By Paul Rayburn August 11, 2025
I just posted this on linked-In but the format limitations make reading the graphs I shared a bit difficult, if not impossible for some. Hopefully the images are easier for you to read in this format. Of course, I’ll take the opportunity to expand a bit on the content with a few more allowable characters than what LinkedIn provides. I start by providing a bit of an introduction to a chat with Chat-GPT and a TLDR (Too Lazy Did Not Read) *scroll down a bit, for those that prefer that briefing and then for the diehards, following the chat, a bit of a summary of some key points.
As some may have noticed, I tend to give a bit of a nautical touch to my posts. This time I’m going to blend in a bit more of a landlubber sport and reference QBs.
*not sure what the difference is between an AVM and an HPI ? I added links to detailed definitions in the links below or jump to the bottom of the page for my brief description and then jump back up here.
Begin Post
Getting a bit tired of armchair QBs who say appraisals are done and AVMs will do all the work? 99% of those are in positions where they don’t really care about the value, they simply want fast and cheap, they want mass mortgage volume!! and the lending industry is pumped on those prospects.
If you read through my chat below, I first got a reading on what an HPI was saying from ChatGPT, a second opinion if you will, and then compared it with a set of comps from that market that I had dropped into a couple scatterplots after I took some time to sort through the garbage data in that market.
Garbage in, garbage out is real. You hear it a lot, but in this case the data is a mess, with incorrectly listed design styles, floor areas, etc.
So, while there may be some markets with good data, that’s not always the case. I’d venture to guess it’s actually more rarely the case.
Also, I know many of my my appraiser cousins in the US are dealing with the GSEs FHA etc. and are being pressed to use tools like an HPI for the sake of making an adjustment, because of some loosely applied logic that says because appraisers were not always making adjustments when the “indexes” were saying something else, then the appraiser must be wrong. Is that poison pill making its way into the hearts and minds of big decision makers on an international level? Hopefully not, but it’s a tempting prospect for some and likely finding sympathizers.
Let’s explore a little further shall we.
By the way when Chat GPT points out how indexes can be less than reliable when the sample size is small, that index I provided it initially is a national model, with algorithms that it uses to correlate macro to almost micro neighborhood markets. The sample size of 15 or so sales I used for the quick 6 month comparison, is certainly a small sample but I’d argue far more effective at reflecting actual market reactions. (Maybe it was thinking the HPI was created from a similarly sparse data set.) The larger data set in the first graph, the red line, is the general area, several hundred sales over the same period and you can see it follows the same “volatile market” (ChatGPTs words not mine) but I would generally characterize it that way based on those visual trends alone.
Before you jump fully onboard the mass modeling boat, make sure there is ample ballast on that ship!
TLDR: I show ChatGPT an HPI and ask for its take on what it sees with minimal prompting. Its response is reasonably expected based on the chart visual (I didn’t fact check every detail or that it was 100% accurate as it was just for this test, but it looks pretty close at a glance). I show it a conforming data set and it immediately determines that the market may very well not be as volatile and the trends and significant shifts indicated by the HPI may not be reliable.
For clarity, the Bot seems to misinterpret that could be a different market, but in fact it is the same competitive market.
Here’s the chat GPT Chat
USER:
Chat-GPT5:
Looking at the past 12 months (July 2024 to July 2025) for Westmount – Townhouse:
- The benchmark price started around $450,000 in July 2024, rose to a peak of $495,600 in November 2024, then dropped sharply to $427,400 in December 2024.
- After that, there was a partial recovery to $472,700 in March 2025, another dip, then a high point of $473,500 in June 2025.
- July 2025 closed out at $451,700, only slightly above where it began a year ago.
Overall trend:
- The market has been volatile, with short-term swings of 8–16% within months.
- Despite some mid-year surges, prices ended the year roughly flat (+0.4% from July 2024).
