Demand Forecasting Methods: 10 Techniques Compared

The 10 demand forecasting methods that matter split into two families: quantitative methods that learn from history (naive, moving average, exponential smoothing, Holt-Winters, ARIMA/SARIMA, Croston's, causal regression, machine learning) and qualitative methods that lean on judgment (sales-force composite, Delphi/expert panel). You don't pick one. You pick a method per demand profile, and the skill is matching the technique to the SKU, not falling in love with a model.

I learned this the hard way running the demand planning function at a $250M manufacturer. We had 8,000 SKUs and a planning team of four. Below I compare all 10 on accuracy, data appetite, and where each one actually fits.

The two families, and why it matters

Every demand forecasting method falls into one of two camps.

• Quantitative — driven by data. Statistical time-series and machine learning. Good when you have history and the future rhymes with the past.
• Qualitative — driven by judgment. Sales input, expert panels, market intelligence. Necessary for new products, step-changes, and anything with no usable history.

The mistake I see most: teams running pure sales gut-feel on mature, high-volume SKUs that statistics would forecast better and cheaper. And the reverse — running statistical models on new-product launches where there's no history to learn from. Match the family to the situation. The standard texts split forecasting the same way, and the methods below map cleanly onto the judgmental and quantitative chapters in Hyndman and Athanasopoulos (2021).

The 10 methods, compared

Here's the full menu in one view. Read "typical accuracy" as relative within a sane demand profile, not an absolute promise.

The quantitative workhorses

Naive and moving average

The naive forecast says next period equals this period. It sounds dumb. It's also the baseline every other method has to beat, and on very stable items it's hard to beat. Moving average smooths the last few periods into one number — useful for slow movers where you want to damp noise, useless when there's real trend or seasonality because it always lags.

Exponential smoothing and Holt-Winters

Simple exponential smoothing (SES) weights recent demand more than old demand. It handles smooth, trendless, season-less items well and almost nothing else. Holt-Winters extends it with separate equations for level, trend, and seasonal index — the classic answer for items with both a slope and a seasonal pattern.

The mechanics are public and worth reading once: the NIST/SEMATECH Engineering Statistics Handbook (2012) lays out the three smoothing equations and the alpha, beta, gamma parameters in plain math. Tune those parameters to minimize squared error and Holt-Winters carries a surprising amount of mid-market demand on its own.

ARIMA and SARIMA

ARIMA models the autocorrelation structure of a series directly — what today's demand tells you about tomorrow's. SARIMA adds a seasonal term. These shine when a series has strong, stable autocorrelation and clean seasonality, and they reward an analyst who knows how to read the diagnostics. They punish one who fits them on autopilot. For most teams, a well-tuned Holt-Winters gets 90% of the benefit with a fraction of the babysitting.

Croston's method: the one for lumpy demand

This is the method everyone forgets, and the omission quietly builds dead stock. Intermittent demand — spare parts, aftermarket, low-volume B2B — has long runs of zeros punctuated by sporadic spikes. Run standard exponential smoothing on that and it systematically over-forecasts, because it spreads the spike across the zero periods.

Croston (1972) solved this by forecasting two things separately: the size of demand when it occurs, and the interval between occurrences. The refinement work by Hyndman and colleagues (2005) gave Croston's method a proper statistical model and led to bias-corrected variants like TSB. If you carry a spare-parts portfolio and you're not using Croston's or TSB, you're over-forecasting your slowest movers right now. I go deeper on this in our guide to forecasting intermittent demand for spare parts.

Causal regression and machine learning

Causal / regression

When price, promotion, weather, or a macro signal moves demand, model the driver directly. Causal regression ties demand to those explanatory variables, so a planned 20%-off promo produces a forecasted lift instead of a surprise. The catch is honest: you need to know the drivers and have clean history for them. The residual error on promo-heavy items is almost always the promo lift itself — so model the lift, don't smooth over it.

Machine learning

ML demand forecasting is oversold for mid-market manufacturers. It earns its keep in exactly three conditions.

1. Scale. Thousands of SKUs where hand-tuning statistical models per item isn't feasible, so one model that learns across the portfolio wins on labor alone.
2. Rich features. You actually have price, promo, weather, web traffic, and macro signals to feed it. ML with no features is just a slower moving average.
3. Clean, deep data. Garbage in, confident garbage out — and ML hides its garbage better than a transparent statistical model does.

The upside is real when those conditions hold. McKinsey (2023) reports AI-driven forecasting can cut errors by 20 to 50 percent and reduce lost sales from unavailability by up to 65 percent. And the market is moving: Gartner (2025) predicts 70% of large organizations will adopt AI-based supply chain forecasting by 2030. If you can't check all three boxes, though, a disciplined statistical-plus-causal approach beats a half-baked ML project — and it's explainable when the CFO asks why the number moved. For the deeper trade-offs, see our primer on machine learning for demand forecasting.

