Adventures in Demand Analysis Using AI

0:00Welcome to the Deep Dive. Today we're jumping into a really fascinating study. It brings together cutting edge AI and that age old puzzle. consumer behavior. Our mission today is to explore how a new approach, basically using AI to really understand product details, is, well, shaking up how we measure price elasticity, which is just how much our buying habits shift when prices change. Exactly. And economists, I mean, for almost 100 years now, going way back to people like Wright and Schultz, and they've been trying to quantify this. Yeah. But the old methods, they were pretty much stuck with simple numbers.

0:36What we're seeing now, though, it's a huge leap. We can capture these subtle qualities in products we just couldn't measure before. It gives us a much, much richer feel for how markets actually work. So this deep dive is kind of a shortcut to understanding this mix of old school economics and new AI. That's a great way to put it, yeah. Econometrics meets machine learning. Okay, let's unpack that. So traditionally, understanding product features meant what? Basic numbers, categories, maybe size, things someone coded by hand? Pretty much. Limited data points. But think about, say, a toy car. Its appeal isn't just the price tag or the brand, right?

1:14It's the look, the shiny finish, maybe the story in the description. How does AI even start to grasp that? Right. And that's exactly where the AI comes in. The key idea here is what they call multimodal product representations. Multimodal meaning? Meaning more than one type of data. So instead of just like the subcategory in a spreadsheet, we're feeding the AI models the full text description and crucially the pictures too. Ah, the images. Yeah. Think of that bright blue SAS BSC Ram 1500 toy truck from the sources. The AI looks at the picture and reads the words. Okay, so it sees the monster truck, reads about it, and somehow turns all that fuzzy subjective stuff into like hard numbers.

1:56How does that work? Well, that's where these things called embeddings come in. You can sort of picture them as dense numerical vectors. And you're saying, there's like coordinates. Exactly like coordinates, but in a really high dimensional space. And these coordinates capture incredibly subtle things, perceived quality, branding vibe, visual details, stuff that, you know, ticking boxes often misses entirely. Wow. And the study uses these really powerful AI models, transformer models, things like Roberta or Llama 3 for the text and BIT for the images. They're specifically built to understand language and visuals deeply.

2:33Okay, that's a huge step. But how does the AI learn these representations? Does someone have to sit there and label millions of toy cars? This one looks high quality. This description is exciting. That sounds impossible. It would be impossible or at least incredibly expensive and slow. That's the really clever part. It uses self-supervised learning. Self-supervised. It teaches itself. Kind of, yeah. Imagine you give the AI a sentence like, well-made masked model truck with masked body. The AI learns by trying to guess the missing words, the masked words from the context. Ah, like fill in the blanks.

3:04Exactly. Same idea for images you hide a patch. The AI tries to predict what's missing. By doing this over and over on massive amounts of unlabeled data, just text and images scraped from the web, It figures out the important features and relationships all on its own. So it learns the patterns without needing human labels. Precisely. Yeah. And that creates these incredibly informative embeddings we were talking about. Okay, so to actually test this whole idea, the researchers needed data, a lot of data. They did. And they found a great data set, nearly 10 ,000 different toy cars sold on Amazon.com.

3:38Collected over, was it nine months in 2023? Yeah, April to December. and they had everything sales ranks prices those text descriptions of pictures the whole package a really rich data set and they had to standardize things of course for quantity how much stuff was selling they used a log transform of the inverse sales rank logarithm of the inverse okay why that it basically turns popularity the sales rank into a smoother measure of sales volume high rank means high sales but the log smooths out the really extreme best sellers makes it easier to model Got it. And price. Similar idea logarithm of the average price over time.

4:13Again, it helps stabilize the data and reveal patterns more clearly. They looked at everything in four-week chunks. Okay, makes sense. So the big question, did these fancy AI embeddings actually do anything useful? Did the AI really get these toy cars? How did they even check that? Good question. They tackled it in two ways. First, qualitatively just looking at the patterns. They visualized the products based on these embeddings. Like plotting them on a chart. Exactly. And when they used both text and image embeddings together, the products formed this really clear full ball shape in 3D with distinct clusters popping out.