- This pattern suggests a lack of sustained upward momentum, with possible sensitivity to seasonal or external factors (e.g., interest rate changes, supply spikes).
Meanwhile, North Okanagan – Townhouse prices (the pink line) appear higher overall but also show a softening trend since late 2023.
If you like, I can calculate the exact volatility percentage and seasonal price change rates to quantify how choppy this market has been. That would make the interpretation much stronger for an appraisal report.
USER:
Chat-GPT5:
Looking at these scatterplots for conforming townhomes in the same market, the story is quite different from the volatile Westmount trend you showed earlier.
Sold Price vs Sold Date (left chart)
- Prices are clustered fairly tightly between roughly $550,000 and $625,000 over the past six months.
- The regression line is almost flat, indicating very little short-term market change — essentially stable prices.
- Even the lowest and highest points are not dramatically far from the mean, which suggests low volatility compared to Westmount.
PPSF vs Size (right chart)
- Strong, clear negative correlation: smaller units are selling at higher prices per square foot, larger units at lower PPSF.
- The spread around the trend line is tight, showing predictable size-to-price relationships.
Key contrast to Westmount:
- Westmount’s benchmark index showed swings of over $45,000 in just a few months.
- This conforming townhome data set shows a much narrower range, meaning the volatility you saw in the Westmount chart is likely due to that specific segment’s composition, supply-demand imbalance, or small sample distortions — not a broad market trend for all townhomes in the region.
If you want, I can quantify the coefficient of variation for both datasets so we can put a number on how much more volatile Westmount has been. That would make the comparison really concrete.
End of chat with Chat-GPT
Clarity on the chat!
For further clarity, I understand that an HPI is not necessarily an AVM. In reality, an AVM in this market would essentially have this same data to work with, or possibly even less detailed than the HPI shown, depending on the data source used.
HPI’s and AVMs have purposes, and they have improved with continually advancing technology. However, there are still very large gaps, and appraisers will continue to be an important part of the valuation space. Appraisers will need to continue to adapt to using improved technology and tools. All of these advances applied in the proper circumstances and not forced in the wrong manner will improve the overall experience for those in the valuation industry and ultimately the end consumer.
If you have read this far you are probably keenly interested in the Real Estate Valuation space. So, speaking of appraisers staying up to speed, you may want to check out George Dell and one of his recent timely posts (they almost all seem timely and at the same time, timeless. This one literally just popped into my mailbox as I was writing this. You can read that here. It digs in a bit, into historical and current causes of friction in the lender/ appraiser interface.
If you really want to begin digging in more on AVMs AVMetrics is one of the only independent AVM testers that does not market its own AVMs. Their newsletter is packed full of information. This Aug 2025 Whitepaper by Sennot Consulting helps to highlight some of the key issues in a fairly brief intro. And his 2023 paper on the AVM bias debate also provides some insights.
HPI vs. AVM – Quick Take
An HPI tracks how the value of a “typical” (mean, median or benchmark) home in a market changes over time. In theory, that trend can be used to update an older sale of a similar home to a current, or conversely, to a historical value. A benchmark home is similar to the “average” home, but rather than use average prices, it uses the most common characteristics of the homes in that market to create the benchmark and then calculate how those prices change. It’s a more advanced method and is the method used by the HPI model I presented to Chat GPT.
An AVM estimates the value of a “specific” home using similar sales and modelled adjustments, often using some of the same math behind HPIs to adjust for differences in sale dates, also known as market condition adjustments.
The catch?
An HPI is generally generic, not fine tuned to a subject property (unless it is a model match for the benchmark) and even when modelled to refined markets, it often becomes less reliable.
An AVM is adapted to analyze a specific property, again the modelling breaks down for those properties in more sparse markets or sometimes even just a few unique or less common features begin to break the model. Jump back up to the story.
Observers versus Doer’s
Next class: We’ll talk about how sometimes being on the sidelines, observing, can produce exceptional insights.
