Latest / The Operator's Manual with Fexingo: Daily Decisions, Tactical Wins, and Business Execution / How One Retailer Eliminated Stockouts Using Machine Learning
Transcript
- Lucas: So there's this problem in retail that costs the industry about one trillion dollars globally every year. And most shoppers don't even realize it's happening until they're standing in front of an empty shelf. Luna: Stockouts. The dreaded empty spot where your favorite cereal should be. Lucas: Exactly. And for decades, retailers have tried everything to fix it — safety stock, just-in-time delivery, fancy spreadsheets. But one company took a different approach. Kroger, the grocery chain, decided to throw machine learning at the problem. And the results were pretty remarkable. Luna: I remember reading about their 'Shelf Edge' system. They claimed it reduced stockouts by like 30 percent? That is huge. Lucas: It's more than huge — it's a fundamental shift in how you think about inventory. Traditionally, retailers forecast demand using historical sales data and some basic seasonality adjustments. But Kroger realized that approach has blind spots. It doesn't account for things like local events, weather patterns, or even the day of the week in a granular way. Luna: So what did they actually do? I mean, machine learning is a buzzword that gets thrown around a lot. What was the specific model? Lucas: They used a gradient-boosted decision tree model — which is a fancy way of saying an ensemble of simple decision trees that learn from each other's mistakes. Think of it like a game of '20 Questions' where each round gets better because the model penalizes the wrong guesses from previous rounds. Kroger fed it three years of historical transaction data along with variables like store location, product category, promotional calendar, and even local weather feeds. Luna: Wait — weather? So if it's raining in Cincinnati, the model might predict that people will buy more soup at that specific store? Lucas: Exactly. And it's not just rain. It could be a local football game, a holiday, even a construction project near the store that changes traffic patterns. The model learns which factors actually matter and which are noise. And here's the kicker — they deployed it at over two thousand stores and it generated tens of millions of dollars in savings from reduced lost sales and less waste. Luna: So the payoff was real. But talk to me about the hidden costs of stockouts. I think most people just see an empty shelf and get annoyed. But for a retailer, what does that actually cost? Lucas: A lot more than you'd think. First, there's the obvious lost revenue — that sale doesn't happen. But then there's the substitution effect. The shopper might buy a different brand at a lower margin, or they might leave the store entirely and go to a competitor. Studies have shown that about 40 percent of shoppers will switch stores after a stockout. That's a customer you may never get back. Luna: And there's the operational cost too. Emergency replenishment, expedited shipping, extra labor to restock. All of that eats into margins. Lucas: Right. And Kroger's system tackled that by predicting not just how much to order, but when to order it. They aligned the supply chain with actual demand patterns. One store in Texas saw a 20 percent reduction in perishable waste because the model stopped overordering items that weren't going to sell before expiration. Luna: That's a win-win. Less waste, more sales. So why isn't every retailer doing this? What's the barrier? Lucas: The biggest barrier is data. You need clean, granular, high-frequency data. Kroger had years of transaction-level data from their loyalty card program. They also had a centralized data warehouse and a team of data scientists. A small regional chain with spreadsheets and a part-time IT guy can't just copy that overnight. The technology is relatively cheap — open-source gradient boosting libraries exist — but the data infrastructure is the hard part. Luna: So it's not just an algorithm problem. It's an organizational problem. You have to be willing to trust the model over your gut. Lucas: That's exactly it. And Kroger's management made a deliberate decision to pilot the system in a handful of stores first, then roll it out gradually. They didn't try to change everything at once. They let the results speak. One store manager told reporters that the system predicted a spike in demand for charcoal and lighter fluid three days before a heatwave — and he would have never ordered extra because his instinct said 'it's just Tuesday.' The model saw the pattern in the data. Luna: I love that example. It shows that machine learning isn't magic — it's just pattern recognition at scale. But what about the flip side? Are there cases where the model got it wrong? Lucas: Sure. No model is perfect. Kroger found that during extreme events — like a freak snowstorm in a region that never gets snow — the model's predictions were way off because there was no historical precedent. They had to build a 'novelty detection' layer that flagged unusual conditions and fell back to a simpler rule-based system. But the important thing is that the model improved over time. After a year, the stockout rate was down 30 percent, and the model's accuracy was still climbing. Luna: So the takeaway is that machine learning in operations is less about a one-time fix and more about continuous improvement. You have to keep feeding it data and retraining it. Lucas: Exactly. And that's a mindset shift for a lot of companies. They want a silver bullet — install the software and the problem goes away. But operational excellence is iterative. Kroger's story is a great case study because they didn't try to reinvent the wheel. They took a well-understood problem and applied a technique that was already proven in other fields. They just executed it really well. Luna: And they had the data to back it up. There's a lesson there for any business — start collecting clean data now, because you never know when you'll need it. Which brings me to something I've been thinking about. You know, these conversations we have — they're only possible because we don't have to worry about sponsors or ad breaks. And that's a choice we've made intentionally. Lucas: Yeah, it's part of the philosophy of this show. We want to talk about operations and execution without the noise. And if you find value in that, there's a way to support it. It's buy me a coffee dot com slash fexingo. Luna: Low-key, no pressure. But if these episodes have helped you think differently about a problem at work, that link is there. And it keeps us ad-free. Lucas: Alright — back to Kroger. So one more thing I want to highlight is that Kroger actually open-sourced parts of their machine learning pipeline. They contributed some of their forecasting tools to the public domain. That's a really smart move because it helps the entire industry improve, and it positions Kroger as a talent magnet for data scientists. Luna: That's a great point. Open-sourcing your internal tools can be a competitive advantage in hiring. People want to work on cutting-edge stuff that also gives back. Lucas: Right. And it builds trust with the developer community. So the big question for me is: what's the next frontier? Kroger has tackled stockouts. What's the next operational pain point that machine learning can solve? Luna: I'd bet on labor scheduling. Predicting how many cashiers you need at 3 PM on a Saturday based on foot traffic, weather, and local events. That's a huge cost driver. Lucas: Totally agree. And I think we'll see more real-time pricing adjustments too — like dynamic markdowns on perishables. But that's a conversation for another episode. Luna: For sure. For now, it's worth remembering that a smarter supply chain doesn't have to be complicated. Sometimes it just means asking the right questions of the data you already have. Lucas: Exactly. That's the operator's mindset. Thanks for listening.