Sales Forecast Calculator (from Your Monthly Sales)

About the Sales Forecast Calculator (from Your Monthly Sales)

Paste or open at least 12 months of sales — revenue or units — and get the next 3 to 24 months (or 4 to 104 weeks) from four methods side by side: a moving average, a linear trend, a damped trend (Holt's method) and Holt-Winters with additive or multiplicative seasons. They are the methods Hyndman and Athanasopoulos teach in Forecasting: Principles and Practice.

Each method is fitted on your history without its most recent months, scored on those months by MAE and MAPE, and the most accurate one is highlighted; two naïve benchmarks show whether any method does better than repeating the last month or last year. The forecast comes with an 80% and a 95% range, and you can add changes you already know about — a price rise, a new branch — on top. Totals by financial year follow the year you choose.

Without a pass the forecast is a free preview: the most accurate method, the scores of the best methods and the first periods of the forecast show, with the chart watermarked; the workbook, the CSV and the summary unlock with a Pro pass. Everything is worked out in this browser.

Common uses

• A shop or a small business planning stock, staff and cash for the next year from its own monthly sales.
• A sales manager checking whether a target is in reach on the trend so far.
• An accountant giving a client a forecast by financial year, with a range rather than a single figure.
• Testing what a price rise or a new outlet would add on top of the usual pattern.

The four methods

• Moving average: the next period is the average of the last k periods; k is 3, 6 or 12 months (4, 8, 13 or 26 weeks), whichever had the smallest one-step error. It follows the level and ignores trend and seasons.
• Linear trend: a least-squares straight line through time, carried forward.
• Damped trend (Holt): a level and a trend smoothed from the data, ℓ(t) = α y(t) + (1 − α)(ℓ(t−1) + φ b(t−1)) and b(t) = β(ℓ(t) − ℓ(t−1)) + (1 − β) φ b(t−1); the forecast h periods ahead is ℓ + (φ + φ² + … + φʰ) b, so the trend flattens out (φ between 0.8 and 0.98).
• Holt-Winters: Holt with a season of 12 months (52 weeks), added to the level (additive: the busy season adds about the same amount every year) or multiplying it (multiplicative: the season grows with the business; needs sales above zero in every period).

The smoothing weights α, β*, γ and φ are chosen from a grid by the smallest sum of squared one-step errors.

How the methods are scored

• About the last fifth of your history (3 to 12 months, 8 to 26 weeks) is kept back; each method is fitted on the rest and forecasts those periods.
• MAE = the average of |actual − forecast|, in your own units. MAPE = the average of |actual − forecast| ÷ actual × 100; periods with zero sales are left out of the MAPE, where a percentage has no meaning.
• The most accurate method is the one with the smallest MAE (MAPE breaks ties). The two benchmarks — next period = last period, and next period = the same period last year — show how hard your history is to forecast: if neither of the four beats them, the page says the history is too irregular.
• After scoring, every method is fitted again on the whole history for the forecast.

Where the 80% and 95% ranges come from

For the smoothing methods the page simulates 2,000 possible futures from the method’s own equations, each period drawing a random error like the method’s past one-step errors, and reads the 10th–90th and 2.5th–97.5th percentiles; for the linear trend it uses the regression’s prediction interval, ŷ ± z σ √(1 + 1/T + (x − x̄)² ÷ ((T − 1) s²ₓ)), with z = 1.28 or 1.96. The ranges assume the future varies the way the past did; a shock the history never had is outside them.

Sources

• Forecasting: Principles and Practice (3rd ed.): Holt’s method and the damped trend
• Holt-Winters’ seasonal method
• Evaluating point forecast accuracy (MAE, MAPE, test sets)
• Forecasting with ETS models and simulated prediction intervals
• Prediction intervals of a regression forecast

Limitations

• A forecast carries the past forward: a new competitor, a lost customer or a change in the market shows in the ranges only as far as similar things happened before.
• Seasons need two full years of history (24 months or 104 weeks), and are scored only when two years come before the test periods.
• Months with no row count as zero sales: add them if they were not really zero.
• Weekly data uses weeks from Monday to Sunday and a season of 52 weeks, so a 53-week year drifts by a day or two.
• The results are projections from your own numbers, not a promise of sales, and not financial advice.

Privacy

On this device; its files download once

Your sales history stays in this browser on this device; nothing you paste or open is uploaded. Your settings (the periods to forecast, the date order and the financial year) are kept in this browser.