Vacancy Duration Prediction
“How long will this unit sit vacant?”
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
How long will this unit sit vacant?
Every vacant day costs 3.3% of monthly rent. Property managers price units based on gut feel and stale comps, leading to either pricing too high (extended vacancy costing $1K-$3K per unit) or pricing too low (leaving $500-$2K per year on the table). For a REIT with 20,000 units and 35% annual turnover, optimizing listing price and timing based on predicted vacancy duration saves $8-15M annually.
Quick answer
Vacancy duration prediction AI forecasts how many days a rental unit will sit vacant at different price points, enabling revenue-maximizing pricing decisions. Graph-based models connect unit attributes, comparable listings, market conditions, and seasonal demand patterns to predict days-on-market per unit. This moves pricing from gut feel and stale comps to data-driven optimization. A REIT with 20,000 units saves $8-15M annually by pricing units at the revenue-optimal point that balances rent level against vacancy duration.
Approaches compared
4 ways to solve this problem
1. Comp-Based Pricing
Price units based on comparable listings in the area. The standard approach used by most property managers and leasing agents.
Best for
Markets with many similar listings where comps are readily available and the unit is not unusual.
Watch out for
Comps tell you what others are asking, not what will actually lease quickly. Asking prices are often aspirational. Also cannot differentiate between a renovated 2BR on the 3rd floor and an unrenovated 2BR on the 1st floor in the same building. These details drive 20-30% rent differences that comps cannot capture.
2. Revenue Management Rules
Apply pricing rules based on occupancy and time: if occupancy is below 92%, drop prices 3%. If a unit has been vacant 30+ days, drop 5%. Simple automation of manager intuition.
Best for
Large portfolios where some systematic pricing discipline is better than pure ad hoc decisions.
Watch out for
Reactive, not predictive. By the time a unit has been vacant 30 days, you have already lost $1,500+ in revenue. Rules also treat all units the same: a ground-floor unrenovated unit and a penthouse with a view get the same 3% adjustment. One-size-fits-all rules leave money on the table.
3. Regression on Unit Features
Train a regression model predicting rent and days-on-market from unit attributes (sqft, floor, renovated, photos) and market features (vacancy rate, seasonal index).
Best for
Mid-size portfolios with enough historical leasing data to train robust models per market.
Watch out for
Treats each unit independently. Cannot capture that listing UNIT01 at $1,950 affects the demand for UNIT02 at $1,350 in the same building (different unit types compete for overlapping tenant segments). Also misses the dynamic nature of the leasing market: yesterday's optimal price changes when a competitor drops their rates.
4. Graph Neural Networks (Kumo's Approach)
Connect units, listings, applications, market data, and seasonal trends into a leasing graph. GNNs predict vacancy duration per unit at different price points, finding the revenue-maximizing rent for each specific unit.
Best for
Large portfolios with diverse unit types where pricing interactions between units, competitive listings, and seasonal dynamics drive leasing velocity.
Watch out for
Requires historical leasing data with actual days-on-market and listing details (photos, descriptions, pricing changes). The model excels when you have 1,000+ leasing transactions per market to learn from.
Key metric: Graph-based vacancy prediction estimates days-on-market within 3-5 days accuracy vs 10-15 for flat models. Revenue-optimal pricing saves $8-15M annually for a 20,000-unit REIT by minimizing the combined cost of vacancy days and under-pricing.
Why relational data changes the answer
Rental pricing is a market equilibrium problem. Unit UNIT02's predicted 35-day vacancy at $1,350 is not just about UNIT02. It is about UNIT02 relative to every other 1BR listing in the market: competing units at $1,300 with 12 photos, the seasonal demand index (March = 85/100, below peak), and the fact that ground-floor unrenovated units take 8 days longer than upper-floor renovated ones in this building. The revenue-optimal price ($1,280) accounts for all of these relationships simultaneously.
Flat pricing models treat each unit as an independent optimization. Graph-based models represent the full leasing market: units compete with each other for the same tenants, seasonal demand affects all units but not equally, and pricing decisions on one unit affect the competitive position of every other unit in the portfolio. SAP's SALT benchmark shows 91% accuracy for graph models vs 63% for gradient-boosted trees on relational tasks. RelBench confirms at 76.71 vs 62.44. In vacancy prediction, this accuracy difference means predicting days-on-market within 3-5 days (graph) vs 10-15 days (flat model). For a REIT pricing 7,000 units per year, that prediction accuracy translates to $8-15M in revenue optimization through tighter pricing that minimizes both vacancy loss and under-pricing.
Pricing a rental unit without considering the market network is like pricing an airline seat without checking what competing airlines charge on the same route. The right price depends not just on your seat (window, aisle, legroom) but on what alternatives the customer has. If your competitor just dropped fares 10%, your same seat is now overpriced. Graph-based vacancy prediction considers the full competitive landscape, not just the unit in isolation.
How KumoRFM solves this
Graph-powered intelligence for real estate
Kumo connects units, listings, applications, market data, and seasonal trends into a leasing graph. The GNN learns vacancy duration patterns from the property network: how pricing relative to comparable listed units affects days-on-market, how seasonal demand patterns vary by unit type and neighborhood, and how listing quality (photos, description) accelerates leasing. PQL predicts expected vacancy days per unit at different price points, enabling revenue-maximizing pricing.
