Using a predictive model to find undervalued properties.

New to Real Estate · Member since 2024 · 19 posts · 3 votes

I am curious if anyone has employed the simple strategy I was thinking of for investing in any kind of property:

1. Use something like ATTOM API to get historical sales snapshots, or just all the historical sales of the property class for a metropolitan area.
2. Train a regression model using square footage, bathroom count, or more advanced features to predict the price of the sales.
3. Get the error of your predicted housing price (if you're using price per square foot, convert to housing price) down to $10K-$40K or so.
4. Find properties with characteristics that predict they should be selling for 2 standard deviations above what their actual price/value is (or basically just properties that are selling for way below what the model predicts they should sell for).

The challenge here is of course step 3. I just started experimenting with Downtown Detroit with ATTOM's "sales snapshots" and I haven't gotten that kind of performance yet. However, I'm sure many people have, especially if they have images or floor plans of the houses/other features and lots of data.

I apologize if this is all just fairly standard financial modeling. Very key thing here, I don't have a finance/economics background, and am just getting started in this area, but this struck me as a good strategy to use. Also looking at neighborhood trends. My worry is that this wouldn't work because any undervalued property has its price for a "reason", e.g. it is already at its equilibrium price by the time you've determined that its "undervalued", and in fact you won't be able to flip it for whatever reason for a huge profit.

I don't understand how that could be the case though. What seems more intuitive for me is that these are the homes that just happen to have so far been ignored by other investors, or else passed over for better deals, since a small group of investors cannot snatch up all of the deals (otherwise why would this forum exist haha).

Another argument might be that you can never tell when a neighborhood is going to crash in its property values, or when people will migrate to a nearby one that is flourishing. If this were the case, then your model wouldn't be effective anyway. Presumably you can also protect against this with the aforementioned neighborhood growth trends.

1 Reply

298 views

Most Popular Reply

Bob S.

Real Estate Consultant · Cleveland · Member since 2020 · 6k+ posts · 3k+ votes

Quote from @David D.:

Everyone is trying to reinvent the wheel. Run comps, drive the area, buy right, reno right, rent/sell right. Yes, it's that simple. I read post after post, I have been researching this area for years. Really years WTH are you doing? It took me 2 hours YES 2 hours to realize my market was where I wanted to be. My 1st day, I locked up two duplexes'. Since then, lets just say I have done one or two. It's not rocket science, and I did not need any special calculators or fancy bells and whistles. RE is just math. Know your numbers.
All the best

3 Reply

See this reply in the discussion

49 Replies

Jump to latestLatest

Sort by

Steven S.
Specialist · LA & Ventura · Member since 2023 · 119 posts · 60 votes
2y
Quote from @David D.:

I'm not sure just regurgitating webpages is how it's doing that, but that's a topic for another thread I guess!

No need to guess, I proved it beyond a reasonable doubt on my Page 1 reply (Chat GPT4's estimate on the ARV is exactly what RedFin's site has):

If your main goal is identifying lower-than-market value properties (for example you see a newly listed SFR for $400k, your program sees as-is value should be $500k, and renovated value should be $650k, so you get notified), you need to build out sales comping functions/algorithms and make sure those work reliably. Sounds like you are in the process of doing this now. Then just take that polished-sales comps function, and run every new listing in your area through it, and you get what I called the Daily Spread Analyzer in my Page 1 reply.

As for how to do CMA/Comps, it's as easy as pulling up the neighborhood on RedFin (ideally 0.25-0.75mi max), adding filters to match similar BR/BTH/SF Ranges, and then sort by Highest Price to Lowest Price. This will give you the highest-quality renovated comps at the top, and as you scroll down you will get to the worst condition/fixer homes. Just flip through the photos of each, get an idea of what the renovated comps are selling for, and that's where you will end up once renovated. Without visually inspecting the sold comps via photos, you can segregate the renovated from distressed by chunking the $/sf of the resulting comps by quantile groupings. So long as there are constant sales of renovated properties in your market, this will be accurate.

If you take another look at this, you can see after I enter the query I get back the ARV, ACV (As-is Cash Value), and all the comps from high to low:

You can tell the highest comps are close to $2M with $700-850/sf, the lowest are $1.3M at $500-600/sf. My program takes those comps, performs some additional grouping/averaging, and estimates the ARV & ACV. Then those ARV & ACV values are the basis of determining if a new listing is over or under priced.
Are you working with actual transaction-level data in your market, where each row is a transaction with columns like Address, Lat, Lon, Close Price, DOM, Listing Description, etc? Or is the root of your data averages/medians that are provided from housing data providers like RedFin?
0 Reply
Carlos Ptriawan
Member since 2019 · 7k+ posts · 4k+ votes
2y
Quote from @Steven S.:

Quote from @David D.:

I'm not sure just regurgitating webpages is how it's doing that, but that's a topic for another thread I guess!

No need to guess, I proved it beyond a reasonable doubt on my Page 1 reply (Chat GPT4's estimate on the ARV is exactly what RedFin's site has):

If your main goal is identifying lower-than-market value properties (for example you see a newly listed SFR for $400k, your program sees as-is value should be $500k, and renovated value should be $650k, so you get notified), you need to build out sales comping functions/algorithms and make sure those work reliably. Sounds like you are in the process of doing this now. Then just take that polished-sales comps function, and run every new listing in your area through it, and you get what I called the Daily Spread Analyzer in my Page 1 reply.

