Investment Scoring
“Which properties offer the best risk-adjusted return?”
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
By submitting, you accept the Terms and Privacy Policy.
Loved by data scientists, ML engineers & CXOs at
A real-world example
Which properties offer the best risk-adjusted return?
Real estate funds evaluate 100+ properties per acquisition, spending $50K-$200K in due diligence per deal. Traditional underwriting models project returns using historical cap rates and comparable sales, missing the demographic and economic graph signals that drive future performance. For a fund deploying $500M annually, improving deal selection by identifying the top-quartile properties saves $20-40M in avoided underperformance over a 5-year hold period.
Quick answer
Investment scoring AI ranks properties by predicted risk-adjusted return using graph-based models that connect property financials to demographic momentum, supply pipeline dynamics, and employment center proximity. Traditional underwriting uses historical cap rates and comparable transactions, missing the forward-looking signals that drive future performance. A fund deploying $500M annually saves $20-40M over a 5-year hold period by scoring deals with graph-based models that identify top-quartile properties before the market prices in their potential.
Approaches compared
4 ways to solve this problem
1. Cap Rate Comparison
Rank properties by going-in cap rate relative to the market average. Higher cap rates signal higher returns (or higher risk). The most basic investment screening tool.
Best for
Initial deal screening to filter out properties that are clearly overpriced relative to their income.
Watch out for
Cap rate is backward-looking: it uses current NOI divided by current price. A property in a growing market with 6% rent growth has a better return outlook than a property in a declining market, even if their current cap rates are identical. Cap rate tells you nothing about where the market is going.
2. DCF / Pro Forma Modeling
Build a discounted cash flow model projecting NOI growth, rent increases, operating expenses, and exit cap rate. The standard underwriting approach for institutional real estate.
Best for
Individual deal evaluation once you have detailed financial data and market assumptions.
Watch out for
Garbage in, garbage out. DCF quality depends entirely on the assumptions: rent growth, expense growth, exit cap rate. These assumptions are usually based on broker opinions and historical averages, not predictive models. Also, each deal is modeled independently, missing portfolio-level correlations and market momentum signals.
3. Regression on Market Features
Train regression models predicting property returns from property features and market-level variables (job growth, population growth, supply pipeline). More data-driven than DCF assumptions.
Best for
Portfolio-level screening across multiple markets where consistent, data-driven deal ranking is needed.
Watch out for
Market features are flat inputs (Austin population growth = 3.2%). Cannot represent the interaction between demographic momentum and supply pipeline (3.2% population growth with high supply pipeline is very different from 3.2% growth with constrained supply). Also misses the timing dynamics: market turning points, absorption rate changes.
4. Graph Neural Networks (Kumo's Approach)
Connect properties, financials, market data, demographics, and transactions into an investment graph. GNNs learn which property-market-demographic combinations produce outsized risk-adjusted returns.
Best for
Funds evaluating large deal pipelines across multiple markets, where the interaction between property financials, demographic momentum, and supply dynamics drives returns.
Watch out for
Requires structured data across property, market, demographic, and transaction sources. The model excels at scoring relative attractiveness across a deal pipeline but still needs human judgment for deal-specific execution risks (title issues, environmental, etc.).
Key metric: Graph-based investment scoring identifies top-quartile deals with 70-75% precision vs 45-50% for traditional screening. Over a 5-year hold, consistently selecting better deals generates $20-40M in incremental value on $500M annual deployment.
Why relational data changes the answer
Real estate returns are determined by the intersection of property-level financials and market-level dynamics. Property INV003 in Nashville projects a 14.8% IRR not because its financials alone are exceptional, but because of the market context: +2.8% population growth creates demand, +3.5% job growth supports rent increases, moderate supply pipeline prevents oversaturation, and comparable transactions at $185K/unit confirm the market is pricing growth. Change any of these connections and the return profile changes dramatically. The same property in a market with declining population and high supply might project 4% IRR.
Traditional underwriting models either ignore these connections (cap rate comparison) or require analysts to manually assess them (DCF assumptions). Graph-based models represent the full property-market-demographic network and learn which combinations predict outsized returns. SAP's SALT benchmark shows 91% accuracy for graph models vs 63% for gradient-boosted trees on relational tasks. RelBench shows 76.71 vs 62.44. For investment scoring, this translates to identifying top-quartile deals with 70-75% precision versus 45-50% for traditional screening. Over a 5-year hold period, consistently selecting better deals generates $20-40M in incremental value on a $500M annual deployment.
Scoring real estate investments from property financials alone is like evaluating a business acquisition using only the income statement. You would miss that the company is in a booming industry, has no competitors, and sits on a transportation corridor that is about to double in capacity. The best investments are defined by their position in the market ecosystem. Graph-based investment scoring evaluates the ecosystem, not just the asset.
How KumoRFM solves this
Graph-powered intelligence for real estate
Kumo connects properties, financials, market data, demographics, and transactions into an investment graph. The GNN learns which property-market-demographic combinations produce outsized returns: how population growth corridors interact with supply pipelines, how employment center proximity affects rent growth, and how comparable transaction patterns signal market turning points. PQL scores each property by predicted risk-adjusted return, enabling funds to focus due diligence on the highest-potential deals.
From data to predictions
See the full pipeline in action
Connect your tables, write a PQL query, and get predictions with built-in explainability — all in minutes, not months.