4:50Clusters, like little neighborhoods of similar toy cars grouping together automatically. Precisely. You could literally see groups emerge, like premium detail replicas of emergency vehicles or iconic movie die cast cars. Really specific stuff. That's cool. And here's the kicker. when they took out the image data and only used text, the whole structure collapsed into more like a stripe on a sphere. Interesting. So the images were adding unique information. Yeah. Definitely. Information the text alone just couldn't capture. They even used generative AI to describe these clusters and make average pictures for each one, and it matched what humans thought the clusters represented.

5:27Pretty neat validation. Okay, that's compelling visually. But what about the numbers? Quantitatively, did the embeddings help predict sales or prices better? Oh, absolutely. The basic data, the tabular stuff alone, it was pretty poor at predicting sales levels. Right. But add in the AI text embeddings. Predictive accuracy jumped significantly. Then add the image embeddings on top of that. It boosted the R squared, that's a measure of explanatory power for quantity prediction, from around 45 % to over 61%. Wow, that's a big jump, from 45 % to over 60%. It's a huge improvement. It clearly shows these AI representations aren't just cool visualizations.

6:04They really add meaningful power to understanding demand. Okay, so this leads us to the main event, right? Estimating that price elasticity of demand, how much sales change when price changes. Why is that usually so tricky? You hear about companies getting pricing wrong all the time. Yeah, it's notoriously difficult. Because simple statistical methods, just running a basic regression of sales on price, they often give you estimates that are just implausibly small. Small meaning. Meaning they suggest prices barely affect sales. Like a 1 % price hike might only reduce sales by, say, 0.1%. Which doesn't feel right.

6:40It doesn't make economic sense. Yeah. If that were true, companies could just keep hiking prices indefinitely, right? Yeah. So it tells us there must be other things going on, confounding factors. Things like maybe the product's quality is improving at the same time its price goes up, or maybe it's getting more advertising visibility. These things mess up the simple calculation, making it hard to isolate the true effect of just the price change itself. Okay, so confounding factors hide the real impact. How did this study try to cut through that noise and get a, well, a truer number for elasticity?

7:11They built a more sophisticated model, what they call a causal dynamic model. It's designed specifically to disentangle these effects. Causal meaning it tries to find the actual cause and effect of price. Exactly. It carefully accounts for things like the product's past sales, what they call lagged quantity in its past prices, alongside those rich AI embeddings. It tries to control for all those potential confounders. Okay. And they used a fancy statistical technique, too. Stuff's called double machine learning or orthogonal inference. It's basically designed to give you robust, unbiased estimates of causal effects, even when you're using complex AI models like these embedding generators.

7:48It helps make sure the AI isn't accidentally introducing new biases. Right. Using AI to avoid AI biases in a way. Sort of, yeah. Using sophisticated methods to control for the complexity. So applying this more robust causal model, what did they find for the average toy car? How sensitive is it to price? Well, first they looked at what's called a homogeneous model, assuming the effect is roughly the same for all toy cars. Okay, the average effect. Right. And they found a clear negative effect. And it was statistically significant. A 1 % increase in price led to roughly a 0.45 % to 0.6 % drop in that quantity signal we talked about.

8:25Okay. A 0.45 % to 0.6 % drop in the signal. What does that translate to in terms of actual demand elasticity? It translates to an elasticity somewhere in the range of, I mean, it is 0.9 to, I mean, it is 1.2. The minus 0.9 to minus 1.2. Is that more plausible? Much more plausible, economically speaking. It suggests demand is sensitive to price, much more than those simpler models often find. Interestingly, while things like past sales and past prices were important confounders they needed to control for, Yeah. The AI embeddings themselves, they weren't strong confounders. They didn't seem to predict price changes directly.