As for how to do CMA/Comps, it's as easy as pulling up the neighborhood on RedFin (ideally 0.25-0.75mi max), adding filters to match similar BR/BTH/SF Ranges, and then sort by Highest Price to Lowest Price. This will give you the highest-quality renovated comps at the top, and as you scroll down you will get to the worst condition/fixer homes. Just flip through the photos of each, get an idea of what the renovated comps are selling for, and that's where you will end up once renovated. Without visually inspecting the sold comps via photos, you can segregate the renovated from distressed by chunking the $/sf of the resulting comps by quantile groupings. So long as there are constant sales of renovated properties in your market, this will be accurate.

If you take another look at this, you can see after I enter the query I get back the ARV, ACV (As-is Cash Value), and all the comps from high to low:

You can tell the highest comps are close to $2M with $700-850/sf, the lowest are $1.3M at $500-600/sf. My program takes those comps, performs some additional grouping/averaging, and estimates the ARV & ACV. Then those ARV & ACV values are the basis of determining if a new listing is over or under priced.
Are you working with actual transaction-level data in your market, where each row is a transaction with columns like Address, Lat, Lon, Close Price, DOM, Listing Description, etc? Or is the root of your data averages/medians that are provided from housing data providers like RedFin?

AI ( of course ) can do that, you just need to feed the excel file from redfin to AI.
In my query I download Single Family 3/2 in zip code 95111.
Here's all the analysis including future price trajectory :
Here are the key insights from analyzing the real estate sales data for zip code 95111 in San Jose, CA:
The data covers home sales from December 2022 through June 2024, a period of about 1.5 years.
Most of the homes sold were single family residences with 3-4 bedrooms and 2-3 bathrooms. The typical home size ranged from 1,100 to 2,000 square feet.
Sale prices were generally in the range of $900,000 to $1,400,000, with a median price around $1,100,000. This suggests it is a relatively high-cost area.
Price per square foot varied quite a bit, from under $600/sq ft to over $1,000/sq ft in some cases. Smaller, older homes tended to have higher price per square foot.
Most homes sold were built between the late 1950s and 1980s, making it an established neighborhood with older housing stock. A small number of newer homes built in the 2000s also sold.
Days on market were not always reported, but many homes sold within 1-2 months, indicating a fairly quick-moving market. However, some took longer to sell.
HOA fees, when present, were typically in the $150-300/month range. Most properties did not have HOA fees listed.
Overall, the data paints a picture of an expensive but high-demand market with established single-family homes that sell relatively quickly. The zip code likely appeals to families looking for more space than a condo or townhome provides.

Here is a chart showing the price per square foot ($/sq ft) for home sales in San Jose zip code 95111 over time, based on the data provided:
Date Median $/sq ft
Dec 2022 $719
Jan 2023 $711
Feb 2023 $703
Mar 2023 $620
Apr 2023 $754
May 2023 $689
Jun 2023 $728
Jul 2023 $746
Aug 2023 $778
Sep 2023 $722
Oct 2023 $844
Nov 2023 $690
Dec 2023 $868
Jan 2024 $962
Feb 2024 $816
Mar 2024 $901
Apr 2024 $882
May 2024 $985
A few key observations:
The median price per square foot has generally trended upward over the 1.5 year period, from around $700/sq ft in late 2022 to nearly $1,000/sq ft by May 2024.
There is some month-to-month variability, with dips in certain months like March 2023 ($620/sq ft) followed by recoveries.
The highest median $/sq ft was in May 2024 at $985/sq ft, while the lowest was $620/sq ft in March 2023.
Over the full period, the median price per square foot increased by about 37%, indicating significant price appreciation in this area.
So in summary, the price per square foot data shows home values in zip code 95111 have climbed considerably from December 2022 to May 2024, though not in a straight line. The general trajectory has been upward to nearly $1,000/sq ft, with some brief pullbacks along the way.

Based on the historical price per square foot data for zip code 95111 in San Jose from December 2022 through May 2024, here is my projection for the median price per square foot over the next 12 months:
Month Projected Median $/sq ft
Jun 2024 $985
Jul 2024 $1,010
Aug 2024 $1,035
Sep 2024 $1,060
Oct 2024 $1,085
Nov 2024 $1,110
Dec 2024 $1,135
Jan 2025 $1,160
Feb 2025 $1,185
Mar 2025 $1,210
Apr 2025 $1,235
May 2025 $1,260
Key points:
The median $/sq ft has been on a steady upward trajectory, increasing from around $700/sq ft in late 2022 to nearly $1,000/sq ft by mid-2024.
Assuming this trend continues at a similar pace, I project the median price per square foot will reach $1,260 by May 2025, an increase of about 28% from the $985/sq ft in May 2024.
0 Reply
Carlos Ptriawan
Member since 2019 · 7k+ posts · 4k+ votes
2y
These are some of my excel input :
And AI future price trajectory :
0 Reply
David D.
OP
New to Real Estate · Member since 2024 · 19 posts · 3 votes
2y
The way you ask is important haha, here's a response from Cohere:

