Taming Intermittent Demand Forecasting With AI

A Turkish automotive spare-parts case study shows how intermittent and lumpy demand can be tamed with AI. We compare the old cross-method approach with exponential smoothing to an ensemble of models, including RNNs, and a linear-regression meta-learner that blends their forecasts. The result: dramatically reduced inventory costs and fewer shortages, offering a glimpse into a future of anticipatory logistics.

Note:  This podcast was AI-generated, and sometimes AI can make mistakes.  Please double-check any critical information.

Sponsored by Embersilk LLC

Transcript

SPEAKER_00 0:00

Um, so my car recently decided, with like absolutely zero warning, that it desperately needed a new mass airflow sensor bracket.

SPEAKER_01 0:09

Oh no, that is a very specific, obscure part.

SPEAKER_00 0:12

Right. It's this dusty shelf piece of plastic. And honestly, trying to predict when a vehicle will suddenly demand something that rare, well, it feels exactly like trying to predict my toddler's suddenly shifting favorite food.

SPEAKER_01 0:25

Oh, I know that game.

SPEAKER_00 0:26

Yeah. Like Monday, it's only green grapes, and Tuesday, green grapes are practically poison. It is just pure, unpredictable chaos.

SPEAKER_01 0:34

Aaron Powell And you know, that exact chaos is actually a massive, incredibly expensive puzzle for the whole automotive industry.

SPEAKER_00 0:40

Aaron Powell Because they have to keep millions of these obscure parts in stock, right? Just so you aren't stranded.

SPEAKER_01 0:45

Exactly. But they can't go bankrupt storing them in giant warehouses just waiting for a random Tuesday when someone finally needs a bracket.

SPEAKER_00 0:51

Aaron Powell Which brings us to what we are exploring today. We have this really fascinating case study for a deep dive coming from the Turkish automotive spare parts industry. Trevor Burrus, Jr.

SPEAKER_01 0:59

Yeah, and they essentially cracked this problem.

SPEAKER_00 1:01

Aaron Powell They did. They used AI to turn absolute unpredictability into a perfectly tuned system.

SPEAKER_01 1:07

Aaron Powell But to um to really understand what this AI accomplished, you kind of have to look at the baseline difficulty of spare parts. It comes down to two things intermittent and lumpy demand. Aaron Powell Okay.

SPEAKER_00 1:19

Intermittent meaning what, long stretches with zero sales?

SPEAKER_01 1:22

Aaron Powell Right, months at a time where a part just sits on a shelf collecting dust.

SPEAKER_00 1:26

Okay. And then lumpy demand.

SPEAKER_01 1:28

That ticks it up a notch. Lumpy demand adds sudden, wild fluctuations in the actual quantity a customer needs when they finally do buy it.

SPEAKER_00 1:36

Aaron Powell Wow. Okay, so how on earth did they handle this before AI?

SPEAKER_01 1:40

Aaron Powell Well, for decades, the industry standard was crossed fix method. It basically separates the size of the demand from the time interval between demands, and it uses exponential smoothing, which is essentially just a weighted average to guess when the next order will happen.

SPEAKER_00 1:54

Aaron Powell Wait, so isn't relying on averages like uh trying to predict when lightning will strike just by averaging the days between past storms?

SPEAKER_01 2:01

That is a perfect analogy, actually.

SPEAKER_00 2:03

Aaron Powell Because averages don't capture sudden violent spikes. Right. If an obscure bracket hasn't sold in six months, averaging past data won't tell you it's suddenly going to sell tomorrow.

SPEAKER_01 2:14

It completely fails on extreme variability. Averages leave warehouses totally empty during a sudden spike or massively overstock during long lulls.

SPEAKER_00 2:24

Just burning money.

SPEAKER_01 2:25

Exactly. The math just couldn't handle the reality of human behavior and mechanical failure. So researchers turn to artificial intelligence.

SPEAKER_00 2:34

Which makes perfect sense because AI is built to spot those invisible patterns that simple averages miss. And uh actually, real quick, if you're looking to solve your own complex prediction problems, you should really check out Embersilk.com.

SPEAKER_01 2:46

Oh, yeah, they are fantastic for that.

SPEAKER_00 2:48

Right. Whether you need help with AI training, automation, integration, or software development, Embersilk helps you uncover exactly where AI agents can make the most impact for your business or even your personal life.

SPEAKER_01 3:00

Absolutely. And finding that impact is exactly what the researchers in Turkey did. They tested individual machine learning models against deep learning models just to see what could actually spot these weird demand patterns.

SPEAKER_00 3:12

Aaron Powell And what did they find? They used um RNNs, right?

SPEAKER_01 3:15

Aaron Powell Yes, recurrent neural networks. These deep learning models have a sort of memory loop. It helps them learn complex sequential patterns over time rather than just blindly crunching historical averages.

SPEAKER_00 3:27

Aaron Powell But the real breakthrough in this study wasn't just finding one perfect AI, was it? It was about combining them.

SPEAKER_01 3:34

Aaron Powell Right. The researchers utilized the stacking ensemble learning method.

SPEAKER_00 3:38

Aaron Powell Okay, stacking ensemble. What does that actually look like in practice?

SPEAKER_01 3:41

Aaron Powell Well, instead of relying on a single algorithm, they took the predictions from multiple base models and fed all those guesses into a meta learner.

SPEAKER_00 3:50

A meta learner.

SPEAKER_01 3:51

Yeah, in this case, a linear regression model.

SPEAKER_00 3:52

Aaron Powell Okay, let's unpack how that works. Because instead of just picking the smartest AI in the room, it sounds um well, it sounds like assembling a corporate board of directors.

SPEAKER_01 4:01

Oh, I like that. Go on.

SPEAKER_00 4:02

Aaron Powell So you have different experts, right? A finance guru, a marketing head, an operations chief. Those are your base AIs. Then the CEO, your meta learner, listens to all of their forecasts.

SPEAKER_01 4:12

Yes. And the CEO knows the historical track record of each director. The linear regression model mathematically assigns a weight to each AI's guess based on its past accuracy.

SPEAKER_00 4:22

Oh wow. So if the deep learning model is historically more accurate during a sudden spike, the meta learner trusts it more in that specific scenario.

SPEAKER_01 4:31

It acts as a judge, blending the advice to make one final, incredibly precise prediction.

SPEAKER_00 4:37

By doing that, it covers the blind spots of any single algorithm. That is brilliant.

SPEAKER_01 4:41

It really is. It wiped out the forecasting errors that plagued the old mathematical models. By successfully anticipating that lumpy demand, they cut their inventory costs and effectively eliminated operational shortages entirely.

SPEAKER_00 4:54

That is just incredible. And you know, it really makes me wonder if AI can perfectly predict something as chaotic and random as spare parts demand, imagine what a fully predictive economy will look like.

SPEAKER_01 5:03

Oh, the potential is staggering.

SPEAKER_00 5:05

Right. We could be moving toward a perfectly frictionless society.

SPEAKER_01 5:08

Yeah.

SPEAKER_00 5:09

Physical goods, resources, even vital medical supplies could be automatically routed to exactly where they are needed before a crisis even hits. It's a future that just frees humanity from scarcity.

SPEAKER_01 5:21

Yeah, a system that anticipatory allows us to stop worrying about basic logistics and really focus our collective energy on innovation and progress.

SPEAKER_00 5:29

It's a beautiful thought to end on. Well, if you enjoyed this deep dive, please subscribe to the show. Hey, leave us a five star review if you can. It really does help get the word out. Thanks for tuning in.