9:01They were more like crucial descriptors of the product rather than factors driving the price changes themselves. OK, that's a subtle but important distinction. The embeddings describe the what, not necessarily the why of price changes. Precisely. So we have a more believable average elasticity now. But this is where the AI really flexes its muscles, right? Moving beyond just one average number. Is every toy car the same? Absolutely not. Yeah. And that, for me, is the most profound finding of the study. This idea of heterogeneous elasticity. Heterogeneous, meaning different. Exactly. Different elasticities for different products.

9:35The AI embeddings, especially when used to group products into those clusters we discussed, revealed that price sensitivity varies dramatically across different types of toy cars. It's not one size fits all at all. Wow. OK, how dramatically? Give me a sense of the variation. Well, for example, they found higher price toys and ones with better sales ranks tended to be more price sensitive, maybe because there are more alternatives or buyers are more price conscious at higher price points. OK, that makes sense. And those AI defined clusters. Products in different clusters showed really different sensitivity levels.

10:07Some clusters had strong negative elasticities. Price mattered a lot. Others were much less affected, suggesting different buying motivations for those types of toys. So instead of that single average number, like making a 1.1 or whatever, what's the full range that AI uncovered? Okay, get this. The study suggests the actual demand elasticity for individual toy cars could range from nearly zero. Zero. Meaning price barely matters for some. Essentially, yeah. Maybe a specific collectible or a must-have movie tie-in for some buyers. Yeah. Price isn't the main driver. All the way to approximately, I guess, 2.8.

10:42Minus 2.8, so a 1 % price hike causes a nearly 3 % drop in demand for those. Exactly. A huge sensitivity. The average across all these different products landed around Negus 2.1, which is also quite sensitive. That range from 0 to Negus 2.8 is massive. It is. And it's statistically significant. It fundamentally changes the picture. It shows these AI embeddings aren't just descriptors. They act as crucial effect modifiers. They tell you how much a price change is likely to matter for which specific kind of product. What really strikes you about seeing a range like that? What strikes me is just how much more useful that is than an average.

11:18An average hides all that variation. This gives you a much sharper tool. But, okay, you always have to ask, even with amazing studies like this, what are the limitations? What assumptions did they have to make? Right. Good point. Every study has them. A key assumption here, for the causal estimates especially, is that there aren't any sort of hidden time-varying demand shocks, meaning things that pop up and affect both price and quantity at the same time, that the model doesn't know about. Like, maybe a sudden viral trend for a specific type of toy car makes demand spike and sellers raise prices simultaneously.

11:53If the model doesn't capture that external shock, it can still slightly misinterpret the price effect. Okay. Now, the researchers actually think this might be less of a problem in this specific data set because they observed that Amazon toy prices often seem quite sticky. Sticky. Yeah, like they stay flat for a while and then jump piecewise constant paths, they called it. They don't seem to respond instantly to every little flicker in demand. That stickiness actually makes the causal estimation a bit easier. Ah, okay. So maybe less of a worry here than in some other markets. Possibly. Right. But it's still an underlying assumption to be aware of.

12:30So even with that caveat, racking this up, this deep dive really shows us a new frontier, doesn't it? Absolutely. by pulling together text images in the usual tabular data and feeding it all into this smart causal framework. These AI embeddings do more than just predict better. That's great. But the real payoff is getting much more believable, much more nuanced estimates of price elasticity, seeing how it differs so much across products. It's a bridge between machine learning and econometrics. It really is a powerful bridge, opening the door to much richer, more granular insights into how markets actually behave.

13:07It's been an incredible journey seeing how AI helps us get under the herd of consumer behavior, not just predicting what sells, but understanding why it sells at a certain price, and crucially, how that reason changes from product to product. And think about the practical side for businesses. You know, instead of one blunt pricing rule for everything, imagine having tailored, data-driven strategies for each product. Knowing how much wiggle room you have, we, our customers, will react strongly. That's powerful. Definitely. And for you, our listener, it's really a glimpse into a future where economic understanding isn't just about averages anymore.

13:41It's built on this incredibly rich, detailed view of the world, all powered by AI. So maybe something to think about. How might these heterogeneous effects, these different sensitivities play out in industries you follow beyond toy cars, fashion, gadgets, streaming services? Something to chew on until our next deep dive.