for inclusion.

Any single signal has an innocent explanation. A retiree paying cash for a downsizer trips the cash signal alone. Anyone who has not updated their mailing address is absentee. Three at once is not a household, it is an investor.

An owner-occupancy flag is a hard veto regardless of score, because someone living in the house is not acquiring another one. Portfolio size then sets rank. Two to thirty properties earns a bonus, since that is a person who answers their own phone and can decide. Above fifty the score is penalised, and at five hundred it is penalised by more than the largest single signal is worth, so an institution can never outrank a local operator no matter how clean its signals look.

Screening out who cannot be sold to

Institutions register one entity per tranche, so a fund often appears in the data as a small holder. The platform groups buyers by brand stem and propagates the largest reported size across the family. It also flags serialised single-purpose naming, address-derived junk names, and buyers whose every observed purchase sits outside the target markets.

One entity reported sixteen properties and zero purchases in twelve months. Its sibling, one letter different in the name, reported 9,735 properties and 2,462 purchases. The sixteen put a fund slice into the band that earns a rank bonus.

Nothing is deleted. Every buyer keeps a verdict and the evidence behind it, and the sellable flag is what narrows the call list. Deduplication is deliberately conservative: the vendor supplies a surname rather than a full name, so nine separate Smith entries are nine different people in nine zip codes. Merging them would invent one buyer with a fabricated portfolio, so only entity names are merged and bare surnames are reported as unresolvable.

The acquisition mandate

Each counterparty's mandate is reconstructed from purchases they have actually made, never from what they say. Price band comes from a percentile range across their observed purchases. Beds, baths, square footage, year built, unit count, zip codes and cities come from the same set. Asset class is derived from observed unit counts: SFR, 2-4 unit, both, or unknown.

A mandate built from two coincidental purchases looks identical to one built from twelve. Acting on the first is how counterparty-first sourcing gets seeded with noise.

So every mandate carries its evidence. A mandate is only usable for search when it has a price band and enough distinct transactions behind it, and same-price same-date purchases collapse to one transaction so a bulk deed transfer cannot inflate the count. Everything below that bar is marked thin and cannot drive a search. Absent unit data is recorded as unknown rather than as one, because defaulting it to single-family would claim a buyer does not do duplexes when nothing supports that.

Sourcing against real demand

With mandates in hand, sourcing runs two ways. Counterparty-first takes one counterparty's revealed mandate and searches its band and zip codes directly. Generic runs the operator band of $40,000 to $400,000 against distress and equity filters: absentee ownership, high equity, tax liens, pre-foreclosure, vacancy, and days on market.

The ordinary sequence is to find a property and then hunt for someone to take it. Starting from a counterparty who exists removes the hunt.

Distress flags combine as a union rather than an intersection. A property that is vacant or tax-delinquent or pre-foreclosure is a lead; one that is all three at once is a rounding error. Combining them as an intersection returned zero results out of 43,744 surveyed properties. SFR and 2-4 unit are searched separately and never blended, because one price-per-square-foot median built from a duplex and a house describes neither, and of 2,448 observed counterparties only 27 buy both. Five units and above is refused outright. Every pull is sorted explicitly and persisted with its candidate IDs, so the same filters return the same shortlist and a deal seen last week can be reopened rather than searched again.

Comparable sales

Comps come from recorded arm's-length sales near the subject, within a square-footage band and a recency window, reduced to a median price per square foot and applied to the subject's own area. Two sources feed it: a county deed register where one is available, and a national comparable sales endpoint everywhere else.

Deed data is stronger evidence, because it carries transaction type and party names, so warranty deeds can be separated from quitclaims and trustee transfers at the source. It covers one county. The national source covers everywhere and carries neither, so arm's-length has to be inferred from price.

The output is an AS-IS value, what the property is worth today in the condition it is in. Screens run in order: a nominal-transfer floor, an absolute price-per-square-foot band, a similarity test against the subject's own size, then a relative screen against the comp set's own median. The relative screen matters most. A $600 per square foot sale on a $20 street is an outlier; on a $220 street it is ordinary, and no fixed band can express that. Inherited fixed floors were rejecting real Detroit sales at $10,100 and $25,500 and reporting no usable sales nearby, on a street with four recorded sales on it.

ARV and the ceiling

After-repair value starts from the as-is comp value and adds an uplift proportional to the repair scope. That figure is then capped by a ceiling taken from the upper quartile of the same comp set, applied to the subject's square footage.

You cannot renovate above the best house on the street. The ceiling is measured from the comps rather than assumed, and when it binds the model says so, because a binding ceiling is the most common reason a rehab loses money.

A second cap holds ARV to twice the as-is value. When that one binds it is a sanity cap rather than a valuation, and it is labelled as such, since a renovation that more than doubles value usually means bad subject square footage or a bad comp set upstream. Where no comp ceiling can be computed, the ARV is shown as rough and not offer-grade rather than presented as a valuation. Blind backtesting, predicting withheld sale prices with the subject excluded from its own comp set, put median absolute error near 13% on the comp-derived value. That is wider than a typical acquisition spread, so it is triage-grade and the model says so.

The tier gate

Before anything is priced, the gate decides whether the automated value can be trusted at all. Above $200,000 it passes. Between $40,000 and $200,000 it passes but is marked low confidence, condition unknown, and not offer-grade. Below $40,000 the value is withheld and the deal routes to a human.

No data source here records property condition, and below roughly $150,000 the spread between a gutted shell and a renovated house on the same street exceeds the entire price. A median price per square foot there summarises nothing, whichever vendor supplies it.

