Manufacturing Forecasting: How to Forecast Demand and Plan Production

Demand Forecasts and Related Planning Outputs

A manufacturer works with several related planning outputs, but each answers a different business question:

Preparing a useful demand forecast requires reliable information on customer activity, product availability, business events, and market conditions linked to demand.

What Data Does a Demand Forecast Use?

A forecast is only as clear as the data it's based on. The business must first decide whether “demand” means customer orders, confirmed orders, invoices, dispatches, or estimated customer demand before stock limitations.

Orders and shipments can show different patterns. Shipments understate demand when stock is unavailable. Orders overstate demand when customers cancel, place duplicates, or increase order quantities during a shortage.

Do not delete unusual transactions automatically. A festival surge, large project order, factory shutdown, or customer closure can represent a genuine business event. Preserve the original record and label the event so its treatment can be reviewed later.

Using Indian Economic Data

External indicators can help explain changes that internal sales data does not capture. Useful Indian sources include:

Before using an official economic series, check whether its base year, basket, weights, or methodology changed during the historical period. Do not combine old and revised series without an official linking method or a documented adjustment. Retain an external indicator only when historical testing shows that it improves forecasts at the required horizon.

Separate Demand Signals From Supply Constraints

Supplier lead times, production capacity, labour availability, planned maintenance, and minimum order quantities affect what the business can produce or purchase. They should not be used to alter expected customer demand.

Keeping demand and supply constraints separate allows management to see shortages, capacity gaps, and unmet demand before approving the production plan.

Forecasting Methods and When to Use Them

There is no single best method for every product. The appropriate starting point depends on whether demand is stable, seasonal, affected by external factors, contains long periods with no orders, or lacks sufficient historical data.

Push, pull, and just-in-time systems are production or replenishment approaches, not forecasting methods.

Intermittent Demand

Spare parts, maintenance items, and low-volume components frequently have long periods with no demand. These patterns can require methods designed for zero-demand periods. Standard smoothing methods and percentage-based error measures can produce misleading results in such cases.

Machine Learning

Machine learning can help when a manufacturer has many related products and reliable information about factors such as price, promotions, customer type, or sales channel. Test it against simpler methods using the same historical periods and error measures. Use it only when it performs more reliably at the required product level and planning horizon.

How to Build a Manufacturing Forecast

1. Define the Decision and Forecast Horizon

Start with the decision the forecast must support. A two-week production schedule, a three-month material purchase, and a two-year capacity investment require different forecast horizons and levels of detail. Define the following:

• What will be forecast
• The unit of measure
• The product and location level
• The forecast horizon
• How often the forecast will be updated
• Who owns the decision
• The cost of forecasting too high or too low

Forecast at the level where the business will act. A product-family forecast can support capacity planning but is usually too broad for scheduling individual stock-keeping units.

2. Create a Consistent Demand History

Use the demand sources identified earlier to create one time-based dataset at the required product and location level. Standardise dates, units of measure, and product codes. Correct duplicate records and separate internal stock transfers from customer demand.

Keep the original records unchanged. Use separate fields to identify stockouts, returns, cancellations, product replacements, and one-time events. This preserves an audit trail while allowing the forecasting model to use cleaned data.

3. Group Products and Shortlist Methods

Use the earlier demand-pattern table to shortlist suitable methods for each product group. Do not apply one method automatically across the entire product range.

Also classify products by value, replenishment lead time, and service importance. A high-value component with a long lead time deserves closer review than an inexpensive item that can be replenished within a day. Greater business importance does not automatically require a more complex model.

4. Establish a Baseline and Backtest Each Model

Compare each candidate model against a simple baseline, such as the previous period’s demand, demand from the same period last year, or a moving average.

Backtesting recreates past forecasting decisions using only the information that would have been available at the time. Generate forecasts at several points in the historical data and compare each one with the actual result. This prevents future information from influencing earlier forecasts.

The testing horizon must match the business decision. If materials must be ordered 8 weeks in advance, good one-week performance does not prove that the model is suitable for purchasing.

5. Measure Forecast Error and Business Results

No single measure explains every forecasting problem.

Weighted Absolute Percentage Error cannot be calculated when total actual demand is zero. Percentage-based measures can also become unstable when demand is very low. In these cases, Mean Absolute Error or Mean Absolute Scaled Error can be more informative.

Forecasting error should be assessed alongside stockouts, order fill rate, expediting costs, obsolete inventory, schedule changes, and inventory value. A statistical improvement matters only when it supports better operational results.

6. Control Manual Adjustments and Show Uncertainty

A manual adjustment can be justified when the model lacks important current information, such as a confirmed promotion, customer closure, or product discontinuation.

Record the original forecast, adjustment, reason, approver, date, and final forecast. After actual demand becomes known, compare the adjusted and unadjusted forecasts. This shows whether human judgement improved the result or introduced additional error.

Where possible, present a forecast range or scenarios alongside the most likely value. An 80% prediction interval represents the uncertainty around the forecast. If the model is properly calibrated, actual demand should fall within such intervals in about 80% of repeated forecasts.

7. Convert the Forecast Into a Production Plan

Consider this illustrative plan for the next month:

The initial production requirement is:

10,000 + 1,200 − 2,400 − 1,600 = 7,200 units

Available capacity is only 6,500 units, leaving a 700-unit gap. Management must decide whether to use overtime, subcontract production, reduce the closing buffer, prioritise orders, or accept backorders.

The forecast estimates customer demand. The production plan determines how the business will respond within its stock and capacity limits.

8. Monitor and Improve the Forecast

Save each approved forecast before actual demand becomes known. After the period closes, compare the forecast with actual demand at the same product level and horizon used for planning. Investigate the forecast when:

• Bias continues in the same direction
• A simple baseline repeatedly outperforms the selected model
• Manual adjustments consistently make results worse
• Stockouts or excess inventory increase despite lower forecast error
• Customer behaviour, prices, products, or sales channels change materially

Possible responses include cleaning the data, changing the product grouping, retraining the model, selecting another method, or revising the review process. Automation can speed up monitoring, but responsibility for the final forecast must remain clear.

Conclusion

Manufacturing forecasting gives managers an informed view of demand before money and capacity are committed. Its value lies in showing what is likely to happen early enough for managers to respond. They can then turn the forecast into a workable production plan based on available resources and operating limits. Before approving the next schedule, test the forecast at the same horizon and product level that the business will use.