How much will the slow season cost you?

Your sales

Use one measure in one currency, for example net sales or gross profit, and enter fixed costs on the same basis. Rows are added into calendar months; refunds as negative amounts net against sales; months with no rows count as zero (you are warned). At least 24 months are needed.

Upload or paste your sales, then click Forecast.

Result

Forecast with ranges

Which months are slow and which are peak

Seasonal index: 1.00 is an average month; 0.60 means 40% below average.

Back-test against "same month last year"

MASE below 1 means better than repeating last year's month in the training years; lower is better. MAPE is the average percentage miss over months with sales.

Model chosen

How it is calculated

Aggregation: rows are summed into calendar months. A half-covered first or last month of daily data is dropped so it is not read as a slump. Holt-Winters (Holt 1957, Winters 1960): three estimates are updated each month, the level, the trend and a seasonal factor for each calendar month, with smoothing weights fitted by Nelder-Mead to minimise the squared one-month-ahead error. Four versions compete: additive or multiplicative seasonality, each with an undamped or damped trend; the one with the lowest MASE on a hold-out (the last 12 months, or 6 if you have under 36) is refitted on everything. Ranges: the spread of the model's own one-month-ahead misses, widened with the horizon by the standard Holt-Winters variance recursion (Hyndman and others, 2008), with a small-sample allowance so the 80% band covers about 80% on simulated data. Back-test: from up to 12 starting points the model is refit on the history so far and forecasts the next 6 months; the same months are also "forecast" as the same month a year earlier. MASE and MAPE are reported for both. The model type was picked using data that overlaps the back-test, so the back-test flatters the model a little. Cash cushion: for each month, max(0, fixed costs minus the low band); the cushion is their sum, and the deepest running deficit is shown too.

Not modelled: promotions, price changes, moving holidays such as Easter, trading-day effects, a one-off shock inside your history, and cash timing (this forecasts the figures you give it, not bank balances). Unverified: checked against simulated series with known answers in unit tests, not against a published dataset or a commercial tool.

Common questions

How do I forecast seasonal sales for a small business?

Use a method that learns the yearly pattern from your own history. Holt-Winters exponential smoothing keeps three running estimates: the current level of sales, the trend, and a seasonal factor for each calendar month, and updates them every month. This tool fits it to the date and amount rows you supply, tries additive and multiplicative seasonality with and without a damped trend, keeps the version that predicted a recent stretch of your history best, and projects it forward with ranges.

How much history do I need?

At least 24 months, which is two full cycles of the seasons. With less, one unusual year cannot be told apart from a real seasonal pattern, so the tool refuses to guess and says how many months it found. Three or more years is noticeably better: the seasonal pattern is then averaged over more years and the ranges are narrower.

What do the 80% and 95% ranges mean?

They are prediction intervals built from how far the model's one-month-ahead guesses missed on your own history, widened the further ahead the month is. About four in five real months should land inside the 80% band and nineteen in twenty inside the 95% band. They describe ordinary month-to-month noise, not a shock such as a closure, a new competitor or a bad storm.

Why does the tool sometimes say not to trust the forecast?

Because it checks itself. It repeatedly hides the last few months, forecasts them, and compares the error with the simplest possible forecast: the same month last year. If the model does not beat that, all its extra machinery adds nothing for your business, and the honest advice is to use last year's figures. That happens with erratic sales, few years of data, or a business that changed a lot.

What is the slow-season cash cushion?

You enter your monthly fixed costs: rent, wages, loan payments, whatever you owe even in a dead month. For each forecast month the tool takes the low end of the range and adds up how far it falls below that cost. The total is the shortfall you should expect to fund from savings or credit in a bad version of the coming months, and the tool also shows the deepest running deficit, which is the cushion you need so the balance never goes below zero.

What format should my file be in?

Two columns, date and amount, one row per day, week or month. Dates can be 2024-03-31, 3/31/2024, Mar 31, 2024 or just 2024-03. Amounts can carry currency symbols and thousands separators, and refunds as negative numbers are netted against sales. Rows are added up into calendar months. Months with no rows are treated as zero sales and flagged. Use one currency and one measure, for example net sales.

Is my sales data uploaded anywhere?

No. The file is read and analyzed in this browser tab. Nothing is sent to a server, there are no accounts or cookies, and the tool never stores your sales. It works offline once the page has loaded.