Most teams shopping for better demand forecasting techniques are solving the wrong problem.
They’re comparing exponential smoothing to ARIMA. Reading up on Holt-Winters. Wondering whether the Prophet is worth a look. All reasonable questions, and for a single, stable, aggregate series, the honest answer is that swapping one of these for another will move your error by a rounding error. You’ll spend six weeks on the migration and get two points of MAPE.
What breaks forecasts at real companies is scale and grain.
One forecast for total monthly demand is easy. Forty thousand forecasts, at SKU by location by week, refreshed every Monday, with promotions and price changes and a new product line that has nine weeks of history, is a different species of problem. That’s where technique choice stops being academic. Simple methods don’t fail because they’re simple. They fail because nobody can run four hundred of them well, so a planner picks a default, applies it everywhere, and then spends their week overriding the output by hand.
So the question worth asking is “how many series do I need to forecast, at what grain, and what drives them.” Answer that and the technique picks itself. “Which algorithm is most accurate” comes second, and a distant second at that.
Let’s go through the options, from the ones you can run in a spreadsheet to the ones you can’t.
Qualitative demand forecasting: judgment, and when it beats math
Qualitative techniques mean asking people. Sales team estimates, executive consensus, the Delphi method, customer surveys, market research.
These get dismissed a lot, usually by people who’ve never launched a product with zero history. When there’s no data, judgment is the only input available, and a good category manager with fifteen years in the business carries real information that isn’t in any table. New product introductions, market entries, a competitor’s bankruptcy, a regulatory change nobody has a data point for yet.
Where it goes wrong is everywhere else. Human forecasts carry systematic bias, and the bias has a direction: sales teams sandbag, marketing inflates, and everyone anchors on last year’s number. Judgment also doesn’t scale. Nobody’s asking a human to opine on 40,000 SKUs every week.
Use qualitative techniques as an input to the forecast and as a sanity check on the output. Don’t use them as the forecast.
Time series techniques: the workhorses
Time series forecasting looks at one thing: the history of the series itself. No external drivers, no promotions, just the pattern of the numbers over time.
Moving average takes the mean of the last n periods. It’s the simplest thing that works, it lags every real change by roughly half your window, and it’s fine for slow-moving items where the lag doesn’t cost you much.
Exponential smoothing weights recent observations more heavily, which fixes some of that lag. Holt-Winters extends it to handle trend and seasonality. For a stable series with a clean annual pattern, this is genuinely hard to beat, and it’s still the backbone of a lot of retail planning for good reason.
ARIMA models the relationship between an observation and its own lagged values and errors. It’s more flexible, it handles a wider range of behaviors, and it demands more from you: stationarity checks, parameter selection, and enough history to fit against. Auto-ARIMA implementations take some of that pain away.
Croston’s method deserves a mention because intermittent demand quietly wrecks a lot of forecasts. Spare parts, slow movers, anything that sells in scattered bursts. Standard smoothing on a series that’s zero most weeks produces a confident forecast of “a little bit, every week,” which is wrong every single week.
The shared limitation is the same across all of them. These methods assume the future looks like the past, in shape if not in level. They can’t know you’re running a promotion in week 12, or that a competitor just opened a store across the street.
Causal and regression techniques: connecting demand to drivers
Quantitative demand forecasting gets more interesting once you link demand to the things that cause it. Price. Promotion depth. Weather. Marketing spend. Competitor pricing. Local events. Macroeconomic indicators, if your product is sensitive to them.
Multiple regression is the standard entry point, and it does something time series can’t: it lets you ask what-if questions. What happens to demand if we cut price 10%? What did the promotion actually contribute, separate from the seasonal lift it happened to overlap with?
The tax is data and discipline. You need history on the drivers, not just the outcome, and you need to keep collecting it. Multicollinearity will bite you. And relationships between demand and its drivers shift, which means the model that fit beautifully in 2024 quietly stops fitting.
Machine learning techniques: everything at once, across everything
Machine learning approaches, gradient boosting in particular, read all of it together. Seasonality, trend, price, promotion, weather, product hierarchy, cross-SKU cannibalization, store attributes, day of week, holiday proximity, and the interactions between them that a regression would need you to specify by hand.
Three things ML does that the other techniques struggle with.
It captures nonlinear relationships and interactions without being told to look for them. Demand response to a 30% discount is not three times the response to a 10% discount, and the effect is different in December than in March. A tree ensemble figures that out on its own.
