Lessons From Analyzing 100 Multifamily Listings

We ran 100 broker-listed multifamily deals through the same structured underwriting process between January and April 2026. The deals ranged from $4M to $48M, spanned 18 metros, and came from a mix of national brokerage firms, regional shops, and direct-from-owner posts. What follows are the patterns that emerged — not about which markets are hot, but about how deals get presented and what that presentation tells you.

Headline finding: the broker pro forma is wrong, predictably

Of the 100 deals, 89 had a broker pro forma NOI at least 8% above the trailing-twelve-month actual. The median gap was 14%. The top decile was 28%.

The mechanism is consistent. Brokers don't fabricate numbers — they make assumptions:

• Rent-to-market. Every unit gets bumped to "market" rent, even units 6 months into a lease.
• Vacancy. Pro forma vacancy is set to "stabilized" — typically 5% — regardless of current trailing vacancy.
• Management fee. Stripped out (you're a sophisticated buyer, you'll self-manage).
• Replacement reserves. Set to zero or $150/unit/year — a fraction of what an institutional underwrite would carry.
• Payroll allocation. For portfolios, allocated based on unit count, not actual time on this asset.

None of this is dishonest. The OM usually says "buyer to confirm." But the cap rate that flashes on the cover ("6.2% cap on pro forma") is meaningfully different from the cap rate on TTM actuals (often 4.8–5.1% on the same property).

Practical takeaway: automated extraction should always pull both the broker pro forma and the TTM actuals, then compute and display both cap rates side by side. The gap between them is itself the headline metric.

Where the deals actually died

Of the 100 deals we put through full underwriting, we pursued 6. The 94 that died broke down roughly like this:

• 38 deals — pricing too high relative to TTM cap rate. The broker number was 6%; the real number was 4.7%; we underwrote 5.4%. Bid would have been 15%+ below ask.
• 22 deals — supply pipeline in the submarket was deep. Even if the property was priced reasonably, projected rent growth couldn't survive 18 months of competitive lease-up.
• 14 deals — capex requirements were materially under-disclosed. A 1985-built garden-style with original windows, original HVAC, and a 35-year roof needs $9–14K per door to compete with new product. The OM mentioned "value-add opportunity" without quantifying it.
• 9 deals — submarket on the wrong side of a demographic trendline. Households in the 22–44 cohort declining, not growing.
• 8 deals — financing math broken at current rates. DSCR < 1.20x at quoted terms.
• 3 deals — title, environmental, or zoning issues that surfaced in early diligence.

The first bucket — pricing — is the biggest, but it's actually the easiest to disposition. You see the gap quickly, you decide whether to lob a low bid, you move on. The harder buckets are 2 and 3: supply and capex. These require local knowledge or a thorough physical that you can't do remotely. They're where deals look fine on paper and break in diligence.

What good deals looked like

The six deals we pursued had four things in common. Not all four every time, but at least three.

1. TTM cap rate within 50 basis points of broker pro forma. The broker had been disciplined about assumptions. This is partly about the broker and partly about the seller; institutional sellers with sophisticated reporting tend to produce more honest OMs.
2. **Rent roll showed concrete loss-to-lease that was executable.** Not just "rents are below market" — but a stack of leases expiring in the next 12 months with documented submarket comps showing the spread.
3. Submarket supply 1.5% of stock or less, declining. Either the cycle had passed or the submarket was too small to attract merchant developers.
4. Capex disclosed and reasonable for vintage. Either recent capex done by seller (and verifiable) or a clear $X/door budget that fit our model.

The interesting thing: none of the six deals we pursued were the lowest-priced. The lowest-priced deals were almost always the ones with the worst pro forma honesty or the deepest hidden capex. Cheap was cheap for a reason 70% of the time.

Distribution of asking caps vs. our cap

We ran the broker asking cap against our underwritten cap for all 100 deals. The distribution:

• Broker asking cap median: 5.8%
• Broker asking cap range: 4.5% to 7.4%
• Our underwritten cap median: 5.0%
• Our underwritten cap range: 3.6% to 6.7%

The 80-basis-point median gap is the cost of the assumption stack described above. The narrowing of the range — broker max was 7.4%, ours was 6.7% — comes from us refusing to underwrite the "pro forma upside" in distressed assets that justified the high broker cap.

The seven deals where our cap exceeded the broker cap were all situations where the seller had under-utilized rent push (often a long-time owner with stable tenant base) and our market research suggested executable upside. Five of those seven were in our pursued bucket.

Things AI did well, things it didn't

Since this exercise ran on our multi-agent underwriting pipeline, we kept notes on where the AI agents added value and where they didn't.

Did well:

• Extracting structured data from OMs (purchase price, NOI, unit mix, year built, occupancy) — ~95% accurate on first pass when prompts forced null on missing fields.
• Flagging mismatches between OM text and rent roll numbers. (The OM said 92% occupancy; the rent roll showed 87%.)
• Drafting the first-pass risk section of the IC memo. Decent baseline, easy to edit.
• Producing market summaries for known metros (Tampa, Nashville, Phoenix, Atlanta).

Didn't do well:

• Estimating capex for specific properties. The model would generate plausible per-door numbers, but they weren't tied to actual condition. Human inspection still drove this.
• Submarket-level supply analysis. Public data is too lagged; the model didn't know about projects breaking ground in 2025–2026.
• Recognizing when a "value-add" listing was actually a distressed property dressed in marketing language. The model took the OM language at face value too often.
• Sponsor-fit analysis. AI can't tell you whether the deal matches your capital stack, your operating partner, your IRR threshold.

What changed for us

Three things changed in our process based on this batch:

We now extract broker assumptions as a separate output. Not just "what does the broker say NOI is" but "what are the assumptions behind that NOI" — pro forma vacancy, market rent for each unit type, expense ratio, management fee inclusion, reserve assumption. This list goes into the IC memo as a "Broker Assumption Stack" section. It's the most-referenced section in committee discussion.

We require a TTM-actual underwrite alongside the pro forma underwrite. Two columns, side by side. We don't accept a deal that's only attractive on the pro forma.

We score every deal on "presentation discipline." A separate 1-5 score on how cleanly the broker presented the financials. Low presentation discipline doesn't kill a deal, but it raises the diligence budget required to confirm what's real.

Why this exercise matters

The point of putting 100 deals through a structured process isn't to find six deals to pursue. It's to calibrate. You learn what a healthy deal looks like by looking at a hundred unhealthy ones. You learn that "8% rent growth pro forma" is a flag, not a thesis. You learn which submarkets consistently produce honest OMs and which produce inflation. You learn which brokers earn the benefit of the doubt.

That calibration is what speeds up deals 101 through 200. The first hundred is the cost. The next thousand is the return.