Forecasting methods compared
Choose a forecast from the data pattern and business decision. Moving averages smooth noise, regression models a relationship, expected value combines probability-weighted scenarios, and learning curves model repeated-task improvement. Validate each against assumptions and a simple baseline.
Match the forecast to the task
- Smooth a stable series
- Use a moving average; it can lag a trend or turning point.
- Model an observed driver
- Use regression; validate the relationship and forecast error.
- Combine possible outcomes
- Expected value = sum of probability × outcome.
A probability-weighted mean is not the most likely result. Show material downside and capacity scenarios separately.
See the decision
Forecast the mean and show the uncertainty
Three exclusive demand scenarios are 4,000 units with 25% probability, 6,000 with 50%, and 9,000 with 25%. Capacity is 7,000 units and cannot expand in this period.
Units of demand; probabilities shown in labels
- Low · 25%4,000
- Base · 50%6,000
- High · 25%9,000
Use 6,250 units as the expected-demand estimate and assess the separate 25% capacity-shortfall scenario.
How to apply these ideas
1. Moving average
Average the selected recent observations.
This is a transparent baseline for relatively stable series. It lags sustained growth and does not automatically handle seasonality. Specify the window and compare errors on future periods rather than selecting the window that best fits history.
Watch for: Using a simple average as a seasonal model.
2. Regression
Estimate an outcome from a supported driver relationship.
Inspect residuals, unusual observations and the range of input data. The fitted coefficient describes the model relationship, not necessarily a causal effect. Avoid extrapolating into a different capacity or market regime without justification.
Watch for: Treating high fit as permission to extrapolate indefinitely.
3. Expected value
Multiply each mutually exclusive outcome by its probability and add.
Probabilities must sum to one for the modeled scenarios. The expected value summarizes a distribution; it may not be an actual possible outcome. Show downside and capacity implications separately when the decision is sensitive to risk.
Watch for: Describing the average outcome as a guaranteed result.
4. Learning curve
Identify whether the stated rate applies to cumulative-average or incremental time.
Under a cumulative-average model, doubling cumulative output multiplies average time by the learning rate. Total time then equals cumulative average time times output. The model should not be applied to unrelated tasks or assumed to continue forever.
Watch for: Applying a cumulative-average learning rate to the last unit alone.
Study the topic
Forecasting methods compared: explanations and applications
Select a forecast method, apply it with consistent inputs and evaluate its uncertainty. The budgeting guide owns the overall planning process and the use of forecasts in linked schedules.
By MiloPrep editorial team. Original teaching explanations and illustrative cases; official curriculum references are listed below.
Start with the horizon, driver and available evidence
A moving average smooths recent observations and can be useful for a relatively stable series. It tends to lag when a trend changes. Exponential smoothing gives the newest observation a defined weight while retaining the prior forecast. A basic version does not automatically account for trend or seasonality; those patterns require a suitable extension or another model.
Regression uses explanatory inputs to estimate the outcome. It is useful when the relationship is sufficiently stable and the future driver is itself known or forecastable. Expected-value scenarios are useful when a few distinct outcomes describe the uncertainty. Learning curves address a different question: how average production time changes with cumulative experience under a specified model.
| Method | Useful input | Main failure to check |
|---|---|---|
| Moving average | Recent comparable observations | Lag and unmodeled seasonality |
| Exponential smoothing | Prior forecast and newest actual | Weight choice and structural change |
| Regression | Credible explanatory driver | Extrapolation, instability and leakage |
| Expected value | Defined outcomes and probabilities | Mean concealing a severe downside |
| Learning curve | Cumulative output and learning convention | Confusing cumulative average with incremental time |
Compare forecasts on the same unseen periods
Use a holdout period or rolling evaluation in chronological order. Fit or choose the method using past information, then compare predictions with outcomes that were not used to tune it. A model can fit historical data closely and still forecast poorly. Include a simple baseline so added complexity has to earn its place.
Mean absolute error measures average error magnitude in the original units. Signed mean error indicates direction under a stated sign convention; positive and negative errors can cancel. Mean squared error penalizes large misses more heavily. Percentage measures can be misleading when actual values approach zero. Choose the metric that reflects the operational consequence, such as stockouts versus excess inventory.