It scales across series. One model can learn across your whole catalog, which means a nine-week-old SKU borrows the pattern of the two hundred similar SKUs that came before it. That’s a real advantage over fitting 40,000 independent models, and it’s the thing that makes SKU-level forecasting practical rather than theoretical.
It handles many drivers without a specification fight. You don’t have to decide up front which twelve variables matter.
Where ML doesn’t win: short, thin, stable series with no drivers. If you have one product, three years of clean monthly history, and no promotions, exponential smoothing is a better use of your time. We’d rather say that plainly than pretend the fancy method wins everywhere. Our take on AI forecasting goes deeper on where the gains come from, and they come from scale and signal breadth, not magic.
Comparison: demand forecasting methods at a glance
| Technique | Typical accuracy | Effort to run | Handles drivers | Best for |
| Qualitative (expert judgment, Delphi) | Low to moderate, high variance | High per forecast | Only informally | New products, no history, market shocks |
| Moving average | Low | Very low | No | Slow movers, rough baselines |
| Exponential smoothing / Holt-Winters | Moderate to good | Low | No | Stable series with clear seasonality |
| ARIMA | Good | Moderate | Limited (ARIMAX) | Longer histories, autocorrelated demand |
| Croston’s method | Moderate | Low | No | Intermittent and lumpy demand |
| Regression / causal | Good | Moderate to high | Yes | Price and promotion sensitivity, what-if analysis |
| Machine learning | Best at scale | Low once automated, high to build | Yes, many at once | Thousands of series, many drivers, granular grain |
The number that decides your technique
Count your series. Then count your drivers.
Few series, few drivers, stable history: use exponential smoothing and go do something else with your week. Adding complexity here is a hobby, not a project.
Many series, few drivers: you need automation more than you need sophistication. Even simple methods need a system that selects and fits them per series without a human in the loop.
Many series, many drivers, and a grain fine enough that human review is impossible: this is where machine learning stops being a preference and starts being the only technique that fits the shape of the problem.
There’s a second number worth counting, and almost nobody does. How much of your planners’ week goes to adjusting forecasts by hand? Across the deployments we’ve seen, that number tends to be most of it. A forecast that gets overridden every cycle is a first draft that costs a salary to finish. If your planners are rewriting the output, the technique isn’t the issue and neither is the accuracy. Trust is. And trust comes from forecasts that explain themselves and hold up week over week, which is also why production planning tends to improve faster than the MAPE chart suggests it should.
How we automate ML-based forecasting
Most planning teams stay on statistical methods for a practical reason. ML forecasting at SKU-location-week grain has traditionally required a data science team: feature engineering, model selection, backtesting, validation, deployment, monitoring, retraining. That’s a headcount problem, and headcount is the one thing planning teams don’t get.
Our Predictive AI Agent does that work. You ask a business question, something like “what will demand look like for each SKU by location for the next 13 weeks,” and the agent handles the rest: preparing your historical sales and inventory data, engineering the features, building and validating the model, and enforcing guardrails against the failure modes that quietly ruin forecasts, like data leakage and overfitting. It forecasts across hierarchy levels, from individual SKUs up to families and categories, so the numbers reconcile up and down the chain instead of contradicting each other in the S&OP meeting.
Your data doesn’t need to be pristine first. Real sales history is messy, hierarchies drift, calendars misalign, and the agent is built to work with what you actually have.
Predictions land where planners work, so the output shows up in the planning system instead of a data science notebook nobody outside the team opens.
Where to start
Pick one category. Not the whole catalog, not the pilot everyone’s arguing about. One category with enough history to backtest and enough revenue that a two-point accuracy gain is worth reporting.
Backtest your current method honestly, on a holdout period you didn’t peek at. Most teams don’t know their real MAPE, they know the MAPE of the forecast after the planners fixed it.
Then compare. If exponential smoothing wins on your category, keep it and spend the saved effort somewhere useful. If it loses badly at the grain you actually plan at, you’ve learned something worth acting on.
The technique matters less than the honesty of the test.
See how our AI forecasts demand from your data
Dror is the VP of Data and Analytics at Pecan AI, where he leads the analytics strategy that powers both customer success and Pecan’s own growth. He joined Pecan as Director of Analytics after years of data leadership roles across tech and fintech, bringing a firsthand understanding of what it takes to make data actually useful for business teams.