The trust floor and the search floor are separate constants, and keeping them separate is the point. They were once the same value, so widening the search band to $40,000 silently promoted every cheap-band value to trustworthy. Whether a value is offer-grade is a fact about condition variance in the data. What to search for is an operator decision. Tying them together made one look like the other.

Repair scope

A tiered cost per square foot, light, medium or heavy, adjusted by a state cost factor and carrying a 10% contingency. On the worked example that is $42 per square foot against a Tennessee factor, producing an $80,078 budget on 1,884 square feet.

Rehab cost tracks square footage and scope, not sale price. The consequence is worth stating plainly: on a cheap house the repair budget can approach the as-is value, so a cheap deal is almost entirely a condition bet, and a wrong scope guess is fatal there in a way it is not at $300,000.

Where interior photographs exist they can raise a repair budget but never cut one. A staged or AI-generated image almost never makes a house look worse, so the failure mode runs one way and the model only lets that evidence move in the direction it cannot be gamed.

Underwriting a single-family deal

With ARV and repairs set, the model computes NOI from market rent less operating expenses, then a debt schedule at a stated LTV and rate, then DSCR and going-in cap rate. On a hold it runs a ten-year DCF, levered and unlevered, with an exit priced off forward NOI.

An acquisition contract still has to make sense to whoever ends up holding it. Pricing without knowing what the property does as a rental means pricing without knowing what a landlord can pay for it.

Every figure that is an estimate is labelled as one with its basis inline. On a vacant property the trailing-twelve statement titles itself a pro forma and carries a do-not-present-as-actuals stamp, because that is what it is.

Underwriting 2-4 unit

Duplexes through fourplexes run the same path with a rent roll instead of one rent. Each unit's rent is entered separately and summed to gross potential rent, then EGI after vacancy and credit loss, then NOI after operating expenses and reserves. Cap rate, DSCR and the DCF follow from there.

Asking for one blended rent is how a fourplex gets underwritten on a single unit. Entering four rents and summing them is the difference between a right deal and a wrong deal that looks like a right one.

Four units is a hard ceiling. At five the property is valued on capitalised income rather than comparable sales, financed on commercial terms, and bought by syndicators rather than by small landlords, so the comps engine, the flip screen and the DSCR screen all stop applying. The platform refuses a 5+ unit property at intake rather than pricing it with machinery that does not fit.

The institutional multifamily model

For larger multifamily the same engine carries unit mix, itemised operating expenses, economic vacancy alongside physical, a debt schedule with a coverage test and debt yield, a ten-year DCF with levered and unlevered IRR, an exit priced off forward NOI, and a four-tier LP/GP waterfall through preferred return, return of capital, GP catch-up and promote.

Economic vacancy is the number that gets skipped. On the worked example it is 15.68% against 6.00% physical, and the gap is concessions, credit loss and non-revenue units. Underwriting to physical vacancy alone overstates EGI before a single expense line is touched.

Debt yield sits beside DSCR because they fail differently. DSCR can be held up by a low rate on a property that does not produce. Debt yield, NOI over loan amount, does not care what the rate is. The workbook is downloadable and every output reconciles to the underlying engine to the cent.

Counterparty-priced MAO

The maximum allowable offer is the lower of two numbers: what the margin tier would allow, and what the best real counterparty can actually pay. The counterparty side is itself the higher of two screens, a flipper at 70% of ARV less repairs, and a financed landlord at the price where NOI covers debt 1.20 times. The disposition spread comes off that to give the offer to the seller.

A margin target says what you would like to make. It says nothing about whether anyone will buy the contract. Pricing to the tier alone produces contracts nobody wants.

On the worked example the tier would have allowed $226,996 and no real counterparty reaches it. The flipper screen sets $188,611, the rental screen sets $189,596, and the rental screen binds, so MAO is $189,596 and the offer to the seller is $179,596 after a $10,000 spread. Which screen binds also inverts with price: DSCR improves as price falls, so cheap deals are priced by the landlord and mid-band deals by the flipper. That changes who to call first.

Matching and disposition

A priced deal is matched back against the stored acquisition mandates on price, beds, square footage, asset class and area. Asset-class mismatch is disqualifying. An unseen zip code is reported as untested rather than as a refusal. The counterparty-facing package is then computed on the counterparty's own basis, and outreach is drafted for a human to send.

Offering a fourplex to a counterparty whose every observed purchase is single-family wastes the first call, and the first call is the one that matters. A zip code we have not seen them buy in is missing evidence, not evidence of a boundary.

The package is generated through a guard that refuses to emit cost basis, MAO or margin, so what goes to a counterparty is computed on their numbers rather than on mine. Nothing sends automatically. Every draft goes to a review queue and a human sends it.

Limits

What the model will not do

It does not substitute a weaker number for a missing one. A vendor returning no data produces an explicit negative rather than a fallback that reads like a result.

It does not withhold a decision because an optional input is absent. A missing confirmation degrades confidence. It never blocks output.

It labels every figure as measured, vendor-supplied or estimated, with the basis stated inline.

It will not treat a photograph it suspects is AI-generated or virtually staged as evidence of good condition. A fake image almost never makes a house look worse, so images may raise a repair budget but never cut one.

And it cannot tell you who ultimately takes title. No deed or assessor record shows that, because the deed usually runs from the original seller straight to the end buyer and the intermediary leaves no trace. A high score means a counterparty sits firmly in the population active acquirers come from. It is call ordering, not qualification. Only the call settles it.

The two worked examples