How much sales history do I need to forecast?

To capture seasonality, at least two full years of history are recommended; with three or more, models distinguish trend from seasonal pattern better. With less history you can work with simple methods complemented by business knowledge, gaining accuracy as data accumulates.

What accuracy can be expected from a sales forecast?

It depends on the business and the aggregation level: forecasting total monthly sales for a stable company can reach 5-10% error, while SKU-level forecasting in volatile businesses can exceed 30%. What matters is measuring error systematically (e.g., with MAPE) and comparing against the previous method.

Is machine learning worth it for sales forecasting?

Only when statistical methods fall short: high SKU volume, influential external variables (price, promotions, weather) and sufficient data. For most companies, a well-implemented, automated time series model offers the best balance of accuracy, cost and maintainability.

Analytics

Sales Forecasting: Methods That Work in Real Companies (Not in Papers)

Okun Data Team · May 6, 2026 · 6 min read

Every company forecasts, even if it doesn't call it that: when the owner estimates "this month we'll sell 10% more than last month," that's a forecast. The question isn't whether to forecast, but with which method — and the right answer depends less on mathematical sophistication and more on your data, your seasonality, and who will use the number.

The four maturity levels of forecasting

Level 1 — Moving average: next month will look like the average of the last three or six. Simple, understandable, and surprisingly decent for stable businesses without strong seasonality. Its limit: it's always late to trend changes.

Level 2 — Simple seasonality: compute an index per month (January sells 70% of average, December 150%) and apply it over the trend. With two or more years of history, this method captures the pattern that dominates most businesses: the calendar. For many companies, getting this level right already transforms purchase planning.

Level 3 — Statistical time series models: methods like ETS or ARIMA (or tools like Prophet) that model trend, seasonality and special events jointly, with confidence intervals. They require more expertise but automate well and produce a band of scenarios instead of a single magic number.

Level 4 — Machine learning with external variables: models that, beyond history, incorporate price, promotions, weather, holidays or macro indicators. Worth it when data volume is large (retail with hundreds of SKUs) and when external variables truly move the needle. To go deeper: machine learning for business.

The most common mistake: forecasting the total

A forecast of total sales hides more than it shows. The real value appears when forecasting by product family, by channel or by key account — because that's where decisions about purchasing, production and sales coverage are made. The rule: forecast at the level where you make decisions, and aggregate upward.

Measure the error, not the hit rate

No forecast is exactly right; the goal is to be wrong by little and know by how much. Measuring error systematically (with metrics like MAPE — mean absolute percentage error) lets you compare methods with data: if the usual moving average yields 18% error and the statistical model 11%, the improvement is concrete and quantifiable. It also reveals where the forecast is reliable (stable products) and where it isn't (launches, fashion), so stock policy can be adjusted to risk.

From spreadsheet to process

A useful forecast is not an annual exercise: it's an automated monthly (or weekly) process, where the model recalculates on fresh data, publishes results to a dashboard, and the team adjusts with the context data can't see — a big new customer, a competitor leaving the market. That back-and-forth between model and commercial judgment is where real accuracy comes from.

Conclusion

Don't start with the most sophisticated model: start by measuring the error of your current method, add seasonality if you're not using it yet, and level up only when the error justifies it. A well-implemented Level 2 forecast used every month is worth more than a machine learning model nobody understands or maintains.

Still planning purchases and production by gut feeling?

We'll build a forecast model on your sales history and benchmark it against your current method. Numbers decide.

Request a demo

Frequently asked questions

How much sales history do I need to forecast?: To capture seasonality, at least two full years of history are recommended; with three or more, models distinguish trend from seasonal pattern better. With less history you can work with simple methods complemented by business knowledge, gaining accuracy as data accumulates.
What accuracy can be expected from a sales forecast?: It depends on the business and the aggregation level: forecasting total monthly sales for a stable company can reach 5-10% error, while SKU-level forecasting in volatile businesses can exceed 30%. What matters is measuring error systematically (e.g., with MAPE) and comparing against the previous method.
Is machine learning worth it for sales forecasting?: Only when statistical methods fall short: high SKU volume, influential external variables (price, promotions, weather) and sufficient data. For most companies, a well-implemented, automated time series model offers the best balance of accuracy, cost and maintainability.