Choosing a forecasting method is a business decision before it is a software decision. This guide shows how to match a method to your planning horizon, available data, error costs, and operating capacity without assuming that more complicated models are better.
How Do You Choose a Forecasting Method?
Start with the decision the forecast supports, then compare a simple baseline with a small number of credible alternatives using historical, time-ordered tests. Choose the least complicated approach that produces useful improvements at the horizons you actually need. Include uncertainty, integration, explainability, and maintenance in the decision rather than selecting on an accuracy score alone.
Define the forecast before choosing the model
Write a short forecasting brief: what you predict, at what level, how far ahead, how frequently, and who will act on it. Weekly product demand, quarterly cash receipts, and annual market growth are different problems even if all three appear in the same planning dashboard.
Separate the forecast horizon from the refresh frequency. A forecast refreshed every morning might still estimate demand eight weeks ahead because supplier lead times determine the decision. Evaluating only tomorrow's accuracy would miss that need.
Specify the target carefully. Recorded sales are not always the same as customer demand: an unavailable product cannot generate a sale, and returns can distort the apparent pattern. Decide whether you need units ordered, units shipped, net revenue, or another measure before comparing methods.
Also identify the cost of different mistakes. Underestimating demand may create shortages, while overestimating it may create waste. A method with slightly lower average error can still be less useful if it repeatedly misses the expensive cases. These costs should shape the evaluation, not be added after a winner is announced.
Match the method family to the evidence
A method is a set of assumptions about patterns that may continue. More records can help estimate those patterns, but volume alone does not tell you which assumptions are appropriate. Thousands of nearly identical observations may offer less useful information than several complete seasonal cycles.
| Method family | Sensible starting situation | Main question to check |
|---|---|---|
| Naive or seasonal naive | A benchmark based on the latest value or matching seasonal period | Does anything more elaborate improve on it? |
| Exponential smoothing | Level, trend, or seasonality in a numeric series | Are these patterns stable enough to extrapolate? |
| ARIMA-style models | Time dependence in a regularly spaced series | Is the history informative at the required horizon? |
| Regression with predictors | Known relationships and available explanatory inputs | Will those inputs be known when forecasting? |
| Machine learning | Many potentially useful predictors and sufficient examples | Does added flexibility improve realistic tests? |
| Structured judgment | Sparse history, a new launch, or a major change | Are assumptions documented and challengeable? |
The forecasting textbook Forecasting: Principles and Practice describes simple benchmarks, including naive and seasonal-naive methods, as useful comparison points (simple forecasting methods). A sophisticated model should earn its extra complexity against these alternatives.
Unstructured data does not automatically make machine learning the right choice. Text or images must first become usable inputs, and their availability, processing cost, and relevance need checking. Conversely, a relatively simple regression may capture a valuable promotion effect if that information is reliable.
Build a fair comparison with time-ordered tests
-
Set the baseline. Why: Without a reference, an impressive-looking score says little about practical improvement. How: Use a last-value or seasonal benchmark appropriate to the series, and record its performance by horizon.
-
Recreate the decision date. Why: Future information can accidentally make an unsuitable model appear excellent. How: At each historical test date, use only records and input versions that would have been available then. Revised economic data and finalized sales totals deserve particular attention.
-
Test several forecast origins. Why: One quiet period may not represent peak season or operational stress. How: Move the training cutoff forward and evaluate repeated forecasts at the required horizon. The textbook's time-series cross-validation discussion explains this rolling-origin approach.
-
Report several useful views. Why: An overall average can hide systematic underprediction or weak performance on important products. How: Review absolute error, directional bias, business costs, and results for meaningful segments. Percentage errors need care when actual values are zero or close to zero.
-
Keep a final holdout. Why: Repeatedly tuning against the same evaluation period can overfit the selection process. How: Reserve a later, untouched period for a final comparison, while recognizing that even a strong result cannot guarantee future performance.
Apply the same information cutoff and target definition to every candidate. A model receiving finalized promotion results should not compete against one receiving only the original promotion plan.
