Volatile Demand Forecasting: 9 Strategies That Work

When demand swings by a third from one month to the next, a single forecast number is a guess dressed up as a plan. Planners respond with spreadsheet overrides, extra safety stock and last-minute expediting, and the plant ends up with stockouts and excess stock at the same time. This guide gives nine strategies for forecasting volatile demand, each tied to evidence you can check, and it is written for plants that still forecast in Excel or from Tally exports.

It starts with how to tell whether an item is really volatile, because the right method depends on the pattern.

What is volatile demand?

Volatile demand is demand that swings widely and unpredictably from period to period. A simple measure is the coefficient of variation, which is the standard deviation divided by the average. An item that sells 80, 120, 60, 140, 90 and 110 units in six months averages 100 with a coefficient of variation of about 0.29 (using the sample standard deviation), while an item that sells between 98 and 102 every month sits near 0.01.

Volatility also has more than one shape, so classify each item before you pick a method. The categorisation scheme from Syntetos, Boylan and Croston uses two measures: the average interval between demands (ADI) and the squared coefficient of variation of the demand sizes (CV²). Its cut-offs of 1.32 and 0.49 sort items into four types.

Pattern	ADI	CV²	What it looks like	Common starting point
Smooth	Below 1.32	Below 0.49	Regular orders of similar size	Standard time-series methods
Erratic	Below 1.32	Above 0.49	Frequent orders with swinging sizes	Combine several models; forecast ranges
Intermittent	Above 1.32	Below 0.49	Occasional orders of similar size	Croston-type methods
Lumpy	Above 1.32	Above 0.49	Occasional orders with swinging sizes	Croston-type methods plus judgement on large orders

The last column is a common starting point, not a rule from the paper. To classify an item, count its demand months: an item with demand in 3 of 12 months has an ADI of 4.

Nine strategies for forecasting volatile demand

Apply these in this order, starting with the cheapest:

1. Classify every item before you choose a method.
2. Set a naive benchmark, and beat it.
3. Separate baseline demand from events and outliers.
4. Use intermittent-demand methods for lumpy items.
5. Combine several models instead of betting on one.
6. Forecast ranges, not single numbers, and size safety stock to the range.
7. Reforecast weekly, and change decisions only when it matters.
8. Control manual overrides.
9. Measure with WAPE, bias and service, not MAPE alone.

1. Classify every item before you choose a method

One method for every item is the most common mistake in volatile demand. Run the ADI and CV² split on your top 200 items from the last 24 months, and give each type its own method and review rhythm.

2. Set a naive benchmark, and beat it

A complex model can lose to "same as last month", especially when demand is noisy. Steve Morlidge's analysis in Foresight found that around 50% of individual forecasts fail to beat a naive forecast. Our guide to demand forecasting error statistics shows how to run that test in a spreadsheet.

3. Separate baseline demand from events and outliers

If a tender deal, a festive-season push or a stockout month sits inside your history, the model learns it as normal demand. The textbook Forecasting: Principles and Practice handles one-off events with spike and dummy variables that remove their effect instead of deleting the data.

In practice, tag events at the transaction level, such as promotion, tender deal, anomaly and stockout period. Record lost sales during stockouts too, because a stockout hides true demand.

4. Use intermittent-demand methods for lumpy items

Spare parts, made-to-order jobs and slow movers have many zero months, where ordinary averages mislead. Syntetos and Boylan tested four methods on 3,000 real intermittent demand series from the automotive industry. A modified Croston method beat a 13-period moving average, simple exponential smoothing and Croston's original method in their out-of-sample comparison.

Results depend on the data, so backtest on your own items before you switch.

5. Combine several models instead of betting on one

In the M4 forecasting competition, 12 of the 17 most accurate methods were combinations, mostly of statistical approaches. The six pure machine-learning methods performed poorly, and none beat the combination benchmark.

