Quick demand forecast trial
See which of seven statistical models works for your sales or demand history: the latest periods are held out for testing, the best model is chosen and refitted on all data to give a forecast with an approximate interval. The result is a baseline; it is not sufficient on its own for a decision.
Data is processed in your browser; nothing is sent anywhere.
Daily, weekly or monthly series; the frequency is detected automatically. The delimiter is detected automatically (tab, ; , |); decimal comma and point are both accepted. Date formats are yyyy-mm-dd, dd.mm.yyyy and yyyy-mm. Without a date column the rows are treated as equally spaced periods.
01
How to use it
A
Paste the date and value columns or upload a CSV; the frequency is detected automatically, change it if needed.
B
Set the forecast horizon and the seasonal period; choose how missing periods are treated.
C
Review the model comparison, the chart and the interval forecast table; download the table as CSV.
02
Which models does the tool test, and how?
Seven models are tried: naive (carries the last value forward), seasonal naive (repeats the previous season), moving average, simple exponential smoothing, Holt (level and trend) and Holt-Winters (level, trend and season; additive and multiplicative). Seasonal models need training data of at least two full seasons; otherwise they are skipped and the reason is shown. The multiplicative model only works if all values are positive.
The smoothing coefficients α, β and γ are found with a grid and local search so as to minimise the sum of squared one-step-ahead errors (SSE) on the training data. The moving-average window is chosen by the same criterion. The last h periods are held out (the test); models are trained on the remaining data, and how well they predict the test period is measured with MAE, WAPE and MASE.
The model with the lowest MAE on the test is chosen and refitted on all data to produce the forecast. A single test window is noisy: if two models are very close, the ranking could change in another window. A longer history and testing over several windows (backtesting) is more reliable.
03
Why is the forecast interval approximate?
The intervals are based on the root mean square error (RMSE) of the chosen model's errors on the test period; errors are assumed to be roughly normal and future errors to be as large as the past test errors. The factor is ±1.28 for the 80% interval and ±1.96 for the 95% interval.
This approach is practical and easy to understand, but it is optimistic in two ways: if the test period is short the error size is poorly estimated, and parameter uncertainty and the uncertainty that grows with the horizon are not accounted for. Read the interval as a rough scale of the uncertainty, not as a firm guarantee.
04
What is a baseline good for, and what does it not show?
A statistical baseline answers the question "what happens if past patterns continue?". It is a good benchmark for checking whether your current forecasting process (planning team, ERP forecast) beats it; a process that cannot beat it should be reviewed.
It knows nothing about promotions, price changes, stock-outs (which are not sales demand), holiday calendars, new product launches or competitor moves. Periods with stock-outs make true demand look low and mislead the model. If such factors drive your demand, you need models with explanatory variables and data preparation.
FAQ
- Is the data I enter sent anywhere?
- No. All calculations run in code inside your browser; the data is neither transmitted to a server nor stored.
- How many periods of data do I need?
- Non-seasonal models need at least 6 periods, but far more is better for a meaningful test. Seasonal models (seasonal naive, Holt-Winters) need at least 2 full seasons in the training data: 24 for monthly, 104 for weekly, 14 for daily (weekly season).
- What happens to missing periods?
- Gaps in the date column count as missing periods. Depending on your choice they are taken as 0 or filled linearly between neighbouring values; the summary shows how many were filled. Use 0 if there really were no sales, filling if the data was lost.
- Does the forecast interval guarantee the actual value will fall inside?
- No. The interval is an approximation derived from the test-period errors; actual values are expected to fall inside at roughly that rate, but it is not guaranteed.
- Why were some models skipped?
- Seasonal models are skipped if the training data is shorter than two full seasons or the seasonal period is 1; multiplicative Holt-Winters is also skipped if any value is zero or negative. The reason is shown in the table.
- Can I use the result directly for ordering or production planning?
- We do not recommend using it alone. The tool gives a baseline that excludes factors such as promotions, price and stock; use it together with human knowledge and, where needed, a more comprehensive model.
From a baseline to a real demand forecast
We combine your sales, stock, price and campaign data into a demand forecast by product and location and measure its accuracy with backtesting. Let's talk about your data in a free discovery call.