Sales & marketing
Sales Forecasting Methods: Choose Inputs and Test a Small Forecast
Choose sales forecasting methods by examining the information available and the period you need to predict. A long list of model names cannot replace that check. This practice guide combines three input-selection cases with a tiny held-out test; the numbers illustrate evaluation rather than establish a reliable commercial forecast.
Match each situation to the evidence it has
Three method-selection cases
New service with no relevant sales history: document assumptions and seek structured input from people with relevant knowledge. Do not claim a seasonal model was fitted to nonexistent history.
Established service with regular weekly sales: compare simple historical benchmarks first; investigate any known changes before assuming older patterns will continue.
Opportunity-level forecast: use current deal values, timing and clearly defined stages. Historical stage probabilities need relevant outcomes; a stage name alone does not supply a probability.
Keep test observations out of the forecast calculation
Forecasting: Principles and Practice distinguishes data used to build a forecast from new observations used to evaluate it. In this invented test, weeks 1–4 are training data: 80, 100, 90, 110 units. Weeks 5–6 are hidden until both methods have produced their two-week forecasts.
Method A holds the last training value at 110 for both future weeks. Method B holds the training mean, (80+100+90+110)÷4=95, for both weeks. Neither method is updated between week 5 and week 6 in this fixed-origin test. These are deliberately simple benchmarks.
Open the held-out results and compare errors
A completed fixed-origin error check
Actual held-out units: week 5=108; week 6=112.
A forecasts 110, 110. Absolute errors: |108−110|=2 and |112−110|=2. Mean absolute error=(2+2)÷2=2 units.
B forecasts 95, 95. Absolute errors: 13 and 17. Mean absolute error=(13+17)÷2=15 units.
A has the smaller error on this two-observation test. This is not evidence that last-value forecasts always outperform means.
Inspect whether the proposed method has adequate inputs
A colleague suggests a yearly seasonal method from these six weekly observations. Reject the proposal for this fixture: six weeks do not contain a yearly pattern to estimate or compare. Another colleague proposes evaluating on weeks 1–4 because those values are already available. That would not be the held-out forecast test described here.
A fuller evaluation would need enough relevant history and test periods for the horizon and situation. A change in product, market or measurement can also make old data less relevant. Keep those concerns visible rather than declaring the winning method from a tiny demonstration ready for production.
Write the result as a bounded conclusion
Completed note: “For the supplied fixed two-week test, the last-value benchmark has MAE 2 units versus 15 for the training mean. Both forecasts were made from weeks 1–4 only. The test is too small to establish stable performance.”
Practice question: if both methods had used week 5’s actual result before forecasting week 6, would this still be the same test? No. It would be a different, updating evaluation and should be labeled and calculated separately.
Questions, answered
Can a method be chosen only because it fits the past?
Past fit does not establish accuracy on new observations. Evaluate genuine forecasts using data that were not used to build them.
Should every business use the same forecasting method?
No. Available data, the required horizon and relevant changes affect which methods can be used and meaningfully tested.
Sources and examples
ResumizeAI team. Examples are fictional teaching material, not customer results. Replace sample details with your own facts.
Your experience. A complete resume.
Build in Resumize with ATS-friendly templates, guided editing and a live preview. Start free, then upgrade when you need more.
Free account required. Free-plan PDF exports include a watermark.