Why Method Selection Is a Portfolio Question
The most expensive mistake in demand forecasting is searching for one method to rule the whole portfolio. A company selling staple products, seasonal items and new launches simultaneously needs three different engines, because the patterns differ structurally: staples are smooth and predictable, seasonal items repeat a shape, and new products have no history at all. Forecasting is a segmentation problem before it is a modeling problem.
The portfolio view also disciplines expectations. A blended accuracy target, say 85% overall, sounds reasonable until you see it is 95% on staples subsidizing 60% on launches, and the launches are where the stockouts and write-offs live. Segment first by product family behavior, then choose the cheapest method that meets the accuracy each family needs, and measure accuracy per family. The rest of this article is the selection map.
The Six Methods, Mapped
The table below is the working map: what each method does, where it fits, and the watch-outs that trip teams in practice. The first four are statistical baselines every team should run regardless of ambition, because they are cheap, transparent and hard to beat on simple patterns. The last two extend the ceiling where data and events justify it.
One reading note: the watch-outs column matters more than the method column. Most forecasting failures in mid-size companies are not modeling failures; they are data hygiene failures (promotions in the history that were never flagged), override cultures (every forecast adjusted until it is an opinion), or granularity mistakes (forecasting at the wrong level of the hierarchy).
| Method | Best for | Data needed | Watch out for |
|---|---|---|---|
| Naive / random walk | Very stable items, quick baseline | One period of history | Blind to trend and seasonality |
| Moving average | Stable items, smoothing noise | Several periods | Lags every turn; window size is a judgment |
| Exponential smoothing (Holt-Winters) | Trend and seasonality, the workhorse | 2-3 seasonal cycles | Parameter drift; needs seasonal rechecks |
| ARIMA-class models | Correlated, autocorrelated series | Long, clean history | Fragile to regime changes; expert tuning |
| Machine learning (boosting, LSTM) | Complex patterns, many drivers, big data | Volume, variety, clean labels | Opaque; easy to overfit; needs MLOps discipline |
| Judgmental (Delphi, structured overrides) | New products, events, thin history | Expert input, analogies | Bias and politics without a structured protocol |
The AI Ceiling, Quantified
Machine learning forecasting earns its hype in specific conditions, and the quantified evidence is worth knowing before the vendor meeting. McKinsey's research on AI in supply chain reports forecast error reductions of 20 to 50% versus traditional methods, driven by the models' ability to use external signals, weather, events, promotions, pricing, that classical methods ignore. The downstream math compounds: the same research ties AI forecasting to up to 65% fewer lost sales and 5 to 10% lower warehousing costs.
The honest caveats are three. Gains concentrate where data is rich and patterns complex; stable simple SKUs barely move. Models need operational care, retraining, monitoring, drift checks, which is an ongoing capability, not a purchase. And the baseline matters: a 30% error cut from a naive start and from a tuned Holt-Winters start are very different projects. Run the cheap baseline first; it is both the control group and the fallback.
Sources: McKinsey, Succeeding in the AI supply-chain revolution, 2021 · McKinsey, 2021
Judgment: Structured or Corrupting
Every forecast meets the sales meeting, and the meeting decides whether judgment is an input or a contaminant. The research pattern is consistent: unstructured overrides, gut-feel percent changes applied to whole families, reliably make forecasts worse, while structured overrides, named events with expected size, direction and duration, make them better. The difference is not whether humans participate; it is whether their knowledge enters as evidence or as authority.
The protocol that works: statistical forecast publishes first, no pre-discussion. Overrides happen on a form: which event, what size, what dates, who owns the claim. Every override is tracked, and a monthly review scores the overriders like any other forecast source, with the top three best and worst adjustments read aloud. Teams that run this loop for two quarters usually find that 80% of the override volume came from three people reacting to three recurring biases, and the fixes become embarrassingly simple.