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
UNITS
| unit_id | property_id | type | sqft | floor | renovated |
|---|---|---|---|---|---|
| UNIT01 | PRO01 | 2BR/2BA | 1,050 | 3 | Yes |
| UNIT02 | PRO01 | 1BR/1BA | 680 | 1 | No |
| UNIT03 | PRO02 | Studio | 450 | 5 | Yes |
LISTINGS
| listing_id | unit_id | asking_rent | photos | listed_date |
|---|---|---|---|---|
| LST201 | UNIT01 | $1,950 | 12 | 2025-03-01 |
| LST202 | UNIT02 | $1,350 | 6 | 2025-03-01 |
| LST203 | UNIT03 | $1,200 | 15 | 2025-03-01 |
APPLICATIONS
| app_id | listing_id | applicant_count | qualified_pct |
|---|---|---|---|
| APP101 | LST201 | 4 | 75% |
| APP102 | LST202 | 1 | 100% |
MARKET_DATA
| neighborhood | avg_rent_1br | avg_rent_2br | vacancy_rate | absorption |
|---|---|---|---|---|
| Oak Park | $1,300 | $1,800 | 5.2% | Positive |
| Downtown | $1,450 | $2,100 | 8.5% | Negative |
SEASONAL_TRENDS
| month | demand_index | avg_dom_1br | avg_dom_2br |
|---|---|---|---|
| March | 85 | 28 | 22 |
| June | 100 | 18 | 14 |
| December | 62 | 42 | 35 |
Write your PQL query
Describe what to predict in 2–3 lines — Kumo handles the rest
PREDICT FIRST(LISTINGS.leased_date, 0, 90, days) - LISTINGS.listed_date FOR EACH UNITS.unit_id
Prediction output
Every entity gets a score, updated continuously
| UNIT_ID | TYPE | ASKING_RENT | PREDICTED_DOM | REVENUE_OPTIMAL_RENT |
|---|---|---|---|---|
| UNIT01 | 2BR/2BA | $1,950 | 16 days | $1,920 |
| UNIT02 | 1BR/1BA | $1,350 | 35 days | $1,280 |
| UNIT03 | Studio | $1,200 | 12 days | $1,250 |
Understand why
Every prediction includes feature attributions — no black boxes
Unit UNIT02 -- 1BR/1BA at Oak Park Apartments
Predicted: 35 days predicted vacancy (revenue-optimal rent: $1,280)
Top contributing features
Asking rent vs market average
+3.8% above market
30% attribution
Low photo count vs similar listings
6 vs 12 avg
24% attribution
Seasonal demand (March = below peak)
85/100 index
19% attribution
Ground floor unit premium penalty
Floor 1 = longer DOM
15% attribution
Not renovated vs renovated comps
Unrenovated
12% 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 vacancy duration prediction
How does AI pricing differ from revenue management software like RealPage or Yardi RentMaximizer?
Traditional revenue management tools use rules and simple models (if occupancy drops, lower price). Graph-based models predict the specific vacancy duration for each unit at each price point, accounting for comparable listings, seasonal demand, unit-specific attributes, and portfolio-level interactions. The key difference is prediction specificity: instead of 'lower prices 3% for all 1BRs,' the model says 'UNIT02 should be $1,280 because of its floor, renovation status, photo count, and the competitive set.'
What is the revenue impact of pricing a unit $50 too high?
On average, $50 above optimal price adds 8-12 days of vacancy. For a $1,350/month unit, 10 extra vacant days costs $450 in lost rent, far exceeding the $50/month rent premium over a 12-month lease ($600). The math is clear: slightly lower pricing that fills the unit faster almost always maximizes total revenue. The model finds the exact tipping point where higher rent stops being worth the additional vacancy days.
How do listing photos affect vacancy duration?
Photo count and quality are among the top 5 predictors of leasing velocity. Units with 12+ professional photos lease 8-12 days faster than units with 6 or fewer amateur photos, all else being equal. This effect is consistent across markets and unit types. For property managers, investing $150-300 in professional photography per unit turn has an ROI of 5-15x through reduced vacancy days.
Can vacancy prediction help with renovation decisions?
Yes. The model predicts vacancy duration and achievable rent for renovated vs. unrenovated units, enabling ROI calculations for specific renovation packages. A kitchen renovation costing $8K that adds $150/month in rent and reduces vacancy by 5 days pays for itself in 4 years. The model quantifies these tradeoffs unit by unit, helping prioritize renovation capital on the units with the highest return.
How does seasonal demand affect pricing strategy?
Seasonal demand creates 30-50% swings in leasing velocity. A 1BR unit priced at $1,350 may lease in 18 days in June (peak season, demand index 100) but take 42 days in December (low season, demand index 62). The revenue-optimal price in December might be $1,250 because the cost of extended vacancy at the higher price exceeds the rent premium. Graph-based models adjust optimal pricing dynamically as seasonal patterns shift, rather than applying static seasonal adjustments.
Bottom line: A REIT with 20,000 units saves $8-15M annually by pricing units based on predicted vacancy duration. Kumo's leasing graph connects unit attributes, market conditions, and seasonal patterns to find the revenue-maximizing price point for each unit.
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.
