MachineLearningforRealEstateUnderwriting:DealSight

Who is the client and what did they need?

Cedarline Homes is a real estate business in the US Midwest. It buys houses, renovates them, and sells them again. Most of the profit on a deal is decided on day one, by the price paid for the house. So the underwriting step, where the team works out what to offer, decides how well the company uses its money.

Cedarline wanted to grow. That meant looking at many more deals, but without hiring underwriters at the same rate. It also wanted offers to be more consistent, and to make real use of 15 years of its own deal history along with current market data.

They asked Xpiderz to design and deliver the system. We call it DealSight. It does not replace the underwriters. It gives them three predictions for every property, and those predictions feed the Maximum Allowable Offer (MAO) calculation that sets the offer price.

• ARV. After Repair Value, or what the house should sell for once the work is done.
• Rehab Cost. What the renovation will cost, split into ten cost categories.
• Time to Sell. The number of days from buying the house to selling it.

How does DealSight work?

Xpiderz handled the full build: the data foundation, the machine learning models, the Salesforce integration, and the MLOps setup that lets the system keep learning after launch.

DealSight uses Cedarline's own deal history together with outside data to make its predictions. Each of the three models has its own design. The ARV model works from comparable sales in the market. The Rehab Cost model predicts ten cost categories at once and is gated by the contractor estimate. The Time to Sell model predicts days from purchase to resale.

Every prediction comes with a SHAP explanation, so an underwriter can see what pushed a number up or down. The results show up inside Salesforce and feed the MAO calculation, taking over from the fixed formula as the main input to offer pricing.

What did the client get out of it?

DealSight is live in production, inside the Salesforce workflow the underwriters already use. In the holdout test, the new models beat the old benchmark on two of the three predictions and came out about even on the third.

• ARV. 31% lower error, down from $24,577 to $16,919.
• Time to Sell. 69% lower error, down from 23.7 days to 7.28 days.
• Rehab Cost. Near parity, at $4,554 against a benchmark of $4,569.
• Three strategies at once. Each model makes its predictions for three strategy variations at the same time, so underwriters can compare options side by side.

These numbers feed the MAO calculation, so underwriters now start from stronger inputs when they decide what to offer. That lets Cedarline judge deals more consistently and take on more of them without overhead rising in step.

The system also gets better with time. Automated ingestion, monthly retraining, evaluation, and drift monitoring mean every new purchase and every market shift can shape the next model version.