Demand Forecasting
We introduce some basic time series forecasting approaches. Next, we study the "accuracy" of the forecasts and how it can be used to decide between the different forecasting approaches. Finally, we discuss how the forecasts can be further improved by addressing bias and accounting for trends and seasonality.
Time Series Forecasting (6 videos)
Three simple forecasting approaches are discussed in the videos below. Please watch them before attempting the questions in the sections below.
Simple Moving Average (Part 1)
Simple Moving Average (Part 2)
Weighted Moving Average (Part 1)
Weighted Moving Average (Part 2)
Exponential Smoothing (Part 1)
Exponential Smoothing (Part 2)
Quality of forecasts (3 questions)
In the videos above, we introduced how simple time series models can be used for forecasting. Next, we will look at how to quantitatively determine the quality of these forecasts, which can be used to help decide between different forecast approaches.
Consider the following demand data along with the 3-period and 5-period simple moving average forecasts:
Forecast bias (1 question)
The performance measures discussed in the previous section can be used to compare between different forecast approaches. Ideally, the errors should be centered around zero. If the expected forecast error is 0, we say that the forecast is unbiased. Else, we say that the forecast is biased.
Modelling trends - Double exponential smoothing (8 questions)
Moving averages and exponential smoothing are backward looking methods where the resulting forecast is some combination of past demand, resulting in biased forecasts when an increasing (or decreasing) trend is present. In particular, we adjust our forecasts as follows:
Trend-adjusted demand forecast = Demand forecast (w/o trend) + Trend forecast
Under double exponential smoothing, the trend-adjusted demand forecast F and trend forecast T are computed as follows:
Ft = α At-1 + (1 - α) Ft-1+Tt-1
Tt = β (Ft - Ft-1) + (1 - β) Tt-1
For the following questions, assume α = 0.1 and β = 0.2.
Modelling seasonality (7 questions)
Demand could also vary from season to season. For example, restaurants at shopping malls generally experience more demand during weekends than weekdays. Here, we illustrate how seasonality can be accounted when forecasting.