All Forum Posts by: David D.

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? 

 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. 

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. 

I don't see a key. Let me send you a message with a working notebook. Just replace the part with "YOUR APIFY KEY HERE" with your actual apify key. 

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

For data size, I'm not using a particularly large amount of transactions so far, and am getting pretty good, but still imperfect valuations of Detroit homes. Nowhere near your error rate of only 4% if that's representative of a typical error in an estimate. If you're doing that using Comparative Market Analysis, then there must be something to the technique surely. I just don't really understand how it works, because it seems to me (intuitively) like it would only introduce a lot of error into the model. I should probably read a book about it realistically!

But yeah to summarize again, the goal here is to take a relatively standard $500,000 home, for instance, which is currently priced at $400,000, and sell it for what it's "actually worth". The main pushback on that idea from some others in the thread I sense is that there's a reason that the home would be priced so much lower, but the idea is to look at homes that don't have those defects.

In my downtown Detroit dataset, for example, the average difference between predicted and actual price of the top 100 "undervalued" deals is $118,000. The idea is that you choose 10 of these randomly, and just visit them. The chances that all of them have some $100,000 list of defects is presumably very low, especially if that $100,000 list of defects is undetectable by an expert. Or put another way, I haven't seen any explanation of how that could be that makes sense to me. 

Carlos mentions that you can also exploit seasonality to get comparable profits like this, and recommends some ranges to seek remodels in (thank you for sharing those!). Again, very curious to try a comp analysis informed by Gen AI, but perhaps not going to be the first thing I end up actually trying in practice haha

These are two different but related strategies it sounds like. Steven is using an ARV model that's homegrown, and ChatGPT is using Redfin's ARV model, probably with its own intuitions about repair values baked in there too.

These are still different from my approach though, I think? I'm assuming Steven is also using a small dataset of Comps. I've been skeptical of Comps since first hearing about them. If there's only a few comps you're using, there's naturally going to be a lot of error in your valuation model, so you're at greater risk of buying a home that really isn't going to net you much of anything. 

That's why I like this traditional ML approach, we leverage lots of data to inform our decision about which house is seriously "undervalued"

I have a couple of questions about both of your approaches though, will be interesting to compare them!

1. In Steven's approach, is the strategy to choose a house that has a high ARV but low current price, or to just select based on the ARV? 
2. Have both of you found that these approaches are accurate in your own investments?
3. Do you think an approach with floorplans (which I just found out Redfin has), would be more effective? 
4. I'm curious about whether you've tried to model the effects of remodeling alone. For instance, let's say I remodel the home completely, have a completely different look inside of it than it had prior, as opposed to the smallest possible remodel. There is presumably a sweet spot in between that has the best "value" on average. Have you explored this? 

 The code already does this. To get the property ID from the urls the code generates, you can set up a new column in excel that splits the url to get the last part (the property id). Or you can keep the url like it is and use power query to pull data from that webpage on the property. 

Yeah, that's just a one liner like this:

url = https://www.redfin.com/TX/Fort-Worth/10725-Lone-Pine-Ln-7610...

the_property_id = url.split("/")[-1]

 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. 

These other aspects of the strategy seem great, really reduces the chances the neighborhood is just going to tank in value or something I would think, though I suspect fitting a curve as opposed to just using the mean may be a bit more effective and less risky. I'll try it out! Would you feel comfortable sharing your approximate "loss rate"? That is, the rough percentage of the investments you make that end up costing you?