Report the estimate with a decision-relevant range
A point estimate does not describe all plausible outcomes. Present scenarios, sensitivity or a statistically justified prediction interval when appropriate. A confidence interval for the estimated mean is not the same as a prediction interval for an individual future outcome. Avoid inventing a probability around a judgmental low/base/high range.
Link the uncertainty to action. If a downside sales outcome breaches the minimum cash balance, show the contingency funding needed rather than hiding the shortfall inside the average. Reforecast when evidence changes, preserve the prior version and record which driver changed so the organization can learn from errors.
Work it through
Worked example: Forecast the mean and show the uncertainty
Three exclusive demand scenarios are 4,000 units with 25% probability, 6,000 with 50%, and 9,000 with 25%. Capacity is 7,000 units and cannot expand in this period.
| Scenario | Probability | Weighted units |
|---|---|---|
| 4,000 units | 25% | 1,000 |
| 6,000 units | 50% | 3,000 |
| 9,000 units | 25% | 2,250 |
- 1
Expected demand = 1,000 + 3,000 + 2,250 = 6,250 units.
- 2
The high-demand scenario exceeds capacity by 2,000 units. Its probability is 25%; the mean alone hides that shortfall.
Use 6,250 units as the expected-demand estimate and assess the separate 25% capacity-shortfall scenario.
Work through the alternatives
More worked cases: forecasting methods compared
Original teaching scenarios. Each case states its assumptions, shows the working and explains the limits of the conclusion.
Compare smoothing with a moving average
The last three monthly demands are 100, 120 and 140 units. The forecast for the latest month was 125. Use a three-month moving average and basic exponential smoothing with alpha 0.4. Next month actual demand later proves to be 150.
| Method | Next forecast | Absolute error against 150 |
|---|---|---|
| Three-month average | 120 | 30 |
| Exponential smoothing | 131 | 19 |
- Moving average = (100 + 120 + 140)/3 = 120.
- Smoothing forecast = 0.4 × 140 + 0.6 × 125 = 131. Absolute errors are |150 - 120| = 30 and |150 - 131| = 19.
What the result means: Smoothing performs better for this one observation, but one observation is not enough to establish a generally superior method. Both may lag a persistent upward trend.
Practice this topic
Check your understanding
Try each short application before opening its explanation. These are original learning checks, not official exam questions or a score prediction.
1. Outcomes are 200 units with probability 0.3 and 400 with probability 0.7. What is expected output?
Show answer and explanation
340 units.
0.3 × 200 + 0.7 × 400 = 340. The planning system must still accommodate the actual possible outcomes.
2. Forecast errors, defined as actual minus forecast, are +20 and -20. What are mean bias and MAE?
Show answer and explanation
Zero bias and 20 units MAE.
Cancellation in the signed average hides the magnitude of the two errors. Report both measures for different purposes.
Questions about forecasting methods compared
- Which forecasting method is always best?
- None. Match the method to data, horizon and decision, then compare out-of-sample error and operational usefulness. More complexity is not automatically more accurate.
- Why report scenarios if I have an expected value?
- The same mean can hide very different downside risks and capacity needs. A decision may depend on the extremes rather than the average.
Sources and scope
Curriculum mapping: Part 1, Section B, Planning, Budgeting, and Forecasting. The official outline assigns 20% to the entire domain, not to this individual technique. These are selected study notes, not complete curriculum coverage.
Official curriculum references establish scope and expected skills. The explanations, diagrams and numerical cases are independent teaching material from MiloPrep, not official questions or source-validated solutions. Curriculum references checked 2026-10-10.
- IMA Content Specification Outlines (effective September 2024) ↗
Supports: Part 1, Section B, beginning on printed page 6. Domain names, weights and curriculum boundaries.
- IMA Learning Outcome Statements (effective September 2024) ↗
Supports: Part 1 B.3: forecast methods; printed page 5. Curriculum scope, not verification of this original worked example.
Independent US CMA study material. Not affiliated with or endorsed by IMA. This covers management accounting, not medical-assistant certification or the separate ICMAI qualification.