Sales revenue LAST 30 DAYS
—Observed sales × recorded prices
Unable to import data
YOUR SALES. IN FOCUS.
Make room for what’s next.
Forecast demand, understand the uncertainty, and keep stock in step.
SYNTHETIC DEMO DATALoading sales historySix products · two years of daily sales
Forecast demand NEXT 30 DAYS
—Units across selected products
Forecast error HELD-OUT WAPE
—Lower is better · final 28-day test
Reorder signals INVENTORY PLAN
—Based on stock and lead times
Sales & demand forecast
Daily units · observed sales and the path ahead
Actual salesForecastPlanning rangeUNITS / DAY
Learning from sales history…
Product outlook
From predicted demand to your next stock decision.
| Product | 30-day demand | Expected value | Test error | Stock plan |
|---|
Planning signals
ACTIONABLEYour stock signals will appear here.
Forecasts estimate observed sales. Lost sales from stockouts aren’t modelled.Models fit in your browser
FROM FORECAST TO FULFILMENT
Stock with a plan.
Set available stock, supplier lead times, and a safety-cover target.
Order estimates cover demand during lead time plus safety-cover days. They assume no incoming orders; update the stock you expect to have available.
Stockout dates are estimated from point forecasts, with no replenishment. They are shown only within the selected forecast horizon. These are planning estimates.
EVIDENCE BEFORE ESTIMATES
Measure the model.
Forecasts earn their place by outperforming a simple alternative.
TIME ORDER IS PRESERVED
Learn from the past. Test on what comes next.
The model is selected on two 28-day validation windows. The final 28 days are held back for an independent test. After evaluation, the selected model is refit on all available sales.
INITIAL TRAINING—Earlier observations
VALIDATION56 daysTwo forecast windows
FINAL TEST28 daysNever used to choose a model
Model comparison by product
Results on the independent final 28-day test.
| Product | Selected model | Regression WAPE | Weekly WAPE | Selected MAE | Range coverage |
|---|
Calendar regression
A Ridge regression learns a trend and weekday patterns. Annual Fourier terms are enabled only when the initial training window includes a full year. The intercept is not regularised.
L2 penalty α = 3 · non-negative predictionsWeekly baseline
The final observed week repeats into the forecast. It is a useful benchmark that can win when a more complex model does not improve validation errors.
Seasonal naïve · 7-day cycleAn honest planning range
Bounds use the 80th percentile of absolute validation errors. Coverage is measured on the final test. Summed product bounds are a planning envelope, not an 80% portfolio confidence interval.
No claim of guaranteed future coverageBRING THE SALES STORY
Your data, your forecast.
Daily sales in. A transparent, evaluated forecast out.
Bring your sales history.
Drop a CSV here or choose a file. Processing stays in this browser; your sales aren’t sent to a forecasting server.
Download sample sales CSV Up to 30,000 rows · 50 products · 8 MBA simple, daily format
| date | product | units | unit_price |
|---|---|---|---|
| 2026-10-01 | Product A | 24 | 129 |
| 2026-10-02 | Product A | 31 | 129 |
| 2026-10-03 | Product A | 0 | 129 |
- date, product, units are required. Unit price is optional.
- Use YYYY-MM-DD and at least 140 continuous daily observations per product.
- All products must share the same final date. Missing days are flagged.
- Duplicate product/date rows are summed. Zero means known zero sales.
- Units and prices must be non-negative. Returns need separate treatment.
Active dataset
Synthetic demonstration data
| Product | Daily records | First date | Last date | Latest unit price | Seasonality |
|---|
Your uploaded sales stay in the browser session.
Reloading restores demo data. Stock settings are stored locally in this browser. Synthetic demo prices, stock levels, and sales patterns are invented for the project.
YOUR MACHINE LEARNING PROJECT
Meet SalesCast-AI.
A complete sales forecasting and inventory-planning project.
A better forecast starts
with a fair test.
Explore a fictional six-product business, bring your own daily sales, and see how a learned forecast compares with simply repeating last week.
Try it in four steps
- Explore the demo.
Choose a product and a 14-, 30-, 60-, or 90-day forecast. Hover over the chart for daily values.
- Inspect the evidence.
Open Model performance. Compare error against the weekly baseline and check test-period range coverage.
- Plan stock.
Change stock, lead time, and safety-cover days. Review order quantities and estimated stockout dates.
- Bring your own sales.
Use the sample CSV format, import continuous daily records, and export the forecasts for analysis.
What’s in the project
APP
Complete browser workspace
Forecast charts, CSV import, model evaluation, and inventory planning.
IPYNB
Executed training notebook
Data generation, scikit-learn models, honest validation, and evaluation plots.
PY
Reproducible ML pipeline
Fixed seed, documented assumptions, and cross-language verification fixtures.
MD
Project report & setup guide
Architecture, methodology, formulas, experiments, and limitations.
RUN LOCALLY
python -m http.server 8000 --directory distThen open http://localhost:8000. No API key is required.