Your data
The relational tables Kumo learns from
PROPERTIES
| property_id | type | units | year_built | market |
|---|---|---|---|---|
| INV001 | Multifamily | 120 | 2010 | Austin |
| INV002 | Office | 50K sqft | 2005 | Denver |
| INV003 | Multifamily | 200 | 2018 | Nashville |
FINANCIALS
| property_id | noi | cap_rate | occupancy | rent_growth_yoy |
|---|---|---|---|---|
| INV001 | $2.4M | 5.2% | 95% | +4.8% |
| INV002 | $1.8M | 6.8% | 82% | -2.1% |
| INV003 | $3.6M | 4.8% | 97% | +6.2% |
MARKET_DATA
| market | population_growth | job_growth | supply_pipeline | rent_trend |
|---|---|---|---|---|
| Austin | +3.2% | +4.1% | High | Rising |
| Denver | +1.5% | +1.8% | Medium | Flat |
| Nashville | +2.8% | +3.5% | Medium | Rising |
DEMOGRAPHICS
| market | median_age | median_income | renter_pct | tech_employment_pct |
|---|---|---|---|---|
| Austin | 34 | $72,000 | 48% | 18% |
| Denver | 36 | $68,000 | 42% | 12% |
| Nashville | 35 | $58,000 | 45% | 8% |
TRANSACTIONS
| txn_id | market | type | price_per_unit | cap_rate | date |
|---|---|---|---|---|---|
| TXN301 | Austin | Multifamily | $210K | 5.0% | 2025-01-15 |
| TXN302 | Denver | Office | $320/sqft | 7.2% | 2025-02-01 |
| TXN303 | Nashville | Multifamily | $185K | 4.9% | 2025-02-20 |
Write your PQL query
Describe what to predict in 2–3 lines — Kumo handles the rest
PREDICT AVG(FINANCIALS.noi, 0, 365, days) / PROPERTIES.asking_price FOR EACH PROPERTIES.property_id RANK TOP 10
Prediction output
Every entity gets a score, updated continuously
| PROPERTY_ID | MARKET | 5YR_IRR | RISK_SCORE | RANK |
|---|---|---|---|---|
| INV003 | Nashville | 14.8% | Low | 1 |
| INV001 | Austin | 12.2% | Medium | 2 |
| INV002 | Denver | 7.4% | High | 3 |
Understand why
Every prediction includes feature attributions — no black boxes
Property INV003 -- 200-unit Multifamily in Nashville
Predicted: 14.8% projected 5-year IRR (Rank #1, Low risk)
Top contributing features
Rent growth trajectory
+6.2% YoY, accelerating
29% attribution
Occupancy stability
97% (above market)
24% attribution
Population + job growth momentum
+2.8% / +3.5%
20% attribution
Moderate supply pipeline (no oversupply)
Medium
16% attribution
Comparable transactions support valuation
$185K/unit
11% attribution
Feature attributions are computed automatically for every prediction. No separate tooling required. Learn more about Kumo explainability
PQL Documentation
Learn the Predictive Query Language — SQL-like syntax for defining any prediction task in 2–3 lines.
Python SDK
Integrate Kumo predictions into your pipelines. Train, evaluate, and deploy models programmatically.
Explainability Docs
Understand feature attributions, model evaluation metrics, and how to build trust with stakeholders.
Frequently asked questions
Common questions about investment scoring
Can AI really predict which real estate investments will outperform?
AI does not predict absolute returns with certainty. What it does well is rank properties by relative attractiveness based on the interaction of financial, demographic, and market signals. Graph-based models identify top-quartile deals with 70-75% precision, compared to 45-50% for traditional screening. Over a portfolio of 20-40 deals, this selection advantage compounds significantly. The model augments human underwriting judgment rather than replacing it.
What is the most important factor in real estate investment returns?
Market selection drives 60-70% of returns, property-level execution drives 30-40%. Within market selection, the interaction between demand growth (population, jobs) and supply constraints is the strongest predictor. A mediocre property in a great market typically outperforms a great property in a weak market. Graph-based models capture this interaction directly rather than requiring analysts to weight these factors manually.
How does investment scoring AI handle different property types?
The model learns different return drivers for each property type. Multifamily returns are driven by population growth, renter demographics, and employment diversity. Office returns depend on job growth, remote work trends, and submarket vacancy. Industrial returns track e-commerce growth and logistics network proximity. The graph structure handles these differences naturally because each property type connects to different market signals through different relationship paths.
How far ahead can the model predict returns?
Useful return predictions extend 3-7 years, matching typical hold periods. Accuracy is highest for 3-5 year projections and degrades for longer horizons due to increased uncertainty in demographic and market trends. For a 5-year hold, the model's IRR prediction is typically within 2-3 percentage points of actual outcome for top-ranked properties. This is substantially better than DCF models, which rely on assumption-driven projections that are often 4-6 points off.
Can investment scoring AI integrate with existing underwriting workflows?
Yes. The model outputs a risk-adjusted return score and rank for each property in the deal pipeline, along with the top contributing factors (demographic momentum, supply dynamics, financial quality). This slots into the deal screening stage, helping acquisition teams focus due diligence on the highest-scored properties. Most funds integrate the model output as a screen before detailed DCF modeling, reducing the number of deals that require full underwriting by 40-60%.
Bottom line: A fund deploying $500M annually saves $20-40M over a 5-year hold by improving deal selection. Kumo's investment graph connects property financials to demographic momentum and supply dynamics, scoring deals by predicted risk-adjusted return rather than historical cap rates alone.
Related use cases
Explore more real estate use cases
Topics covered
One Platform. One Model. Infinite Predictions.
KumoRFM
Relational Foundation Model
Turn structured relational data into predictions in seconds. KumoRFM delivers zero-shot predictions that rival months of traditional data science. No training, feature engineering, or infrastructure required. Just connect your data and start predicting.
For critical use cases, fine-tune KumoRFM on your data using the Kumo platform and Research Agent for 30%+ higher accuracy than traditional models.
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
