You’ve probably been in that meeting. The exec asks for a sales forecast. You know a simple statistical model will do the job perfectly. But the room wants to hear about “AI.” So, you nod, fire up Python, and build a neural network to predict something that barely moves.
We are exhausting ourselves chasing the hype cycle. We take simple, stable data and force it through a 10-million-parameter transformer because a linear model feels too basic to put on a slide deck. But here is the dirty secret of real-world forecasting: A neural network will happily memorize the noise and call it a pattern.
When you have limited data—which is most of us—deep learning models overfit. They hallucinate trends that don’t exist. They break when the underlying environment shifts slightly, and worst of all, they offer no explanation for why they failed. You are left staring at a black box, unable to debug the mess.
This is exactly why ARIMA refuses to die. It is old. It is linear. It is aggressively boring. And it is exactly what you should be using.
ARIMA doesn’t try to be smart. It relies on first principles: stationarity, autocorrelation, and parsimony. It assumes that the future will probably look a lot like the past, just with some trend and seasonality mixed in. In stable environments, that assumption isn’t a flaw; it’s a feature. Complexity is a liability, not a feature. If a model can’t explain why it predicted a spike in Q3 sales, it shouldn’t be trusted with your budget.
When you use ARIMA, you get theoretical guarantees. You get out-of-sample performance that doesn’t degrade the second the market shifts. You get a model you can debug, explain, and trust. You aren’t guessing at hyperparameters or hoping the gradient descent converges. You are applying a mathematically sound, transparent process to a real-world problem.
The industry has convinced us that novelty equals progress. But true engineering isn’t about using the newest tool; it’s about using the right one. Most business forecasting doesn’t need to learn the structure of human language. It just needs to know what happened last Tuesday.
So the next time someone suggests throwing an LSTM at a dataset with 50 rows of monthly data, push back. The future of forecasting isn’t a black box with a billion parameters. It’s a transparent, boring model you actually understand. Embrace the boring. It’s what works.
FAQ
Q: But what if I actually have millions of data points and highly volatile, non-linear data?
A: Then deep learning might be the right tool. ARIMA shines in low-data, stable regimes. If you have massive, chaotic datasets, neural networks can capture complexities that linear models miss. But be honest with yourself: most business forecasting does not involve massive, chaotic datasets.
Q: Why do data scientists still push deep learning for simple forecasting then?
A: Because novelty gets funding, promotions, and conference talks. Nobody gets a raise for deploying a 50-year-old statistical model that just works. The hype cycle demands innovation, even when the 'innovation' performs worse out-of-sample than a baseline.
Q: Is ARIMA actually better, or are you just being a contrarian?
A: It is practically better. In low-data or stable regimes, ARIMA's parsimony gives you theoretical guarantees and out-of-sample performance that deep learning cannot match. A model you can debug and explain is infinitely more valuable in a production environment than a black box that might break tomorrow.