The qualitative methods

Sales-force composite

Reps roll up their account-level expectations into a forecast. It's the right tool for lumpy B2B and project demand, where a single rep knows about a deal no statistical model can see. The danger is optimism bias and sandbagging, so discipline it: measure whether the reps' input actually beats a naive baseline. Often it doesn't.

Delphi / expert panel

A structured, anonymous round-robin among experts that converges toward consensus without the loudest voice winning. Use it for genuinely new products and step-changes where there's no history to fit. Pair it with analog modeling — borrow the launch curve of a comparable SKU — and you have a defensible number. Our guide to new product demand forecasting with no data walks through the analog approach end to end.

How to choose: a decision rule that works

Forget model worship. Segment your SKUs first, then assign a method per segment. An ABC-XYZ inventory analysis gives you exactly the volume-and-variability grid you need to do this.

• Smooth, high-volume A-items (CV < 0.5): Holt-Winters or exponential smoothing is plenty. If price and promotion swing demand, layer causal regression on top. Don't reach for ML here — the lift over a well-tuned statistical model is usually small and the maintenance cost is real.
• Intermittent / spare parts (CV > 1.0): Croston's or TSB. This is the single most common error I see on aftermarket portfolios.
• Promo- and price-driven items: causal regression or ML with the drivers fed in. Model the lift directly.
• New products: qualitative. Analog/like-modeling off a comparable SKU's launch curve plus a sales-force or expert input.
• Lumpy B2B / project demand: sales-force composite, disciplined with forecast value added.

The method nobody lists: ensemble + FVA

The highest-accuracy approach isn't on the list. It's running several methods, picking the best per SKU automatically (a champion-challenger setup), then measuring forecast value added (FVA) so human overrides are only kept when they beat the machine.

FVA is the change in forecast accuracy attributable to each step or participant in the process. The Foresight reality check on FVA (2013) showed something that should worry every planning leader: judgmental overrides frequently make the forecast worse, not better. In the teams I've run, this two-step discipline moved accuracy more than swapping any single algorithm. The model matters less than the process around it — which is also the through-line in our 9 tactics to improve forecast accuracy and the full FVA how-to guide.

The bottom line

Demand forecasting methods aren't a menu where one wins. Smooth items want exponential smoothing or Holt-Winters. Lumpy items want Croston's. Promo-driven items want causal models. New products want judgment. And a large, feature-rich portfolio is where ML pays off. The real edge is matching method to demand profile, then policing it with FVA so you keep only the human touches that actually help.

Not sure which methods fit your portfolio? PlanForge runs a free planning-maturity and stranded-inventory teardown. We profile your SKUs by demand variability, tell you which method belongs on each segment, and show where your current approach is building dead stock. Book a 30-minute call and we'll map your portfolio together.

Frequently asked questions

What is the most accurate demand forecasting method?

There's no single most accurate method — accuracy depends on the demand profile. Holt-Winters and ARIMA lead on smooth seasonal items, Croston's leads on intermittent demand, and machine learning leads on large feature-rich portfolios. The most accurate setup overall is an ensemble that runs several methods and picks the best per SKU automatically.

What is the difference between quantitative and qualitative forecasting?

Quantitative forecasting uses historical data and statistical or machine learning models, so it needs usable history to work. Qualitative forecasting uses human judgment — sales input, expert panels, market intelligence — and is the right choice when there's no relevant history, such as a new product launch. Most mature planning functions blend both, applying judgment on top of a statistical baseline.

Which forecasting method works best for intermittent or spare-parts demand?

Croston's method or its bias-corrected variant TSB. Standard exponential smoothing systematically over-forecasts lumpy demand because it spreads sporadic spikes across the zero periods, which quietly builds dead stock. Croston's fixes this by forecasting demand size and the interval between demands separately.

Do I need machine learning for demand forecasting?

Only if you meet three conditions: thousands of SKUs, rich explanatory features like price and promo, and clean deep data. McKinsey (2023) reports AI-driven forecasting can cut errors 20 to 50 percent when those conditions hold. Without all three, a well-tuned statistical-plus-causal approach is usually more accurate and far easier to maintain and explain.

How do I choose a forecasting method for my SKUs?

Segment first, then assign a method per segment rather than picking one model for everything. Use an ABC-XYZ analysis to sort items by volume and variability: smooth high-volume items get exponential smoothing or Holt-Winters, intermittent items get Croston's, promo-driven items get causal regression, and new products get qualitative judgment. Then measure forecast value added so you keep only the overrides that beat a naive baseline.