Worked example: choosing for a replenishment decision
Consider a hypothetical distributor ordering six weeks before delivery. It has three years of weekly sales, a promotion calendar, and records of stockouts. Its team compares seasonal naive forecasts, exponential smoothing, and a regression using planned promotions.
All numbers in this example are hypothetical. Across repeated six-week-ahead tests, the average absolute error is 24 units for the benchmark, 21 for smoothing, and 19 for regression. However, during promotion weeks the regression systematically underpredicts, and its inputs depend on a calendar that sales staff often update late.
The team does not automatically select the lowest average error. It checks whether the apparent improvement survives when only the promotion information actually available six weeks earlier is used. It also measures errors for fast-selling products, where shortages matter most.
Suppose the realistic regression error becomes 23 units. Smoothing may then be the better operating choice: it improves on the baseline without relying on an unreliable input. The team can retain documented promotion overrides while testing whether better calendar discipline makes regression worthwhile later.
The lesson is not that smoothing wins generally. It is that the best method depends on the real decision and information supply chain, not a model label or a laboratory score.
Include uncertainty and explainability
A point forecast is a central estimate, not a promise. Ask whether the method can produce intervals or useful scenarios and whether those outputs have been checked against historical outcomes. The textbook explains prediction intervals as ranges that express uncertainty under a forecasting model.
Intervals are conditional on assumptions and do not automatically cover unprecedented disruptions. For a supplier shutdown or a regulatory change, a separate scenario may be more honest than stretching a normal forecast beyond its evidence.
Explainability requirements also depend on the decision. A planner may need to distinguish a seasonal increase from a promotion effect; an auditor may need reproducible inputs, model versions, and approval records. Do not assume an explanation chart proves that a relationship is causal.
Choose the level of transparency with the relevant business, risk, and compliance owners. Avoid blanket claims that every organization in an industry must use one particular model type. Applicable obligations depend on jurisdiction, use, and the consequences for people.
Check operating fit and total cost
Compare the whole workflow, not just training speed. Include data preparation, licenses, computing, monitoring, documentation, staff time, and the effort required when an input feed fails. A cheap model with fragile manual preparation can be expensive to run.
Check that outputs reach the actual planning process in usable form. This might mean a scheduled export rather than a complex live integration. Confirm units, product identifiers, refresh times, and which version is authoritative when systems disagree.
Assign an owner who can explain unexpected changes and a fallback that works without the most fragile inputs. Set review triggers such as sustained bias, missing data, or a material business change. Retraining should follow evidence and a controlled process, not a belief that models automatically improve as new records arrive.
Finally, document why the selected method won and which conditions would justify reconsidering it. That record makes future changes more disciplined than periodically buying a more elaborate tool.
Questions About Forecasting Method Choices
Can I combine several forecasting methods?
Yes. An average or weighted combination may reduce reliance on one model's weaknesses. Evaluate the combination as another candidate, including its operational complexity, rather than assuming that combining models always improves results.
What if there is no history for a new product?
Use relevant analogues, explicit assumptions, and structured expert judgment. Explain where the analogy may fail and update the forecast as actual orders arrive. A narrow numeric interval is difficult to defend when the evidence is sparse.
Should different products use different methods?
Sometimes. Stable high-volume items and intermittent sellers can behave differently. Grouping similar series can make evaluation manageable, but maintain consistent reporting and avoid creating more model variants than the team can support.
How often should we reconsider the chosen method?
Use a regular review schedule alongside event-based triggers. A new sales channel, changed lead time, or persistent error shift can justify review sooner. Replacing the model every time one week is inaccurate usually creates noise rather than improvement.
Match Methods to Data and Decisions
The right forecasting method is the one that supports a defined decision with credible tests, understandable uncertainty, and sustainable operating effort. Start simple, challenge candidates fairly, and demand practical value from added complexity.
Next steps: Review the predictive analytics introduction, prepare an AI forecasting implementation checklist, or explore the Predictions hub.
Sources
- OTexts — Some simple forecasting methods
- OTexts — Time series cross-validation
- OTexts — Prediction intervals
This guide is general education, not professional or financial advice.