Who's Innovating Without Adding AI?

[Carsten]

This is interesting, several people asking the same thing.

Do the OEMs just send you new stock and you go with it?
What about used cars? Just buying what is a good deal?

Who does the sourcing for stock for your lots?

 

[jon.berna]

I have been reading and thinking about this and one of the challenges I see is using different data/metrics for the wrong things. If you have time series sales data that will almost certainly be the best for predicting future sales. Any metrics that are upstream of that will lack the same value and correlate to sales anyways with just more noise. Sure there are 100 other ways to do this to seek a specific question about an audience or the performance of a geo, or a campaign but past sales is the best data point for future, that gets you like 70% of the way there.

The next 10-15% is real time data that shows pipeline devoid of human error, the remaining 15-20% is what determines your accuracy and this is based on 100s of things you'll go crazy trying to map, collect and clean. Most people never get the 70% right and focus on the remaining 30% which is 10x harder.

I recommend to just start forecasting model-level sales 1-3 months out and track your MAPE, keep improving your model and learn all the different approaches. If you want post them up in a new thread how you built it.

 

[Jon Singo]

The problem with any sort of advanced vehicle sourcing system for used, at least right now, is there's no inventory to source. All the data I collect and use to try and make better purchasing decisions leads me to the same conclusion everyone else has. Need car, everyone needs car, only 10 of car, and 100 dealers need car. It's a land grab and it's only going to get worse as 2025 lease turn ins are projected to be down again.

In a perfect world where the lanes are filled with dealer trades and fleet vehicles I could get behind a product to do this. Hell, Carmax has had something similar for years and I've considered creating a version for dealers on multiple occasions.

You might be on to something if you could use this algorithm to assist with service and private party acquisition based on your personalized need/demand/etc. As it sits, the auctions are a seller's game and that isn't going to change anytime soon.

Ultimately, success is going to look like being able to buy anything and everything while learning how to make a buck via retail, wholesale, or fleet. Going to have to start getting creative.

 

[jon.berna]

A lot of the prior conversation was about model-level new sales. Totally agree on with you on used being volatile based on lack of supply. I would say even if there's a lack of supply as you know the further your perspective the more predictable used gets. But yeah 1 dealer in 1 month it's pretty challenging. As we used to, I still like the approach of using the first 72 hours of engagement data after a car goes live, the dealers prior history and the market's history to smooth out.

This approach is similar to how you can use early player stats in baseball to determine reference players with similar trajectories. The first 72 hours shows you the early career, the dealer and market history give you the comps. If you track time series engagement data you can do this. For the nerds: Career Trajectories

 

[Ryan Everson]

Good idea, difficult to execute. You would need some sort of reference to the market dynamics you were dealt at that timestamp, a lot of what @jscole86 mentioned. Hurts my head thinking about it.

I learned just yesterday that companies exist just to do dealer stock ordering and my first thought was "okay... but what are they getting access to and looking at to even make that decision"? What does this service cost?

Sort of did this a while ago for some dealers running our custom order tool to help understand what trims, engines, colors being selected. But we were all in a vacuum at the time.

Not to mention taking into account the ever changing list of constraints regarding what you can order.