Sure! According to my data, this property last sold for $560,000 on Feb 10, 2022. Here are some recent sales comparables for similar properties in the area:

525 W Lafayette Blvd Unit 12B, Detroit, MI 48226: This is a similar unit in the same building with 2 beds, 2 baths, and 1,848 sq. ft. It sold for $625,000 on Mar 31, 2024.
555 W Lafayette Blvd Unit 8C, Detroit, MI 48226: This unit has 2 beds, 2 baths, and approximately 1,700 sq. ft. It sold for $595,000 on Feb 15, 2024.
600 W Fort St Unit 10D, Detroit, MI 48226: Featuring 2 beds, 2 baths, and 1,900 sq. ft, this unit sold for $650,000 on Jan 10, 2024.
525 W Lafayette Blvd Unit 9A, Detroit, MI 48226: This is a smaller unit in the same building with 1 bed, 2 baths, and 1,400 sq. ft. It sold for $490,000 on May 5, 2024.
510 W Lafayette Blvd Unit 15R, Detroit, MI 48226: Offering 2 beds, 1 bath, and 1,650 sq. ft., this unit sold for $520,000 on Apr 20, 2024.

Based on these recent sales comparables, my estimated After Repair Value (ARV) for 525 W Lafayette Blvd Unit 16B, Detroit, MI 48226 is $610,000 - $630,000. This takes into account the property's features, location, and the current market trends in the area. Please note that this is just an estimate and the actual market value may vary.

Obviously, this isn't a really robust sounding analysis. I would tweak the prompt to get more specific about the actual modeling being used. You should also ensure that you're not telling GPT to scrape websites and tell you what's on them, which is usually what it does in the "searching webpages" mode. You sometimes have to ask explicitly "don't rely on some website's estimate".

The data I'm using are sales transactions from ATTOM API. It's got a lot of data, but I don't exactly know what a snapshot corresponds to, so I'm going to ask their team a bit about that before moving forward with it further.

As for the comp methodology, that sounds pretty straightforward to do. If it's that accurate, that sounds amazing. I'm curious about an approach that pulls those images of the newly renovated homes on redfin, selects the "nicest" 10 of those for each neighborhood of interest, and then of all those, selects the highest quality potential homes.

Something else I'm curious about is identifying "bad remodels". That is, can we say for a home that the cause of its most recent low sale value was that the most recent remodels were bad/ugly/low quality? If so, that seems like a great home to buy and remodel again. I kind of like the first strategy more though.
0 Reply
James Wise
Real Estate Broker · Cleveland Dayton Cincinnati Toledo Columbus & Akron, OH · Member since 2013 · 31k+ posts · 20k+ votes
2y
Quote from @David D.:

I am curious if anyone has employed the simple strategy I was thinking of for investing in any kind of property:

1. Use something like ATTOM API to get historical sales snapshots, or just all the historical sales of the property class for a metropolitan area.
2. Train a regression model using square footage, bathroom count, or more advanced features to predict the price of the sales.
3. Get the error of your predicted housing price (if you're using price per square foot, convert to housing price) down to $10K-$40K or so.
4. Find properties with characteristics that predict they should be selling for 2 standard deviations above what their actual price/value is (or basically just properties that are selling for way below what the model predicts they should sell for).

The challenge here is of course step 3. I just started experimenting with Downtown Detroit with ATTOM's "sales snapshots" and I haven't gotten that kind of performance yet. However, I'm sure many people have, especially if they have images or floor plans of the houses/other features and lots of data.

I apologize if this is all just fairly standard financial modeling. Very key thing here, I don't have a finance/economics background, and am just getting started in this area, but this struck me as a good strategy to use. Also looking at neighborhood trends. My worry is that this wouldn't work because any undervalued property has its price for a "reason", e.g. it is already at its equilibrium price by the time you've determined that its "undervalued", and in fact you won't be able to flip it for whatever reason for a huge profit.

I don't understand how that could be the case though. What seems more intuitive for me is that these are the homes that just happen to have so far been ignored by other investors, or else passed over for better deals, since a small group of investors cannot snatch up all of the deals (otherwise why would this forum exist haha).

Another argument might be that you can never tell when a neighborhood is going to crash in its property values, or when people will migrate to a nearby one that is flourishing. If this were the case, then your model wouldn't be effective anyway. Presumably you can also protect against this with the aforementioned neighborhood growth trends.

Thumbs down on this idea.
1 Reply
Steven S.
Specialist · LA & Ventura · Member since 2023 · 119 posts · 60 votes
2y
Carlos, I see you took a group of comp data and passed it to Ai. If you wanted to repeatedly do this, and your source data is available/downloadable, why wouldn't you just make a spreadsheet or program that would take this list of comps you provided and calculate everything you want from the real source data? Why pass it to an Ai for a half-baked summary when everything it told you (Median & Avg data only) is available on RedFin, with real data, for free? See this https://public.tableau.com/shared/W5RZNC9GC?:display_count=n...

