Demand Forecasting for Indian Distributors: Data Requirements and Pilot Evaluation

Demand forecasting estimates future demand so a distributor can plan purchasing and stock allocation. A useful forecast is tied to a decision: which product, at which warehouse, for what period, with which lead time? A forecast for the whole business may look accurate while giving a planner little help with the products that cause shortages.

This guide proposes a pilot for an Indian distributor using existing sales and inventory records. Examples are illustrative planning exercises, not Brainguru client results. The objective is to compare a proposed method with the current process before connecting recommendations to replenishment.

Choose the Decision and Forecast Horizon

Start with one product family and the warehouses that serve it. Agree whether planners work in days, weeks or months, and which decision the forecast supports. The forecast horizon should relate to lead time and the review cycle; a daily estimate is not automatically useful for an order that arrives several weeks later.

Separate demand forecasting from replenishment. The forecast estimates an outcome. A replenishment decision also considers current stock, incoming orders, minimum order quantities, storage limits and service objectives. Do not turn a predicted quantity into a purchase order without those rules and the planner’s review.

Prepare a Minimum Data Dictionary

• Product: stable SKU identifier, unit of measure, pack size and any product replacements.
• Location: warehouse identifier and the sales or fulfilment region it represents.
• Time: transaction date and the meaning of each date, including order, invoice and dispatch dates.
• Quantity: sales, returns and cancellations recorded separately, with a defined aggregation method.
• Availability: stock levels or flags showing when a product could not be supplied.
• Planning context: promotions, known price changes, lead times and relevant events available when the forecast is made.

Review identifiers before choosing a model. A product renamed in the ERP may appear to be two unrelated items. A case quantity mixed with individual units can create an apparent demand spike. A weekly total that includes transfers between warehouses may measure internal movement rather than customer demand.

Sales Are Not Always the Same as Demand

When a product is unavailable, recorded sales can understate demand. An apparently quiet week may therefore reflect a shortage, not reduced customer interest. Mark stockout periods and document how the pilot treats them. If lost demand is not observed, avoid presenting an estimated adjustment as a known fact.

Returns, bulk orders and cancelled orders also need a stated policy. Keep the original record and the transformed quantity, so a planner can see what the model used. For a large one-off order, a reviewed event flag may be more useful than silently letting it alter every future estimate.

Illustrative Pilot: One Product Family Across Two Warehouses

A fictional distributor selects a stable product family with weekly replenishment reviews. The team assembles dated sales, availability and lead-time records, then defines a forecast output containing SKU, warehouse, target week, predicted quantity, forecast date and model version. The planner also sees the recent history and any missing-data flags.

The pilot uses the same product and time boundaries for both the existing planning method and the proposed model. It includes ordinary weeks, promotions and periods with limited stock. New products are identified separately instead of being hidden inside a score dominated by established products.

The final screen should show what the forecast is for, when it was made and which information was available at that time. A forecast made before a promotion cannot fairly be compared with another generated after actual sales became known.

Compare With a Simple Baseline

A baseline is a method you can explain, such as using the previous comparable period, a moving average or the current planner’s estimate. Select one that reflects the business pattern and document its limitations. The proposed AI model needs to justify extra data preparation, operating costs and maintenance against that baseline.

Do not change the baseline after seeing which result makes the new model look strongest. Keep a record of the chosen method, forecast dates and evaluation rules. If the simple method performs well enough for the decision, using it can be a reasonable outcome of the pilot.

Test in Time Order and Avoid Information Leakage

Train or configure the method using earlier information and test forecasts for later periods. Repeat the evaluation across several forecast windows. Microsoft’s forecasting evaluation documentation describes held-out testing and rolling backtests; it also explains why a short test period may provide an unreliable estimate.

Information leakage occurs when the model receives something that would not have been known at prediction time. Examples include final month-end totals, later cancellation outcomes or a promotion plan added after the forecast date. Check the availability date of each input, not just its transaction date.

Use realistic forecast horizons in the backtest. Evaluating tomorrow’s demand does not establish performance for the next six weeks. Keep late-arriving records and corrections visible so the team understands the difference between a historical database view and what planners actually knew.

Measure Error and Business Consequences

• Absolute error: the size of the difference between forecast and actual quantity, in the agreed units.
• Bias: whether the method systematically over- or under-forecasts a product group.
• Segment results: performance by warehouse, product family, sales frequency and availability.
• Planning outcomes: shortages, excess inventory, urgent purchases and the effort needed to review recommendations.

Percentage error needs care where actual demand is zero or very small. Select and document a metric appropriate to the data rather than reporting a flattering percentage without its denominator. The scikit-learn evaluation documentation provides definitions and limitations for common regression metrics.

As an illustrative example, forecasting 120 units when actual demand is 100 creates an absolute error of 20 units and an over-forecast of 20. Whether that is acceptable depends on margins, lead times and the cost of holding or running out of stock. One error score cannot answer that commercial question alone.

Use Planner Review Before ERP Write-Back

Start with a report or read-only planning interface. Show forecast history, inventory context and the reasons for any exception flag. Record whether the planner accepts, changes or rejects a suggestion, together with a reason that can support later analysis.

When adding ERP integration, agree identifiers, update frequency, permissions and retry behaviour. A duplicate message must not raise a duplicate purchase order. An unavailable feed should trigger an explicit stale-data warning. Procurement authority and stock adjustments remain governed by the agreed business process.

When the Pilot Should Pause or Expand

Pause when source data changes without explanation, errors worsen for an important product group or planners repeatedly reject recommendations for the same reason. Investigate the process and data before assuming that a more complex model will solve the issue.

Expansion should have its own evaluation. Results for stable products do not establish performance for intermittent demand, new launches or another warehouse. A proposal should identify the data owner, acceptance tests, refresh process and ongoing costs. See logistics and supply chain AI development for the broader implementation scope.

Frequently Asked Questions

How much sales history is needed for demand forecasting?

It depends on seasonality, product turnover and the forecast horizon. Check whether the history covers relevant business patterns and availability changes; a fixed number of months is not a universal readiness test.

Should we forecast sales or demand?

Define the quantity required for the planning decision. Recorded sales may understate demand during stockouts, so document availability and any assumptions used to estimate unmet demand.

What is a forecasting baseline?

It is an agreed comparison method, such as a moving average, previous comparable period or the current planning process. Evaluate the new model against it on the same products and time windows.

Why should evaluation follow time order?

A future forecast must use information available at the prediction date. Time-based testing helps identify whether the method works on later periods without benefiting from future information.

Can a forecast automatically create a purchase order?

The forecast alone does not include every replenishment constraint. Start with planner-reviewed recommendations; any write-back needs separate authority, validation, duplicate handling and an audit trail.

What should we do with a new product?

Treat it separately from established products. Review available product information and comparable history, make assumptions visible and evaluate its forecast as a distinct group.

Related Reading and Services

• Forecasting model development
• AI pilot scoping and evaluation
• AI automation in Indian business
• AI agents for operations

Discuss a Scoped Pilot With Brainguru

Brainguru Technologies provides custom AI development from Noida for businesses in India and internationally. If you are evaluating this workflow, share the current process, representative data, system documentation and the person responsible for review. Explore the relevant industry AI service or request a free consultation. Deliverables, costs and support are agreed for the project.