Forecasting demand and dynamic pricing in real estate markets

Pricing changes absorption. Cut rent and units lease faster (higher absorption, lower rent per unit). Raise rent and they lease slower. The model's job is to find the price that maximizes total revenue over time, accounting for how long units sit and what future demand looks like.

The revenue-per-available-unit lens

Hotels track RevPAR (revenue per available room). Multifamily operators think similarly: revenue per available unit across the whole property. A model that pushes rents too high wins on headline rent but loses on occupancy, and total revenue falls. The optimization target is the product, not the price alone.

What the demand model actually looks at

A demand forecast for a single property blends several signal categories.

Internal pipeline signals

• Current occupancy and the number of upcoming lease expirations
• Traffic: tours, website visits, and leads per week
• Conversion rateConversion rateThe percentage of visitors or prospects who complete a desired action (purchase, sign-up, contact form), calculated as conversions divided by total opportunities.Voir la définition complète → from lead to signed lease
• Renewal probability for existing tenants (renewals are cheaper than turnover)

Seasonality

Leasing demand is highly seasonal in most US markets. Spring and summer see far more moves; December is slow. A model that ignores this will overprice in January and underprice in June. Seasonality is usually the single strongest pattern in the data.

Submarket signals

A submarket is a defined geographic slice of a metro (for example, a specific set of neighborhoods or ZIP codes). Signals include local job growth, new supply coming online (a competitor delivering 300 units nearby suppresses your pricing power), and comparable asking rents.

Macro context

Interest rates, for-sale housing affordability (when buying is expensive, more people rent), and regional migration trends.

A simplified feature set

Here is the kind of tabular data a pricing model consumes, one row per unit per day:

Note what is NOT here: the resident's name, demographic data, or anything that could introduce discrimination risk. That exclusion is deliberate, and we will return to why.

From forecast to price recommendation

The forecast estimates demand. The pricing engine turns that into a number.

A common approach: for each candidate price, the model estimates the probability the unit leases within a target window. Multiply expected occupancy by rent across a time horizon, and you get expected revenue for each price. Pick the price that maximizes it.

In practice, operators layer in business rules on top of the model:

• Floors and ceilings so recommendations stay sane
• Maximum daily change (no 15 percent overnight swings that anger prospects)
• Renewal discounts to retain good tenants and avoid turnover costs (a vacancy plus make-ready plus new leasing commission can easily exceed a month of rent)

The human still approves. Most systems recommend; a revenue manager reviews exceptions. Full autopilot is rare because a bad feed (a competitor mistakenly listed a unit at half price) can poison the model.

🎬 [VIDEO: "How Revenue Management Works in Multifamily" — youtube.com — a plain-language walkthrough of automated rent pricing for apartments]

The RealPage antitrust problem

You cannot teach this topic honestly in 2026 without the legal context. RealPage, the dominant revenue management vendor, became the subject of a US Department of Justice antitrust lawsuit alleging its software helped landlords coordinate rents by pooling competitors' nonpublic data. Several cities and states have passed or proposed restrictions on algorithmic rent-setting software.

The legal question is not "is dynamic pricingdynamic pricingAutomatically adjusting prices in real time based on demand, competition or user behaviour to optimise revenue, margin or conversion.Voir la définition complète → bad?" It is narrower: does software that ingests competitors' private data and nudges many landlords toward similar prices function as a coordination mechanism? Using public asking rents is standard practice. Sharing confidential future pricing across competitors is where regulators focus.

You can follow the case through the DOJ Antitrust Division's public case documents. The takeaway for operators: know what data your vendor uses, and whether it is public or shared confidentially.

Fair housing and the discrimination trap

The US Fair Housing Act prohibits housing discrimination based on protected characteristics (race, national origin, familial status, disability, and others). A pricing model that uses ZIP code or other proxies can produce disparate impact: a neutral-looking variable that correlates with a protected class and produces discriminatory outcomes.

This is why serious operators exclude demographic features and audit models for proxy effects. "The algorithm did it" is not a legal defense.

Building intuition: a worked example

Imagine a 200-unit property. It is late May (peak season, week 22). You have 12 vacant two-bedroom units and 18 expirations coming in the next 45 days. Weekly leads are strong at 25, conversion is running 12 percent, so you expect roughly 3 leases per week from that traffic.

Current asking rent is 2,400 dollars. The model sees strong seasonal demand and low competitor availability nearby. It estimates that at 2,475 dollars, absorption slows only slightly (you still lease in about five weeks) while capturing 75 more dollars per unit per month across a 12-month lease.

The recommendation: raise to 2,475. The revenue manager checks that no competitor just dropped a large block of units and approves.

Now flip it to January (week 3). Same property, but leads fall to 10 per week and you still have 12 vacant units. Holding at 2,400 means units sit for months, and each empty month costs a full 2,400. The model recommends dropping to 2,325 to accelerate absorption. Losing 75 dollars per month hurts less than losing an entire month of rent.

That asymmetry (vacancy loss versus rent optimization) is the heart of the discipline.

Where these models break

Thin data. A single small property does not generate enough leases to train a robust model. Vendors pool data across many properties, which is exactly what creates the antitrust scrutiny. The value and the legal risk share a root cause.

Regime shifts. Models trained on stable years fail when conditions break. A sudden wave of new supply, a local employer layoff, or a rate shock can invalidate historical patterns. Human oversight matters most precisely when the model is least reliable.

Feedback loops. If every operator in a submarket uses similar software reading similar signals, prices can move together even without explicit coordination. Regulators and researchers are actively studying whether this alone raises rents.

Key Takeaways

• Vacancy is the enemy, not low rent. An empty unit loses full rent permanently, so pricing optimizes total revenue over time, balancing rent against absorption speed.
• Seasonality is the strongest signal. Peak leasing season supports higher rents; winter demands discounts to keep units moving. Any model ignoring this fails.
• Public data is defensible; shared confidential data is not. Know whether your vendor uses public asking rents or pooled private competitor pricing, because that distinction drives the antitrust risk.
• Exclude protected-class proxies and audit for disparate impact. Fair housing law applies to algorithms, and "the model decided" is not a defense.
• Keep a human in the loop. Models fail hardest during regime shifts and bad data feeds, which is exactly when a review layer earns its keep.