I'm not sure what comps/data you uploaded (looked like all 95111, but the Ai analysis seems wrong/blended). You asked about the 95111 Zip Code, but the $/sf values it gave you seem to be a blend of San Jose itself AND the 95111 Zip Code. The 95111 Zip Code doesn't have a Median $/sf of anything over $750/sf, but the Ai analysis says it does!

And David, I'm impressed a query on Cohere could pull viable sales comps, impressive! However you get the sales comps data, that's great. Get all the data you can, and build some analysis tools that will help you make quick decisions. I wouldn't leave any interpretation to an Ai, these are your dollars you are spending after all!

If you guys are seriously considering Ai for Sales Comps Analysis/Market Analysis/etc, you are using Ai wrong and are either just unaware or cognitively dissonant of the pitfalls. I mean Carlos, you and I have talked about seasonality affecting prices in a repeatable way (buy in Winter, sell in Spring/Summer for ~10% gains), do you see any such consideration from the Ai in it's pricing predictions below?:

- No, just pure linear growth of +$25/mo applied to the wrong $/sf (95111 is ~$700/sf currently, not $900/sf+).

I'm wondering how much you both have actually used Ai to accomplish something that's measurable (like coding projects for example, if they don't run right, you have instant measurable feedback). I've used it on lots of software projects across varying complexity levels, and while yes it will be able to make a simple program or identify an obvious issue, it just can't do anything of real complexity. To non-technical people (no coding ability), it's mind-blowing and magical to get a game of snake coded up for you, or similar toy program.

To me, someone who tries to use it constantly, I am continuously disappointed. It will remove parts of the code that were essential, it will import non-existent libraries and call non-existent functions, and just generally confidently hallucinate something that sounds like it should work. I can recognize this because I actually try and put the code it outputs to use, then I see these disappointing things, and re-evaluate my trust/confidence in these Ai models. Same applies to your Ai market recommendations, except you would have bought a home thinking the resale is $900/sf+ when in reality in 95111 the median is $700/sf at best. You just never know what the heck this thing is doing, until you put it to the test.
1 Reply
Carlos Ptriawan
Member since 2019 · 7k+ posts · 4k+ votes
2y
Quote from @Steven S.:

Carlos, I see you took a group of comp data and passed it to Ai. If you wanted to repeatedly do this, and your source data is available/downloadable, why wouldn't you just make a spreadsheet or program that would take this list of comps you provided and calculate everything you want from the real source data? Why pass it to an Ai for a half-baked summary when everything it told you (Median & Avg data only) is available on RedFin, with real data, for free? See this https://public.tableau.com/shared/W5RZNC9GC?:display_count=n...

I'm not sure what comps/data you uploaded (looked like all 95111, but the Ai analysis seems wrong/blended). You asked about the 95111 Zip Code, but the $/sf values it gave you seem to be a blend of San Jose itself AND the 95111 Zip Code. The 95111 Zip Code doesn't have a Median $/sf of anything over $750/sf, but the Ai analysis says it does!

And David, I'm impressed a query on Cohere could pull viable sales comps, impressive! However you get the sales comps data, that's great. Get all the data you can, and build some analysis tools that will help you make quick decisions. I wouldn't leave any interpretation to an Ai, these are your dollars you are spending after all!

If you guys are seriously considering Ai for Sales Comps Analysis/Market Analysis/etc, you are using Ai wrong and are either just unaware or cognitively dissonant of the pitfalls. I mean Carlos, you and I have talked about seasonality affecting prices in a repeatable way (buy in Winter, sell in Spring/Summer for ~10% gains), do you see any such consideration from the Ai in it's pricing predictions below?:

- No, just pure linear growth of +$25/mo applied to the wrong $/sf (95111 is ~$700/sf currently, not $900/sf+).

I'm wondering how much you both have actually used Ai to accomplish something that's measurable (like coding projects for example, if they don't run right, you have instant measurable feedback). I've used it on lots of software projects across varying complexity levels, and while yes it will be able to make a simple program or identify an obvious issue, it just can't do anything of real complexity. To non-technical people (no coding ability), it's mind-blowing and magical to get a game of snake coded up for you, or similar toy program.

To me, someone who tries to use it constantly, I am continuously disappointed. It will remove parts of the code that were essential, it will import non-existent libraries and call non-existent functions, and just generally confidently hallucinate something that sounds like it should work. I can recognize this because I actually try and put the code it outputs to use, then I see these disappointing things, and re-evaluate my trust/confidence in these Ai models. Same applies to your Ai market recommendations, except you would have bought a home thinking the resale is $900/sf+ when in reality in 95111 the median is $700/sf at best. You just never know what the heck this thing is doing, until you put it to the test.

Steven we know we can always tune in that aspect to make it better. I have been using this predictive modelling even before AI exist just solely using Zillow ZHVI. I am agile enough to expect AI can be better.
I ask again to the AI whether they see pattern:

Based on this excel sheet, what is seasonality pattern that you see ?

