Lecture 3 Flashcards
What is forecast?
A calculation or estimate to predict future events. It is based on past and present data.
What is the purpose of forecasting?
To explain as much of the systematic variability as possible and quantify the unsystematic variability to eventually support a decision.
What is the importance of forecasting?
Forecasts are the basis for nearly all planning decisions in a SC. Poor forecasts cause high inventory costs, underused capacity, and poor customer service.
What are the three golden rules of forecasting?
- The forecast is always wrong
- errors always occur due to the presence of randomness and uncertainty. Forecasts will never be a solution to our problems, but it is something to help in our decision making. It is more a range not an exact number. - The longer the forecast horizon, the harder it is for the forecast to be accurate
- easier to know what happens tomorrow than in a month, but sometimes we have to make planning decisions on long-term - Aggregate forecasts are more accurate
- because it enables us to get a big picture
the relative accuracy of the forecast will tend to be higher when we aggregate across products or locations.
There are two main distinctions for forecasting methods, could you name them?
Qualitative and quantitative forecasting methods. Which one depends on the type of data we want to analyse and predict.
What are the 4 different qualitative forecasting techniques?
- Panel of experts/executives
- Delphi method
- Sales force composite
- Market research
Explain the forecasting with panel of experts/executives.
This technique is based on emotions, intuitions of the company’s executives which have experience. It can involve interviews, discussions and/or surveys.
Explain the forecasting with the Delphi method. What is the difference between the Delphi method and the panel of experts/executives?
The Delphi method involves a systematic and interactive discussion between experts on a given problem or situation. The difference from the panel of experts/executives is that with the Delphi method a consensus is required to determine the final results.
interviews separately -> general conclusion -> consensus?
Explain the forecasting with the sales force composite.
The sales force composite involves interviews and surveys completed by the sales representative of each particular region or area. So, regional customer needs are forecasted based on experience and intuition of the company’s salesperson.
Explain the forecasting with market research
Market research involves qualitative studies on consumer behaviour. Based on interviews, focus groups, and surveys distributed among customers. It aims to understand the changing customers’ needs.
What are the time series components?
Level: an “average” around which observations vary
Trend: a predictable increase or decrease in the level over time
Seasonality: a pattern of predictable and recurring shifts in the level
Random noises: unpredictable variations in the data
What are the different types of trends?
Upward, downward, up-to-down, down-to-up trend
What does a good forecast contain?
It explains level, trend, seasonality, and describes random poses. It should be robust and not sensitive (reacting quickly to random noise).
What is systematic and unsystematic variability?
systematic: trend and seasonality -> explain
unsystematic: random noises -> quantify
When do you apply the moving average? What is moving average? What is the effect of N?
Time series with no trend nor seasonality, so we focus on forecasting the level. It is the average of the most recent N observations. A larger N is used if you want smoother forecast, and a smaller N is used for a more responsive forecast.
When do you apply exponential smoothing? What is exponential smoothing? What is the effect of the weighted average (alpha)?
Time series with no trend nor seasonality, so we focus on forecasting the level. It is the weighted average of the previous forecast and the last demand point. The weighted average can be between 0 and 1. A smaller alpha is used for a more smooth forecast, and a larger alpha for a more reactive forecast (chases demand). It is important to decide which level we have in the data. A changing level (up down, not a lot of noise) should have a high alpha (fast learning) and a stable level (more noise) should have a low alpha (slow learning).
When do you apply Holt’s method? What is Holt’s method? How do we get level and trend at timepoints 0?
Time series with trend but no seasonality. Holt’s method predicts both the level and trend, so it has a level component and a trend component each with its own weighted average (alpha or beta). You can get L0 and T0 using linear regression. Follow D(t)=a+bt, then L0 is the intercept and T0 is the slope.
When do you apply seasonal decomposition? What is seasonal decomposition? How would the forecast steps be?
Time series with seasonality but no trend. Seasonal decomposition removes the seasonal effect from data before conducting a forecast using the seasonal relatives.
- compute the season averages
- compute the overal average of the seasonal averages
- compute the seasonal relatives (season average/overall average)
- divide sales data by its seasonal relatives -> deseasonalized data
- apply forecasting methods like MA or ES on the deseasonalized data
- incorporate the seasonality again by multiplying with seasonal relatives
How do you calculate the forecast error? What different types of forecast error do you have? What can you do with the forecast errors?
You calculate forecast error by subtracting the original demand from the forecast. The other types are absolute forecast error and absolute percentage forecast error. You can sue forecast errors to determine the accuracy and bias of the forecast.
What is the difference between the measures of accuracy and the measures of bias?
The measures of accuracy tell us how big the error is and the measures of bias tell us what direction the error is.
What do the measures of accuracy tell you? What are the three different measures of accuracy?
Measures of accuracy tell you about the average size of the error.
Mean absolute deviation (MAD)
Mean squared error (MSE)
Mean absolute percentage error (MAPE)
What is the difference between the measures of accuracy and the measures of bias?
The measures of accuracy tell us how big the error is and the measures of bias tell us what direction the error is.
What do the measures of bias tell you? What are the three different measures of bias?
It tells you if the forecast tends to be constantly too high or consistently too low.
Mean forecast errors (MFE)
Running sum of forecast errors (RSFE)
Tracking signal (TS)
What do the measures of accuracy and bias mean?
MAD
MSE
MAPE
MFE
RSFE
TS
- MAD: average of absolute errors over a period
good if MAD is small - MSE: average of squared errors over a period
large MSE means higher presence of large errors - MAPE: average of absolute percentage forecast error over a period
forecast accuracy is often defined as 1-MAPE
it is useful when demand range is wide - MFE & RSFE
positive -> too optimistic
negative -> too pessimistic - TS
unbiased if TS is between -0.5 and 0.5
