Meet releaRent, our new proprietary rent model

Hi there,

Rent is one of the most important numbers in an investment analysis. It is also one of the easiest to get wrong.

When we first built relea, we used a widely adopted third-party rental model. It gave us a useful starting point, but as more real properties moved through relea, its limits became hard to ignore. We saw markets with little or no coverage. We saw studio estimates exceed rents for full one-bedroom units. We saw the same address represented more than once in the underlying comp set, which could skew the result. Most importantly, we could not inspect, explain, or improve the model ourselves.

So we built our own.

We are proud to introduce releaRent, our proprietary model for estimating current market rent. It now powers the default rental estimate in relea and is the first in a planned series of data and modeling improvements. It demonstrates what relea can do with its own data and insights: make every assessment more relevant, transparent, and defensible.

Instead of returning one unexplained number from one source, releaRent builds a current local comp set from multiple nationwide rental-inventory sources, then normalizes, deduplicates, filters, and analyzes those listings as a distribution.

That difference matters.

Rental inventory is fragmented. A listing visible through one national rental surface may be absent from another, especially in thinner markets or for particular property types.

releaRent draws from multiple sources with meaningful live US rental inventory rather than depending on a single feed. The goal is not to collect the largest possible pile of listings. It is to achieve viable local density: enough currently available, relevant rentals to form a defensible comp distribution for the specific property being analyzed.

Getting there required more than connecting additional sources. We first built an inventory audit to understand what each source actually contributed. Rental addresses are often written differently across listing surfaces, so releaRent normalizes street-address variations and combines the normalized address with the bedroom configuration to identify the same rental appearing more than once. When records collide, the model keeps the richer version rather than allowing a widely syndicated listing to receive extra weight.

We then measured each source's coverage, its overlap with every other source, its exclusive inventory, and what disappeared when that source was removed. The analysis found meaningful overlap between some of the broadest sources, with the most overlapping source pairs sharing roughly 10% of their combined inventory on average. Address-based normalization and deduplication removed about 9% of merged source records on average before the model calculated a rent distribution.

What remained was a broader unique index, not an inflated listing count. Across the 28-profile evaluation, releaRent retained about 76 usable comps per profile on average, compared with about 15 for third-party rent model A, or roughly 5.2 times the usable inventory depth. The two comp sets had only about 1% average overlap, and 91.8% of their combined inventory appeared only in releaRent's source network.

That density translated into coverage. The multi-source model produced a usable comp set for 100% of the evaluated profiles. Third-party rent model A, a separate established rental model, covered 85.7%.

That does not mean every market will always have enough comparable inventory. The model requires a minimum usable sample and falls back safely when the comp set is too thin.

Rental markets can move quickly, and seasonality matters. An estimate built from rents observed months ago can miss what renters are seeing today.

releaRent pulls live available rental inventory when an analysis runs. That gives the model a current view of what renters can actually find in the market at that time, rather than relying only on a long historical look-back. This matters in markets where rents are moving quickly or where seasonal inventory changes the available comp set.

Current asking rent is not the same as a signed lease, so releaRent does not pretend it is. We treat it as timely market evidence, filter it carefully, and use the median of the resulting distribution rather than a simple average.

More data only helps when it is relevant data. releaRent filters comps using:

This addresses several problems we saw with the old approach. A studio should not be treated as a one-bedroom. An apartment should not quietly anchor the estimate for a detached house. The same rental syndicated across multiple listing surfaces should not receive multiple votes.

The model also removes statistical outliers when there is enough evidence and requires a minimum usable sample before releaRent can carry its name.

A local rental market is a range. Two otherwise similar units can rent differently because of condition, finishes, floor plan, parking, amenities, block quality, or timing.

releaRent therefore calculates the 25th percentile, median, and 75th percentile of the filtered comp distribution. The median remains the central estimate used in the analysis today, with the comp count and search radius providing context for the evidence behind it. The fuller distribution gives us a stronger foundation for future confidence and range displays.

This distribution-based foundation also gives us a path to make the model more specific over time. Planned improvements include accounting for renovation and finish quality, amenities, neighborhood quality, micro-location, listing age, concessions, and seasonal patterns. These can become modeled adjustments to an observable distribution instead of arbitrary changes to a black-box average.

We tested the model across 28 property and unit profiles in 11 metro areas, spanning single-family homes, townhomes, apartments, and small multifamily units. The evaluation included current lease rents from a real operating portfolio where those figures were available.

The clearest result was coverage: the evaluated multi-source model returned enough usable comps for all 28 profiles, while third-party rent model A covered 24 of 28. We subsequently refined the source mix for production based on inventory contribution, overlap, and reliability.

The additional inventory was not simply counted and averaged. Each profile was filtered independently and converted into a rent distribution. Some profiles retained more than 100 relevant observations after filtering, while thinner profiles still had to clear the model's minimum evidence threshold.

On properties with known lease rents, the results also showed why a transparent comp set is more useful than blind reliance on one estimate:

The model did not win on every property, and we do not want to imply otherwise. The same evaluation exposed segments that still need calibration, particularly some Chicago apartment profiles. That is an advantage of owning the model and its evidence: we can see where it is strong, see where it is not, and improve it against real outcomes.

releaRent is the headline, but it was not the only thing we shipped:

We also opened relea beyond the private test and introduced self-serve plans. If you missed that announcement, you can review the options here (25% launch discount available!).

releaRent is the first proprietary model in relea, but the larger goal is broader: combine better data with transparent methods so the numbers in an analysis are easier to trust and easier to challenge.

We are already using the evaluation results to improve how the model handles property condition, apartment submarkets, neighborhood differences, amenities, concessions, and seasonality. As those improvements prove themselves, they will become part of releaRent without changing how you work.

If you have a property where the market-rent estimate looks especially right or wrong, contact us. Those examples are some of the most valuable inputs we can get.

Try it for yourself now!

Thanks for building with us,

Simon
Founder, relea