Based on the provided real estate data for San Jose, CA 95111, there appears to be a seasonal pattern in home sales activity and prices:

Peak Season (Spring/Summer)
- The data shows a higher number of home sales occurring in the spring and summer months, particularly from March through August.
- Home prices also tend to be higher during this peak selling season, likely due to increased buyer demand.
Off-Peak Season (Fall/Winter)
- There are noticeably fewer home sales recorded in the fall and winter months, from around September through February.
- Home prices also tend to be somewhat lower during this off-peak period compared to the spring/summer months.
This seasonal pattern is quite common in real estate markets, driven by factors such as:
- Warmer weather and longer days make it more convenient for buyers to go house hunting in spring/summer.
- Many families prefer to move during the summer when children are out of school.
- The spring market is fueled by annual bonuses/tax refunds providing down payment funds.
- Sellers often list their homes in spring to take advantage of the higher buyer demand.
However, it's important to note that this is just a general observed trend. There can be significant variations year-to-year and exceptions to the pattern based on local market conditions, economic factors, interest rates, inventory levels, and other variables
I am just saying at the end of the day, the small data set of statistic modelling for real estate would be entirely change to AI even now and future as we don't even need to code anymore. This is just global trend. Like Peter question or Dave question, if you want to find potential 2 standard deviation purchase, buy single family that's double than the DOM of the market and make a visit. Been using 'manual' predictive models just by following ZHVI alone and Redfin sold data in excels.
0 Reply
Carlos Ptriawan
Member since 2019 · 7k+ posts · 4k+ votes
2y
Quote from @David D.:

Quote from @Carlos Ptriawan:

Quote from @David D.:

A Redfin export seems pretty useful, but don't they usually use very aggregated data? I'm not sure which export to use as I didn't find historical sales.
it seems like you dont know how to create the dataset. Here's how to do it:
1. open redfin.com
2. search to buy, enter your query: Single Family 2/1 in zip code 95123 for example
3. Choose sold homes
4. In redfin output choose "Table"
5. Then click download, it would create csv file/excel file
6. once you have it in there you can do so much , you can input the data to your prediction model, create your own chart, tableau , SalesForce CRM or directly put to AI software/models.

These are all free.

The difference between Redfin is the data, Redfin has 24 months data while Privy has 5 years data of last sold homes.
I think Lance's data from Resiclub has more 10 years information.

Finding a two standard deviation homes shall not be too difficult when you combine these data/tools.
0 Reply
Carlos Ptriawan
Member since 2019 · 7k+ posts · 4k+ votes
2y
Quote from @Account Closed:

Quote from @Carlos Ptriawan:

Quote from @David D.:

Quote from @Carlos Ptriawan:

Quote from @David D.:

A Redfin export seems pretty useful, but don't they usually use very aggregated data? I'm not sure which export to use as I didn't find historical sales.
it seems like you dont know how to create the dataset. Here's how to do it:
1. open redfin.com
2. search to buy, enter your query: Single Family 2/1 in zip code 95123 for example
3. Choose sold homes
4. In redfin output choose "Table"
5. Then click download, it would create csv file/excel file
6. once you have it in there you can do so much , you can input the data to your prediction model, create your own chart, tableau , SalesForce CRM or directly put to AI software/models.

These are all free.

The difference between Redfin is the data, Redfin has 24 months data while Privy has 5 years data of last sold homes.
I think Lance's data from Resiclub has more 10 years information.

Finding a two standard deviation homes shall not be too difficult when you combine these data/tools.

"4. In redfin output choose "Table"
5. Then click download, it would create csv file/excel file"

Some we people we joint venture with have found that Redfin does not allow downloadable files in their area. (Just a note to those who read this)

Yes in some market the sold data is not public information, for example in Texas it's not available. Long time ago I tried to find these from TX MLS but not available. Also the data from each MLS is litte bit different from one to another. May need to ask Lance from Resi how he gathers more information in specific market.
In "CA MLS", the "private" remark field data has lot more (hidden) information , we use to be able to read it but they close it again. There are lot of detail to particular house on this field.
So while I can find 2 standard deviation house, usually there's hidden explanation in primate remark why such and such house has lower selling price.
1 Reply
Steven S.
Specialist · LA & Ventura · Member since 2023 · 119 posts · 60 votes
2y
Quote from @Carlos Ptriawan:
Quote from @Steven S.:

Carlos, I see you took a group of comp data and passed it to Ai. If you wanted to repeatedly do this, and your source data is available/downloadable, why wouldn't you just make a spreadsheet or program that would take this list of comps you provided and calculate everything you want from the real source data? Why pass it to an Ai for a half-baked summary when everything it told you (Median & Avg data only) is available on RedFin, with real data, for free? See this https://public.tableau.com/shared/W5RZNC9GC?:display_count=n...

I'm not sure what comps/data you uploaded (looked like all 95111, but the Ai analysis seems wrong/blended). You asked about the 95111 Zip Code, but the $/sf values it gave you seem to be a blend of San Jose itself AND the 95111 Zip Code. The 95111 Zip Code doesn't have a Median $/sf of anything over $750/sf, but the Ai analysis says it does!

And David, I'm impressed a query on Cohere could pull viable sales comps, impressive! However you get the sales comps data, that's great. Get all the data you can, and build some analysis tools that will help you make quick decisions. I wouldn't leave any interpretation to an Ai, these are your dollars you are spending after all!

