AAE5110 Chap.5 Demand Forecasting and Uncertainty
Demand Forecasting and Uncertainty
Forecasts carry a centre and a spread
A demand forecast is a distribution rather than one central estimate. This chapter calculates expectation and dispersion, aggregates independent variance correctly, and interprets capacity quantiles as planning choices rather than promises. It then builds and challenges time trends through residual patterns, structural breaks and simple benchmarks.
The final drills connect probability tails to asymmetric shortage and excess costs, preserving the distinction between forecast accuracy and the economic objective used to select capacity.
Independent variation aggregates at the variance level
Expected values add. Variances add when random demands are independent, so the standard deviation of a sum is the square root of the total variance.
Positive correlation weakens that diversification effect because high outcomes arrive together. State the dependence assumption before using the square-root rule.
A total standard deviation is still expressed in the original demand unit, which makes it suitable for a capacity buffer after the model is chosen.
Quantiles are operating thresholds
A ninetieth-percentile capacity lies at or above modeled demand in ninety per cent of cases and below demand in the upper ten per cent. It does not imply a ninety per cent load factor or forecast accuracy.
The economically suitable percentile depends on the cost of shortage relative to excess. Spill can forgo fare or disrupt an itinerary, while unused capacity disappears at departure. Those asymmetric consequences, not the mean alone, select the tail.
Trend models inherit their sample
A time trend summarises an average path but does not explain the mechanism producing demand.
Residual cycles suggest seasonality, widening spread suggests changing variance, and a level shift suggests a structural break. Plot and inspect residuals before extrapolating, compare the fitted model with a simple benchmark, and justify the sample when policy, capacity or market conditions changed. A more complex fit earns its place only when it improves the decision beyond the estimation period.
What this chapter covers
- 01
Random variables
- 02
Expected value
- 03
Variance and standard deviation
- 04
Independent aggregation
- 05
Normal quantiles
- 06
Trend forecasting
- 07
Residual diagnosis
- 08
Asymmetric error cost
Worked example · free
Set capacity for a service target
- 2Identify the requested quantile.
- 4Substitute mean, z and standard deviation.
- 2Interpret the threshold.
Key terms
- Random Variable
- A numeric value assigned to each possible outcome.
- Expected Value
- The probability-weighted centre of a random variable.
- Variance
- Expected squared deviation from the mean.
- Standard Deviation
- Dispersion expressed in the original unit.
- Quantile
- A threshold with a stated share of outcomes at or below it.
- Residual
- Actual demand minus fitted demand.
Demand Forecasting and Uncertainty FAQ
Why do independent standard deviations not add?
Independent variances add, and standard deviation is the square root of the total variance. Adding standard deviations directly exaggerates uncertainty except in special dependence structures.
What does a capacity percentile mean?
A ninetieth-percentile capacity is at or above modeled demand in ninety per cent of outcomes and below demand in ten per cent. It does not imply a ninety per cent load factor.
How do residuals challenge a trend model?
A seasonal cycle, widening spread or persistent run of one sign shows structure the fitted line missed. A level shift may indicate a new regime that older data should not dominate.
Exam move
Define the demand variable and unit, combine means and variances only under stated dependence, draw the relevant tail, and translate the probability into a capacity or inventory consequence. Compare any fitted trend with a simple benchmark and inspect residuals before extrapolating. Practise independent totals with unequal standard deviations so variance addition becomes deliberate rather than automatic.
For each percentile question, shade the tail and write its complement before reading a z value. Keep forecast accuracy separate from decision loss by assigning different costs to shortage and excess, then observe how the chosen percentile moves. Build a residual plot for an original time series containing seasonality, widening spread and a level break, and identify which pattern invalidates a simple trend.
End each model with the sample boundary, forecast horizon and the operating action that uses the result. Compare an independent total with a positively correlated case. The difference reveals why the square-root benefit is conditional and prepares you to challenge the aggregation assumption.
Working through Demand Forecasting and Uncertainty in AAE5110? Sia is AskSia’s AI Economics tutor — ask any AAE5110 Demand Forecasting and Uncertainty question and get a clear, step-by-step explanation grounded in how AAE5110 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.