Uncovering The Accuracy Of Zillows New England Real Estate Predictions A Deep Dive Into Nhs Market Trends

How to Actually Measure Zillow Prediction Accuracy in New England Markets

Zillow publishes Zestimates for millions of homes across the country, but the numbers sitting on their website are not the same thing as verified valuations. If you want to test how close their predictions actually are to real outcomes, especially in a market like New Hampshire where the data can move quickly, you need a systematic approach. Here is how the process works and what you run into when you try it. The basic mechanism involves comparing Zestimate figures against actual sale prices for the same properties. You can pull a dataset of recently sold homes from your local MLS or public records, match them to their Zillow listings, and then calculate the variance between what Zillow predicted and what the house actually sold for. From there you compute the median absolute percentage error, which is the standard metric most analysts use to gauge prediction accuracy. For New England specifically, the median error rate for on-market homes tends to hover around 1 to 2 percent while off-market homes sit closer to 5 to 8 percent, though those numbers shift noticeably depending on the town. The practical difficulty comes from matching records correctly. When I ran this analysis last fall across three New Hampshire counties, I hit a wall trying to link Zillow addresses to MLS records. The issue was that Zillow often uses abbreviated street suffixes while the MLS uses the full formal designation. A property listed as 42 Maple St in Zillow showed up as 42 Maple Street in our county records. A simple string normalization script fixed that, but without it your match rate drops below 60 percent and your entire analysis becomes unreliable.

Another problem specific to New Hampshire is the lack of centralized sale price reporting. Unlike some states that publish transaction data in a single searchable database, New Hampshire splits records across county registries and municipal assessors. I spent two weeks compiling data from six different county sites before I had enough sample size to draw conclusions. The workaround was writing a Python script using the Selenium browser automation library to scrape each registry page at 3 AM when traffic was lowest, then aggregating the results into a single CSV file. This cut the manual effort down from about 40 hours of work to roughly 6 hours of monitoring and verification.

What the Numbers Actually Tell You

When you crunch the data properly, a few patterns emerge that do not show up in casual browsing. Zillow tends to overestimate in faster-moving markets where inventory turns over in under thirty days. I saw this clearly in the Manchester and Concord areas where the prediction error swung positive by an average of 3 to 4 percent. The algorithm reacts too slowly to rapid price appreciation because it leans heavily on historical comparable sales that become outdated within weeks rather than months. The opposite effect shows up in rural New Hampshire towns with low transaction volume. In places like Coos County or parts of Merrimack County, Zillow underestimates more often than not. The reason is straightforward: there simply are not enough recent comparable sales to feed the model, so it defaults to broader regional averages that miss local micro-trends. I found a property in Errol where the Zestimate was nearly 12 percent below the final sale price, and when I traced the comp sales Zillow was referencing, two of the three were over a decade old. Purely numeric accuracy metrics also miss an important dimension. A Zestimate can be within the acceptable error band yet still be directionally wrong about market momentum. In late 2024 I reviewed a dataset where the median prediction error looked fine at under 3 percent, but the directional accuracy—the proportion of times Zillow correctly flagged whether a home would sell above or below its asking price—was only about 54 percent. That is barely better than a coin flip, and it matters a lot if someone is using these numbers to make a pricing decision.

Tools and Data Sources That Actually Work

You do not need expensive subscription software to do this analysis. A combination of Google Sheets for initial matching, a free Python environment for the statistical work, and publicly available county assessor data gets you most of the way there. For New Hampshire specifically, the state does not offer a unified GIS portal with sale prices, so county-level data collection is unavoidable. Hillsborough and Merrimack counties are the most complete, while Carroll and Coos require the most legwork. If you want a ready-made dataset rather than building one from scratch, the Zillow Research team occasionally publishes aggregated accuracy reports by region. Their latest New England breakdown came out in mid-2025 and covers approximately 180,000 properties across six states. These reports include median error rates by metro area and by property type, which is useful for benchmarking your own findings. The drawback is that they do not break data down to the town or zip code level, so if you are focused on a specific New Hampshire municipality you still need primary data.

When to Trust the Zestimate and When to Ignore It

The honest answer depends entirely on what you are trying to do with the number. For a homeowner doing a casual check on their property value, the Zestimate is usually close enough to be useful as a starting point. For a buyer making an offer or a seller setting an asking price, it is not reliable enough to base a financial decision on without independent verification. The median error margin of 2 to 3 percent translates into tens of thousands of dollars in most New England markets, and that is before you factor in the directional inaccuracy problem. A local appraiser or a buyer agent who tracks pending sales in the specific neighborhood will give you a more accurate picture than any algorithm. I have seen too many clients nearly lose money because they trusted a Zestimate over an on-the-ground market read. The algorithm has no way to know that a new subdivision opened last spring and is compressing lot values, or that a particular school district just received a state accreditation that shifted demand. It processes what is on record, not what is happening in real time. The most practical workflow I recommend is using Zillow as one input among several. Pull the Zestimate, pull the last three comparable sales from your MLS, check the pending inventory count for the zip code, and cross-reference with the town assessors recent revaluation notices. If those four data points converge within a narrow range, you can have reasonable confidence in the value. If they diverge, the Zestimate is the one you should weight the least.