If you guys are seriously considering Ai for Sales Comps Analysis/Market Analysis/etc, you are using Ai wrong and are either just unaware or cognitively dissonant of the pitfalls. I mean Carlos, you and I have talked about seasonality affecting prices in a repeatable way (buy in Winter, sell in Spring/Summer for ~10% gains), do you see any such consideration from the Ai in it's pricing predictions below?:

- No, just pure linear growth of +$25/mo applied to the wrong $/sf (95111 is ~$700/sf currently, not $900/sf+).

I'm wondering how much you both have actually used Ai to accomplish something that's measurable (like coding projects for example, if they don't run right, you have instant measurable feedback). I've used it on lots of software projects across varying complexity levels, and while yes it will be able to make a simple program or identify an obvious issue, it just can't do anything of real complexity. To non-technical people (no coding ability), it's mind-blowing and magical to get a game of snake coded up for you, or similar toy program.

To me, someone who tries to use it constantly, I am continuously disappointed. It will remove parts of the code that were essential, it will import non-existent libraries and call non-existent functions, and just generally confidently hallucinate something that sounds like it should work. I can recognize this because I actually try and put the code it outputs to use, then I see these disappointing things, and re-evaluate my trust/confidence in these Ai models. Same applies to your Ai market recommendations, except you would have bought a home thinking the resale is $900/sf+ when in reality in 95111 the median is $700/sf at best. You just never know what the heck this thing is doing, until you put it to the test.

Steven we know we can always tune in that aspect to make it better. I have been using this predictive modelling even before AI exist just solely using Zillow ZHVI. I am agile enough to expect AI can be better.
I ask again to the AI whether they see pattern:

Based on this excel sheet, what is seasonality pattern that you see ?

Based on the provided real estate data for San Jose, CA 95111, there appears to be a seasonal pattern in home sales activity and prices:

Peak Season (Spring/Summer)

The data shows a higher number of home sales occurring in the spring and summer months, particularly from March through August.
Home prices also tend to be higher during this peak selling season, likely due to increased buyer demand.

Off-Peak Season (Fall/Winter)

There are noticeably fewer home sales recorded in the fall and winter months, from around September through February.
Home prices also tend to be somewhat lower during this off-peak period compared to the spring/summer months.

This seasonal pattern is quite common in real estate markets, driven by factors such as:

Warmer weather and longer days make it more convenient for buyers to go house hunting in spring/summer.
Many families prefer to move during the summer when children are out of school.
The spring market is fueled by annual bonuses/tax refunds providing down payment funds.
Sellers often list their homes in spring to take advantage of the higher buyer demand.

However, it's important to note that this is just a general observed trend. There can be significant variations year-to-year and exceptions to the pattern based on local market conditions, economic factors, interest rates, inventory levels, and other variables
I am just saying at the end of the day, the small data set of statistic modelling for real estate would be entirely change to AI even now and future as we don't even need to code anymore. This is just global trend. Like Peter question or Dave question, if you want to find potential 2 standard deviation purchase, buy single family that's double than the DOM of the market and make a visit. Been using 'manual' predictive models just by following ZHVI alone and Redfin sold data in excels.
You spoke only to the seasonality (or previous lack thereof), and not the incorrect $/sf metrics which it confidently hallucinated. Of course when you specifically ask 'what seasonality pattern do you see' it will try and detect the pattern, but that's a part of my point. It will feed you BS for any complex analysis/task, and the more specific you are, the more specific and convincing the BS is (this is EXACTLY what I run into on Medium-High complexity code projects).

If it truly understood this seasonality pattern, it would have applied it to the previous uninstructed $/sf estimate for the future. The fact it didn't, and only applies this reasoning when prompted, proves it's not as intelligent as you seem to think. Even after I specifically explain a complex loop in my code, and exactly what should be refactored/adjusted, it is unable to assist. I mean, it will happily assist, but the amount of hallucinations and incorrect code it writes compels me to state it is unable to assist.

If you write up a prompt with a bunch of specific instructions, trying to guide the Ai to weigh everything of consideration, it will surely sound more convincing than if you didn't include those stipulations, but the underlying problem is it has no real understanding of such a complex task and will just give it's best guess/hallucination regardless of how specific you are (as proven by your 95111 zipcode $/sf prompt where it provided you a $/sf much higher than that of the actual 95111 zipcode). This was a basic task of calculating the avg $/sf of closed sales in the 95111 zipcode, and it completely failed, so what makes you think specifying more details will make it more accurate? The underlying architecture of these models makes them best for chatbots & translators (matching word meanings/uses across various languages and translating based on positional/usage probability).

I understand in 10 years there might be an Ai that can do this, but right now, these tools are useless for achieving anything of substantial complexity (meaning anything from medium-large coding projects to properly valuing a home, or even calculating the proper $/sf avg in a zipcode), so I don't understand why you repeatedly advocate for their efficacy in this REI space?
0 Reply
David D.
OP
New to Real Estate · Member since 2024 · 19 posts · 3 votes
2y
Quote from @Carlos Ptriawan:

Quote from @David D.:

Quote from @Carlos Ptriawan:

Quote from @David D.:

A Redfin export seems pretty useful, but don't they usually use very aggregated data? I'm not sure which export to use as I didn't find historical sales.
it seems like you dont know how to create the dataset. Here's how to do it:
1. open redfin.com
2. search to buy, enter your query: Single Family 2/1 in zip code 95123 for example
3. Choose sold homes
4. In redfin output choose "Table"
5. Then click download, it would create csv file/excel file
6. once you have it in there you can do so much , you can input the data to your prediction model, create your own chart, tableau , SalesForce CRM or directly put to AI software/models.

