About
How Prysm Works
Methodology and approach
What is Prysm
Prysm is a free, instant property valuation tool for Arizona residential properties. It provides comparative market analysis estimates by combining recent comparable sales data with a statistical model and AI-generated narrative analysis.
Prysm is designed for homeowners researching their property's value, buyers evaluating potential purchases, investors screening opportunities, and anyone who wants a quick, data-driven estimate without waiting for a formal appraisal or CMA from an agent.
The tool is operated by an individual as a personal project. It is not affiliated with any real estate brokerage, lending institution, or appraisal firm. Valuations are for informational purposes only and are not appraisals.
The Valuation Model
When you enter an address, Prysm runs through a multi-step process to produce an estimate. Here is how each stage works.
Step 1
Comparable Sales Retrieval
The model queries the Rentcast API to pull recent property sales within a defined radius of the subject property. This returns a set of raw comparable sales including sale prices, dates, square footage, bedroom and bathroom counts, lot sizes, and geographic coordinates.
Step 2
Sanity Checks and Filtering
Raw comps are filtered through several quality gates. Sales missing critical data (price, square footage, or coordinates) are removed. A square footage gating filter removes comps that are drastically different in size from the subject property, since a 900 sq ft home is not a meaningful comparable for a 4,000 sq ft property. Price-per-square-foot (PPSF) outlier detection identifies and removes sales with anomalous pricing that would skew the estimate.
Step 3
Ridge Regression Adjustment
The model fits a ridge regression on the comparable set to estimate the relationship between square footage and sale price. This allows the model to adjust each comp's price to what it would theoretically sell for if it were the same size as the subject property. Ridge regression is used instead of ordinary least squares because it handles small sample sizes and multicollinearity more gracefully with its regularization penalty.
Step 4
Composite Scoring
Each comp receives a composite quality score based on three factors:
Recency
30%
More recent sales are weighted more heavily, since they better reflect current market conditions.
Size Similarity
40%
Comps closer in square footage to the subject receive higher scores, since size is the strongest price predictor.
Bed/Bath Match
30%
Comps with similar bedroom and bathroom counts are scored higher, reflecting layout and functional similarity.
Step 5
Distance Weighting
An exponential decay function is applied to each comp based on its geographic distance from the subject property. Comps that are physically closer contribute more to the final estimate than distant ones. This reflects the reality that hyper-local factors like street, subdivision, and immediate neighborhood have a meaningful impact on property values.
Step 6
Top Comp Selection and Outlier Removal
The comps are ranked by their combined composite score and distance weight. The top 10 comps are selected. An iterative outlier removal pass then checks for any remaining price outliers within this final set, removing comps whose adjusted prices fall far outside the group's distribution. This ensures the final estimate is based on a tight, coherent set of comparables.
Step 7
Weighted Valuation
The final estimated value is computed as a weighted average of the adjusted prices of the selected comps, where the weights combine the composite quality score and distance weight. This means that the most similar, most recent, closest comps have the greatest influence on the final number.
Confidence Grading
Every valuation includes a confidence grade that reflects how much trust you should place in the estimate. The confidence grade is derived from three factors:
- Comp count: More comparable sales provide a stronger statistical basis for the estimate. A valuation based on 8 or more comps is more reliable than one based on 3.
- Price spread: A tight cluster of adjusted prices indicates strong market consensus around the estimated value. A wide spread suggests more uncertainty.
- Proximity: Comps that are geographically close to the subject provide more relevant pricing signals than distant ones.
Robustness Grading
In addition to confidence, each valuation receives a robustness grade that evaluates the quality and diversity of the underlying data:
- Coefficient of Variation (CV): Measures the relative dispersion of comp prices. A low CV indicates consistent pricing among comps; a high CV signals heterogeneous data.
- Herfindahl-Hirschman Index (HHI): Measures whether the valuation is overly dependent on a small number of comps. A balanced weight distribution across comps produces a more robust estimate than one dominated by a single comparable.
- Adjustment burden: Quantifies how much the regression model had to adjust comp prices to account for size differences. Large adjustments introduce more uncertainty, so lower adjustment burden indicates more naturally comparable sales.
AI Narrative Analysis
Each valuation report includes a narrative summary generated by Anthropic's Claude AI. The AI receives the valuation data, comparable sales details, confidence metrics, and robustness metrics, and produces a plain-language interpretation of the results.
The narrative is designed to make the statistical output more accessible by highlighting key patterns, explaining the basis for the estimate, noting potential concerns, and providing context for the confidence and robustness grades. The AI does not have access to information beyond what the model provides, and its narrative should be read as an interpretation of the data, not independent analysis.
How Prysm Differs from Zillow and Redfin
Zillow's Zestimate and Redfin's estimates use massive proprietary datasets, tax records, MLS feeds, and machine learning models trained on millions of transactions nationwide. They are powerful tools with broad coverage.
Prysm takes a different approach:
- 01Transparent methodology. The entire model logic is described on this page. You know exactly how the number was derived, what weights were used, and what data went into it. Zillow and Redfin treat their models as proprietary black boxes.
- 02Comp-level detail. Prysm shows you every comparable sale that contributed to the estimate, its weight, its adjusted price, and its distance from the subject. You can evaluate the comps yourself.
- 03Quality metrics. Confidence and robustness grades give you a structured way to evaluate how much trust to place in the estimate, rather than just a single number.
- 04AI interpretation. Rather than just numbers and charts, you get a narrative analysis that explains the estimate in context.
- 05No account required. No sign-up, no data collection, no lead generation. Enter an address and get a result.
The tradeoff is that Prysm has a narrower dataset (Rentcast vs. MLS), covers only Arizona, and uses a simpler model. For a quick, transparent estimate, it fills a useful niche.
Limitations
No automated valuation model is a substitute for professional judgment. Prysm has specific limitations you should be aware of:
- Arizona only. The model is designed and tested for Arizona residential properties. It does not support other states.
- No interior condition assessment. The model cannot evaluate renovations, upgrades, damage, deferred maintenance, or any factor requiring physical inspection.
- Data dependency. The estimate is only as good as the comp data available from Rentcast. In areas with few recent sales, the model may produce less reliable estimates or fail to generate one entirely.
- Residential focus. The model is designed for single-family homes, condos, townhouses, and small multi-family properties. It is not suitable for commercial, industrial, agricultural, or vacant land valuations.
- Lagging indicator. Because the model relies on closed sales, it reflects where the market has been, not necessarily where it is going. In rapidly shifting markets, estimates may lag current conditions.
- No forecasting. Prysm does not predict future values, appreciation rates, or market trends.
- Unique properties. Highly unique properties (custom builds, historic homes, properties with unusual features) may not have meaningful comparables, resulting in unreliable estimates.
For important financial decisions, always consult with licensed professionals. See our Valuation Disclaimer for full details.