M4 used 100,000 series across many fields, not only manufacturing demand, so treat it as a guide to method design. Averaging two or three simple models is a cheap way to apply it.

6. Forecast ranges, not single numbers, and size safety stock to the range

The forecasting textbook says point forecasts can be of almost no value without the accompanying prediction intervals. A range tells you how much buffer a lead time needs: safety stock is a z-score times the forecast error times the square root of the lead time, as the worked example in our statistics guide shows.

The z-score comes from the cost of a miss. In an MIT lecture on the newsvendor model, an extra unit costs 1 if it is wasted and earns 4 if it sells. The optimal in-stock probability is then 4 ÷ (4 + 1), or 80%, so set service levels by the cost of a miss against the cost of excess, not by habit.

7. Reforecast weekly, and change decisions only when it matters

More frequent reforecasting is not automatically better, especially if it whipsaws MRP. This is our recommended rhythm, not a research finding: refresh demand signals daily, publish an operational reforecast weekly, and hold a monthly consensus review.

Change purchase and production decisions only when a forecast moves outside a band you agree in advance. High cadence with low noise keeps the plant from chasing every swing.

8. Control manual overrides

Fildes, Goodwin, Lawrence and Nikolopoulos analysed more than 60,000 forecasts from four supply-chain companies. Planners' adjustments helped on average at three of the four, but small adjustments often reduced accuracy and upward adjustments were much less likely to help.

Put guardrails in place: reason codes for every override, approval when a change exceeds your band, and a monthly check of whether overrides beat the baseline.

9. Measure with WAPE, bias and service, not MAPE alone

Mean absolute percentage error is undefined when actual demand is zero and extreme when it is close to zero, which describes exactly the items that make demand volatile. Track WAPE for the units you actually make, bias for direction, and fill rate or stockout days for the service you deliver.

Our guide to demand forecasting methods for manufacturers goes deeper on MAPE, WMAPE and bias.

Which forecasting tools are most accurate for volatile demand?

No tool wins everywhere, and the evidence shows why. M4 found that combinations beat single methods, and the intermittent-demand test found that the best method for intermittent series was not the standard one.

So judge any tool on your own items with a backtest against a naive forecast, and check that it handles intermittent demand, event tags and forecast ranges. Our guide to the top demand forecasting tools for manufacturers compares ten options.

Do you need to replace your ERP to forecast volatile demand?

Usually not. Keep your ERP, whether SAP, Odoo or Tally, as the system of record for items, stock and orders, and add a planning layer that produces the baseline, the event adjustments and the ranges.

Our demand planning software works that way: it runs on top of SAP, Odoo or Tally and drafts the purchase for your approval. If your ERP cannot store lead times or event tags at all, that is a different problem, and our manufacturing ERP page covers it.

Volatile demand forecasting: common questions

What is volatile demand?

It is demand that swings widely and unpredictably between periods. Measure it with the coefficient of variation, and classify each item as smooth, erratic, intermittent or lumpy before you choose a method.

How do you forecast volatile demand?

Classify each item, beat a naive benchmark, tag events, combine several models and forecast ranges. Then reforecast weekly, control overrides and track WAPE, bias and service.

What is the best method for intermittent or lumpy demand?

Croston-type methods are the usual starting point for items with many zero months. A modified Croston method beat standard alternatives on 3,000 automotive series, but backtest on your own data because results depend on the series.

How often should I reforecast?

Refresh demand signals daily, reforecast operationally every week, and hold a monthly consensus review. Change decisions only when the forecast moves outside an agreed band.

How much safety stock do I need for volatile demand?

Safety stock is a z-score times the forecast error times the square root of the lead time. Double the forecast error and the buffer doubles, which our forecast error statistics guide shows with a worked example.

Conclusion

Volatile demand is not solved by one smarter model. Classify your items, beat a naive forecast, forecast ranges and control overrides. To see how Likwid does this on your own data, talk to our team, or read about our demand forecasting software.