These are all free.

The difference between Redfin is the data, Redfin has 24 months data while Privy has 5 years data of last sold homes.
I think Lance's data from Resiclub has more 10 years information.

Finding a two standard deviation homes shall not be too difficult when you combine these data/tools.

This is very cool, but I'm only getting 119 homes or so for the past 2 years in downtown and midtown Detroit. ATTOM is getting me well over 1000 for the same time period. Is there a setting I'm using wrong possibly?

I would use privy but want to experiment with more open tools for now.
0 Reply
Carlos Ptriawan
Member since 2019 · 7k+ posts · 4k+ votes
2y
Quote from @David D.:

Quote from @Carlos Ptriawan:

Quote from @David D.:

Quote from @Carlos Ptriawan:

Quote from @David D.:

A Redfin export seems pretty useful, but don't they usually use very aggregated data? I'm not sure which export to use as I didn't find historical sales.
it seems like you dont know how to create the dataset. Here's how to do it:
1. open redfin.com
2. search to buy, enter your query: Single Family 2/1 in zip code 95123 for example
3. Choose sold homes
4. In redfin output choose "Table"
5. Then click download, it would create csv file/excel file
6. once you have it in there you can do so much , you can input the data to your prediction model, create your own chart, tableau , SalesForce CRM or directly put to AI software/models.

These are all free.

The difference between Redfin is the data, Redfin has 24 months data while Privy has 5 years data of last sold homes.
I think Lance's data from Resiclub has more 10 years information.

Finding a two standard deviation homes shall not be too difficult when you combine these data/tools.

This is very cool, but I'm only getting 119 homes or so for the past 2 years in downtown and midtown Detroit. ATTOM is getting me well over 1000 for the same time period. Is there a setting I'm using wrong possibly?

I would use privy but want to experiment with more open tools for now.

well just download these 26K sold SF in Detroit,MI for last 5 years then filter the data on excel:
https://www.redfin.com/city/5665/MI/Detroit/filter/sort=lo-d...
privy is good when you want to do more intelligence where there're more "flipping" activity.
0 Reply
Bob S.
Real Estate Consultant · Cleveland · Member since 2020 · 6k+ posts · 3k+ votes
2y
EVERYONE is trying to reinvent the wheel, why???? Run comps and buy it if it makes sense. So many spend months and even years analyzing (WTH are they doing?) and never even do a deal. If you are brand new and never done a deal this is a total waste of time. RE is about one thing, math. just know your numbers. KEEP IT SIMPLE. I have done 100s and 100s and 100s of deals, bought from nearly every avenue there is all withOUT any fancy tools. Any market anywhere in America, I (as many others on BP ) can tell you in under 20 minutes if it's a deal or not. Buy right, reno right, sell/rent right, it's not rocket science.

Good luck
1 Reply
David D.
OP
New to Real Estate · Member since 2024 · 19 posts · 3 votes
2y

well just download these 26K sold SF in Detroit,MI for last 5 years then filter the data on excel:
https://www.redfin.com/city/5665/MI/Detroit/filter/sort=lo-d...
privy is good when you want to do more intelligence where there're more "flipping" activity.
This only gets me 350 or so homes when I hit "download all" (9 pages of rows). Either way though, it seems to be good quality data (my error is quite a bit lower now).
0 Reply
Bob S.
Real Estate Consultant · Cleveland · Member since 2020 · 6k+ posts · 3k+ votes
2y
Quote from @David D.:

well just download these 26K sold SF in Detroit,MI for last 5 years then filter the data on excel:
https://www.redfin.com/city/5665/MI/Detroit/filter/sort=lo-d...
privy is good when you want to do more intelligence where there're more "flipping" activity.
This only gets me 350 or so homes when I hit "download all" (9 pages of rows). Either way though, it seems to be good quality data (my error is quite a bit lower now).

WHY do you need to look at 26k homes let alone 350? WAY to much overthinking and " analyzing"
Just run comps,
0 Reply
Robert Ruschak
Real Estate Broker · IN · Member since 2017 · 70 posts · 12 votes
2y
Quote from @Jonathan Greene:

There are companies with billions of dollars in war chests doing this all the time. There is no reason to think that this will help you invest in real estate. You are basically trying to take the real estate out of real estate.
Predictive analytics are best used as guides to the work. You sound (I could be wrong, but probably not) like you want to use your background to eliminate some of the guesswork, but real estate investing is learned by being at properties, talking with other investors, going to meetups, etc.
Predictive data for investors is wholly flawed because it can't see inside of the houses.

I agree and visiting the properties and talking with real life investors/owners.
1 Reply
David D.
OP
New to Real Estate · Member since 2024 · 19 posts · 3 votes
2y
Quote from @Bob S.:

Quote from @David D.:

well just download these 26K sold SF in Detroit,MI for last 5 years then filter the data on excel:
https://www.redfin.com/city/5665/MI/Detroit/filter/sort=lo-d...
privy is good when you want to do more intelligence where there're more "flipping" activity.
This only gets me 350 or so homes when I hit "download all" (9 pages of rows). Either way though, it seems to be good quality data (my error is quite a bit lower now).

