Prelim 2 – Module 5: Forecasting Flashcards
Method: Historical Analogy
Type:
Data Required:
Relative Cost:
Horizon:
Type: subjective
Data Required: experience
Relative Cost: high
Horizon: medium to long
Method: Regression
Type:
Data Required:
Relative Cost:
Horizon:
Type: causal
Data Required: all past data
Relative Cost: medium
Horizon: medium
Method: Moving average
Type:
Data Required:
Relative Cost:
Horizon:
Type: time series
Data Required: recent data
Relative Cost: very low
Horizon: short
Method: Exponential Smoothing
Type:
Data Required:
Relative Cost:
Horizon:
Type: time series
Data Required: last forecast and smoothed value
Relative Cost: very low
Horizon: short
How do we calculate an N-period moving average forecast?
-> what is MAt
-> what is At
MAt = the N-period moving average at the end of period T
At = actual observation for period T
Forecast formula
Ft+1 = MAt
Strengths of a moving average forecast
- only need N observations to make a forecast
- very inexpensive and easy to understand
Drawbacks of a moving average forecast
- does not consider observations older than N periods
- gives equal weight to last N observations
Exponential smoothing advantages
- old data are never dropped but have progressively less influence
- don’t need to keep any historical information; only need most recent smoother value
St =
Smoothed value at end of period T
𝑆𝑇=𝑆(𝑇−1)+𝛼∗(𝐴𝑇−𝐹𝑇 )=(1−𝛼)∗𝑆(𝑇−1)+𝛼∗𝐴𝑇
At =
Actual observation for period T
Ft+1 = St
Forecast for period T+1
New smoothed value =
Old smoothed value + α*observed error
Which of the following two statements is true?
1. When alpha is small (i.e., closer to 0), older data points have more weight in determining the forecast.
2. When alpha is large (i.e., closer to 1), older data points have more weight in determining the forecast.
- When alpha is small (i.e., closer to 0), older data points have more weight in determining the forecast.
𝛼 = 1
Naive forecast (i.e., use last actual value)
When is the weighted average age of your data the same under exponential smoothing and moving average forecasts?
WA of data under exponential smoothing is usually lower than moving average forecasts.
They can be made equal by adjusting the smoothing parameter in exponential smoothing to match the time span of the moving average window.
How do we measure forecast accuracy?
Cumulative forecast error (CFE)
Mean absolute deviation (MAD)
Mean squared error (MSE)
Mean absolute percentage error (MAPE)
Cumulative forecast error (CFE) purpose
Should be close to zero. Otherwise, forecast is biased.
Mean absolute deviation (MAD) purpose
Common measure of error. Gives equal weight to all errors.
Mean squared error (MSE) purpose
Gives extra penalty to large error values
Mean absolute percentage error (MAPE) purpose
Use percentages to prevent large forecast values from dominating the accuracy measure.
Exponential smoothing with seasonal adjustment
Data has a ___
cycle of L periods (hours, days, months, seasons, etc.)
Exponential smoothing with seasonal adjustment use seasonality index lt to ___
deseasonalize data in each cycle
How to use Solver to optimize smoothing parametres
Set objective: choose the performance score that you want to minimize
By changing variable cells: choose your smoothing parametre or smoothing parametres
subject to the constraints:
- all smoothing parametres <= 1
- all smoothing parametres >= 0
Use GRG Nonlinear solving method
Optimizing the smoothing parameters minimizes whichever metric we choose ___
on the historical data
Overfitting
An overly complex model that tells you more about the idiosyncrasies of historical data without having value for future predictions
How do we avoid overfitting to pick among various candidate models?
- divide our data into training and testing data sets
- fit our candidate models based on the training dataset
- see how they perform on the testing dataset