- Statistical baseline first, published before discussion
- Overrides on a form: event, size, dates, named owner
- Monthly scoring of override accuracy, read aloud
- Analogy-based forecasting for new products: closest prior launches, scaled
Measuring Forecast Accuracy Without Fooling Yourself
Accuracy measurement has its own failure modes, and the two most common are level confusion and bias blindness. Level confusion: forecasting at aggregate level where errors cancel (a portfolio can look 95% accurate while half its SKUs are 60%), so measure at the decision level, the SKU-location where stock is actually positioned. Bias blindness: average percentage errors hide systematic over- or under-forecasting, which is the expensive kind, because bias means inventory or service damage accumulates in one direction.
The working dashboard is short: MAPE per product family at the decision level, bias per family (signed, not absolute), and exception counts, how many items moved into or out of tolerance this month. Pair it with the inventory math the forecast feeds: the safety stock guide shows how accuracy drives buffer size, and the supply chain statistics roundup provides external benchmarks to compare against.
A 30-Day Forecasting Reset
Thirty days is enough to move a mid-size portfolio from one-size-fits-all to segmented baselines. Week one: segment the portfolio by pattern (stable, seasonal, short-lifecycle, new) and clean the history, flag every promotion and stockout so the models learn demand, not supply. Week two: run the three cheap baselines per family and publish honest accuracy. Week three: fix the worst family with its best-fit method and install the override protocol. Week four: wire the forecast into safety stock and the S&OP meeting, and set the monthly scoring rhythm.
The reset is also the curriculum of the seven-day Supply Chain & Logistics Optimization program, which runs it on your own SKU data with daily missions. For the strategic layer around forecasting, network and inventory decisions, our supply chain optimization definition and scenario planning guide cover how forecasts feed the bigger bets.
Frequently asked questions
▸What are the main demand forecasting methods?
Six families cover practice: naive (last period repeats), moving average, exponential smoothing (including Holt-Winters for trend and seasonality), ARIMA-class statistical models, machine learning models (gradient boosting, LSTM), and judgmental methods (sales input, Delphi panels). Most real companies should run a portfolio: statistical baselines for the bulk, judgment for events, machine learning where data volume justifies it.
▸What forecast accuracy should I expect?
It depends on the product family more than the method. Stable A-items commonly reach 85 to 95% accuracy at monthly level; seasonal and short-lifecycle items live at 60 to 80%; new products start near guessing and need analogies. The discipline is segmenting: one accuracy number across a mixed portfolio hides exactly the families where the money leaks.
▸How much can AI improve demand forecasting?
McKinsey's supply chain research reports AI-driven forecasting reduces errors by 20 to 50% versus traditional approaches, which can translate into up to 65% fewer lost sales, 5 to 10% lower warehousing costs and 25 to 40% lower administration costs. The gains concentrate in families with rich data and complex patterns; stable simple items barely improve, which is why segmentation still comes first.
▸Should judgment be part of a statistical forecast?
Yes, but structured, not casual. The evidence consistently shows unmanaged overrides corrupt forecasts, while structured adjustments for named events (promotions, product changes, market entries) improve them. The rule: statistical baseline first, judgment overrides only with a stated reason, a size, and a tracked owner, then override performance reviewed monthly like any other forecast input.
▸How do I choose a method for a product family?
Ask four questions: how much history exists, is the pattern simple or seasonal, how much data volume and variety is available, and what does the decision cost when the forecast misses. Low history and high stakes push toward judgment and analogies; rich history and stable patterns make exponential smoothing hard to beat; big data and complexity justify machine learning. The table in this article maps it.
▸What should I do with the forecast once I have it?
Wire it into the decisions it exists for: safety stock and reorder points (our guide to calculating safety stock shows the math), capacity and staffing plans, and S&OP review. A forecast that only lives in a planning tool is a report. Teams that want the full loop practiced on their own data run the seven-day Supply Chain & Logistics Optimization program.