WHY do you need to look at 26k homes let alone 350? WAY to much overthinking and " analyzing"
Just run comps,

What do you mean by running comps? From the sources I've seen, it involves looking for 20-30 or so similar homes to the major MLS listed features of your home, then taking the average price per square feet and using that as your target "market price per square foot". I can't see how that could possibly be a better estimate than just fitting a line using all the features and several hundred properties.

Then again, Steven's tool sounds highly accurate, so there must be something to this technique. Is there a more involved way to do it?
0 Reply
Bob S.
Real Estate Consultant · Cleveland · Member since 2020 · 6k+ posts · 3k+ votes
2y
Quote from @David D.:

Quote from @Bob S.:

Quote from @David D.:

well just download these 26K sold SF in Detroit,MI for last 5 years then filter the data on excel:
https://www.redfin.com/city/5665/MI/Detroit/filter/sort=lo-d...
privy is good when you want to do more intelligence where there're more "flipping" activity.
This only gets me 350 or so homes when I hit "download all" (9 pages of rows). Either way though, it seems to be good quality data (my error is quite a bit lower now).

WHY do you need to look at 26k homes let alone 350? WAY to much overthinking and " analyzing"
Just run comps,

What do you mean by running comps? From the sources I've seen, it involves looking for 20-30 or so similar homes to the major MLS listed features of your home, then taking the average price per square feet and using that as your target "market price per square foot". I can't see how that could possibly be a better estimate than just fitting a line using all the features and several hundred properties.

Then again, Steven's tool sounds highly accurate, so there must be something to this technique. Is there a more involved way to do it?

you are making this much more complicated than it needs to be. You do not need to run 20- 30 closed sales to get a comp,
0 Reply
Steven S.
Specialist · LA & Ventura · Member since 2023 · 119 posts · 60 votes
2y

If you are just looking for your next project to buy, I agree with Bob. If you want to build a tool to help you identify opportunity faster, you have to basically take those manual underwriting skills and codify them somehow.

My program's approach is basically my manual approach, it just doesn't have any photos/videos to work with. Thankfully, in an active market like the SFV & West-LA, there are so many transactions of high & low end property that you actually don't need photos/videos, the market values it as it should be upon sale, and my program simply reverses that logic.

If you want to learn exactly how I extract the comps from the larger dataset and calculate the ARV/ACV values based on that comp data, checkout my older post here & go through the comments: https://www.biggerpockets.com/forums/48/topics/1130680-i-mad...
-This should give you a great idea of how to implement some logic to analyze your own transaction-level data.

It is accurate, because the market is reliable at valuing homes on a $/sf basis. Here's another recently closed, renovated home run through my program with the query "19707 Corbin Ln, Winnetka, CA 91306 5/4/2350-3000,0.65,255":

2-4 good, recent comps are all you need to properly estimate the ARV of your project, definitely not 20-30.

Here's how I manually underwrite a flip if the numbers look promising (note the number of comps used, the $/sf avg of the comps, the $/sf avg of the green 'High' percentile graph, and the resulting $/sf of my estimated Sale Price):

0 Reply
Dean Romero
Member since 2025 · 7 posts · 0 votes
1y
Hi guys, I have been following your conversation and I was wondering if you have ever put together a model (either MLS or Redfin) that looks at homes that have been sold/re-sold within an 18mo period? I am looking for properties that have been flipped and attempting to gauge the original vs. flipped price and bounce that against the interior changes made as part of the rehab.
Thanks!
0 Reply
- Steven S.
  Specialist · LA & Ventura · Member since 2023 · 119 posts · 60 votes
  1y
  Quote from @Dean Romero:
  
  Hi guys, I have been following your conversation and I was wondering if you have ever put together a model (either MLS or Redfin) that looks at homes that have been sold/re-sold within an 18mo period? I am looking for properties that have been flipped and attempting to gauge the original vs. flipped price and bounce that against the interior changes made as part of the rehab.
  Thanks!
  
  Finding homes flipped in the last 18 months would yield developers/investors in the area (potential buyers for your under contract or owned distressed property). It's not useful as a developer trying to value things (just run comps), it's most useful to a wholesaler trying to find buyers for their 'locked-up' properties. I think PropStream has a 'recently flipped' toggle if that's what you're after.
  
  Whether you look at flips from developers or newly renovated properties from owners, the source of the flipped/renovated house doesn't matter for determining value. Just take closely matching renovated comps, and closely matching as-is comps, to make sure you are buying for the right price, and are estimating dispo at or just below the other top comps.
  
  No 'model' will adjust value based on interior changes automatically, you'll have to draw up your own model to check/account for these things, but at the end of the day it's a lot more efficient to just run comps and put closely matching, similar comps into your spreadsheet model like I do above. Then just build to match the top-comps quality/finishes/appliances, etc.
  0 Reply
Dean Romero
Member since 2025 · 7 posts · 0 votes
1y
Thanks Steven!
0 Reply

Join the conversationCreate a free account to reply, vote on answers and follow this thread.