Demand Forecasting You Can Run Today, Not in 12 Months

Those are enterprise platforms designed for global Fortune 500 supply chains, with 6 to 18 month deployments and 6-figure annual licenses. Our service is fractional: model plus people, days not quarters, suited to teams under $1B revenue or pilots inside larger organizations. We give you the same calibrated forecasts and S&OP cadence without the platform cost or rip-and-replace project.

Minimum 24 months of historical demand at the granularity you want to forecast (SKU, product family, channel or location). A simple two-column CSV with date and value works to start, and a multi-series file covers many SKUs or locations at once.

For ongoing operation we set up the data feed from your ERP, WMS or BI system as part of onboarding, quoted per system.

It depends on data quality and demand volatility, so we do not quote a number before we have seen your data. Every forecast reports MAPE and bias (whether the model systematically over- or under-forecasts), measured on a holdout backtest: the most recent periods are hidden from the model, forecast blind and compared with actuals.

During onboarding we run the same backtest against your current method or a naive baseline, so any improvement is measured on your own data rather than taken from an industry benchmark. WAPE is being added to the standard report.

Today every series is forecast with Prophet, a decomposable time-series model that separates trend and yearly seasonality and returns an 80% prediction interval. Accuracy is measured out of sample: we hold back the most recent periods (about 20% of the history), forecast them blind and compare the result with actuals.

Next on the roadmap is a model tournament: Prophet, ETS, SARIMA and a seasonal naive baseline run on the same backtest, the best one is picked per series, and every candidate's score is shown so your planners can challenge the choice. Promotions and other external drivers as model inputs come after that.

Seasonality: yearly patterns are detected automatically from history. They need at least two years of data to be reliable, which is why we ask for 24 months to start.

Promotions and one-off events: the current model learns only from history, so promotional spikes and stockout dips show up as unusual points in the data. We are adding planner controls to adjust or exclude those periods before the model fits, with every change logged. Promotion calendars as model inputs follow later.

New products: a statistical forecast needs history. For launches without history we agree a manual approach with your planners during onboarding, for example using a comparable product as a reference.

The forecast is the statistical input to your weekly or monthly S&OP review. Each series comes with its accuracy and bias, so the review can focus on series with high error or a systematic bias instead of every line.

Results are available in a hosted dashboard, as CSV and through a REST API; connecting them to your spreadsheets, BI or ERP is set up during onboarding. Automated exception alerts (Slack, Teams) and scenario comparison are on the roadmap.

Data is processed and stored on AWS in eu-central-1 (Frankfurt), encrypted at rest and in transit. Every forecast is fit only on that customer's own series: we do not pool data across clients and we do not train shared models on your numbers. Uploaded data and results are kept for 24 hours by default.

For enterprise customers we can scope a dedicated deployment in your own AWS account. We sign an NDA before discovery and a DPA before any production data leaves your perimeter.

Two parts. A one-time setup fee covers discovery, a data audit, the model build and a backtest report on your data, with the first forecast in about two weeks. After that a monthly retainer covers retraining, planner support and a regular S&OP review.

Tiers are sized by the number of series you forecast, not by user seats. The minimum term is 6 months on the entry tier and 12 months on larger tiers. That is still a fraction of an enterprise planning platform, without a multi-quarter implementation.

Yes. The forecasting model is industry-agnostic; what changes per industry is how the data is prepared and which effects planners need to watch, for example holiday calendars in food and beverage or supply interruptions in pharma.

The onboarding phase is short enough to validate fit on your own data before the retainer term starts.