Demand & Sales Forecasting

A demand forecasting service for importers and distributors turns your order history into sales forecasts by product, customer and month, and feeds them into purchasing and planning inside the ERP you already run, such as Priority. How do you know it works? Every forecast is graded against actual sales, and against the method you plan by today.

DataWise builds demand and sales forecasting systems for manufacturers, importers, and distributors, trained on your real order history and graded openly against actuals every month. One engagement of this kind ran for three years as an embedded monthly service for a US-market industrial manufacturer, feeding directly into leadership planning meetings. We deliver forecasts leadership can plan on, not model demos.

In that engagement we forecast demand for hundreds of points of sale, worth several million dollars a month, working with the CEO and CTO. The company manufactures and ships on the forecast we built, three months ahead.

What does DataWise actually build in a forecasting engagement?#

DataWise builds multi-horizon forecasting systems: monthly and annual sales forecasts, quarterly product-family demand forecasts, or item-level projections, depending on the decision they need to feed. Under the hood we use ensemble time-series stacks that combine statistical models, gradient-boosted models, and neural architectures, tuned per horizon, because no single model family wins at every horizon on real business data. Around the core forecast we have shipped, in real engagements:

• customer clustering
• dedicated forecasting for mega-customers
• stockpiling detection
• churn analysis
• an explainability layer
• geographic sales analytics

Every engagement starts with a characterization phase that defines the decision the forecast serves, the data available, and the success metric, before any model is trained.

How do you know the forecast is actually working?#

This is the question that separates a forecasting service from a slide deck, and DataWise answers it with a ledger. In the three-year manufacturing engagement, forecast performance was tracked openly against actuals every month for three years, with 3-month and 12-month error tracking in a yearly performance ledger. Predictions were graded, not just made. The same discipline applies to every forecasting engagement: the accuracy metric is defined during characterization, measured on real data before rollout, and reported against actuals on an agreed cycle after it. If the model stops earning its place in the planning meeting, the ledger shows it first.

How does the backtest on your last 12 months work?#

Before you commit to a forecasting project, you can check what a forecast would have given you over the past year. We build it on the history you already have and test it on your last 12 months, month by month, as if those months had not happened yet. A backtest takes between a few months and a year, depending on the task and what it has to answer. It is scoped and quoted per client: the scope and a fixed price are agreed in writing before it starts, and it begins with a scoping conversation.

What you provide

• Order lines or invoice lines exported from your ERP, by item, customer and date, covering the last 12 months and the years before them. An export is enough; nothing is installed and nothing in your systems changes.
• Your item list with product families, so results can be broken down the way you plan.
• The forecasts or purchase plans you made for those 12 months, if they were saved. If not, tell us the rule you plan by, for example last year plus a percentage, and we rebuild it.
• One conversation to name the decision the forecast should serve: purchasing, targets, or both.

How much history do you need? It depends on the task. A forecast that learns your seasonality needs a few years of data, and one year is not enough for that. A shorter history can still serve other forecasting tasks. Have less? Tell us what you have and we will look together at what your data can support.

What you get

• A month-by-month ledger for the 12 months: our forecast, your current method, and what actually sold, side by side.
• The error at the horizons you actually plan on. In the three-year engagement described above, those were 3 and 12 months ahead.
• Results broken out where the decisions live: by product family, fast movers against slow ones, and your largest customers.
• A written verdict: whether a forecast does better than your current method on your data, where it does not, and what it would take to go further. If it does not do better, we say so and you stop there. The findings stay with you.

How accuracy is graded against your current method

Each forecast is frozen as of the month it would have been made, using only what was known then, so the test cannot peek at the answer. Our forecast and your current method are scored on the same months with the same four numbers: how big the misses are, which way they lean, the same pair weighted by value, and the comparison with a naive baseline such as the same month last year. We agree the measure with you before the run, so no one moves the goalposts afterwards, including us. How to grade a forecast honestly walks through the method step by step.

This is how we worked for three years for a US-market industrial manufacturer: every monthly forecast graded against actuals at 3- and 12-month horizons, in a ledger that sat on the table in leadership planning meetings (the full case). The difference is direction. There, forecasts were graded going forward, month after month. Here, the grading runs backwards over a year that has already happened, before you commit to anything.

Planning on experience vs. planning on a graded forecast#

Which industries and data situations does this fit?#

DataWise has built and validated forecasting on real data across manufacturing and industrial operations and import, distribution, and retail contexts. For example:

• Multi-region quarterly demand forecasts for a global water-technology manufacturer across five sales regions (EU, UK, US, Australia, China), unifying direct sales with inter-company stock transfers.
• A GPU-accelerated commodity-price forecasting engine for feed procurement and agri-business, validated against real feed-mill purchasing history.
• Business-cashflow forecasting built directly from raw bank-account transaction history.

We have also designed and specified forecasting systems for construction-supply procurement and for sea-freight and currency exposure. The common requirement is not big data, it is real transaction history: order lines, invoices, or bank movements over enough time to learn seasonality.