Forecasting

Dr. Everette S. Gardner, Jr.

Judgment exercises
Exercise 1
Finished files are the result of years of scientific study combined with the experience of years.
How many times does the letter F appear in the sentence above? Count them only once; do not go back and count them again. ________ How confident are you in your answer? Rate your confidence on a scale of 0 to 100, where 0 means that you are sure you are wrong, and 100 means that you are sure you right. _________

Forecasting

Judgment exercises (cont.)

Exercise 2
Threatened by a superior enemy force, the general faces a dilemma. His intelligence officers say his soldiers will be caught in an ambush in which 600 of them will die unless he leads them to safety by one of two available routes. If he takes the first route, 200 soldiers will be saved. If he takes the second route, theres a one-third chance that 600 soldiers will be saved and a two-thirds chance that none will be saved. Which route should he take? ________

Forecasting

Judgment exercises (cont.)

Exercise 3
The general again has to choose between two escape routes. But this time his aides tell him that if he takes the first, 400 soldiers will die. If he takes the second, theres a one-third chance that no soldiers will die, and a two-thirds chance that 600 soldiers will die. Which route should he take? _______

Forecasting

Judgment exercises (cont.)

Exercise 4
Linda is 31, single, outspoken, and very bright. She majored in philosophy in college. As a student, she was deeply concerned with discrimination and other social issues, and participated in anti-nuclear demonstrations. Which statement is more likely? a. Linda is a bank teller. b. Linda is a bank teller and active in the feminist movement. _______

Forecasting

Human biases in forecasting

Company Politics
Forecast what the boss wants to hear

Overconfidence
Confidence has no relation to accuracy

Wishful thinking
Optimistic forecasts more probable

Success/failure attribution
Good forecasts due to skill, bad due to chance
Forecasting

Human biases in forecasting

Gamblers fallacy
Bad luck and good luck will balance out

Data presentation
Misleading graphs/tables easily accepted

Conservatism
Refusal to accept drastic change

Forecasting

Forecasting methods
Human judgment
Subject to bias and inconsistency Models usually beat humans

Time series forecasting

Based on analysis of past history Cheap and easy On average, most accurate method Should always be attempted

Forecasting

Forecasting methods (cont.)

Regression modeling
Based on causal relationships Expensive and difficult Must forecast independent variables

Growth or market development models

Based on assumed growth patterns Cheap and easy Difficult to validate

Forecasting

Can your data be forecasted by a model?

Use common sense
Abrupt turning points usually impossible to predict

Compare your accuracy to a nave benchmark

Forecast for next period is the same as the data this period If you cannot beat a nave benchmark, forecasting is usually futile
Forecasting

10

Forecast profiles
Nonseasonal Constant Level Linear Trend Exponential Trend Damped Trend
Forecasting

Additive Seasonality

Multiplicative Seasonality

11

Simple exponential smoothing

(1) Error in t = Actual data Forecast for t (2) Forecast for t+1 = Forecast for t + (Error in t) To get started: Set first forecast equal to mean of first few data. Smoothing weight (): In practice, is usually 0.30 0.50. Effects of extreme values: If = 0, the forecast never changes. If = 1, this is a nave or random walk model.
Forecasting

Simple.xls

12

Error measures for evaluating forecast models

MAD = Mean absolute deviation (error)
MSE = Mean squared error MAPE = Mean absolute percentage error

Forecasting

13

Smoothing a linear trend

(1) Error in t = Actual data - Forecast for t (2) Level at end of t = Forecast for t + h1(Error in t) (3) Trend at end of t = Trend at end of t1 + h2(Error in t) (4) Forecast for t+1 = Level at end of t + Trend at end of t

To get started, set: Initial trend = Average growth in first four data Initial level = First data observation initial trend
Search for weights in the following ranges: Level weight (h1) 0.10 to 0.90, increments of .10 Trend weight (h2) 0.05 to 0.30, increments of .05
Forecasting

Trendsmooth.xls

14

The general trend model

(1) Error in t = Actual data Forecast for t (2) Level at end of t = Forecast for t + h1(Error in t)

(3) Trend at end of t = (Trend at end of t 1) + h2(Error in t) (4) Forecast for t+1 = Level at end of t + (Trend at end of t)
Long-term forecasting: Forecast for t+2 = Forecast for t+1 + 2(Trend at end of t) Forecast for t+3 = Forecast for t+2 + 3(Trend at end of t) Trend possibilities: If < 1, the trend is damped. If = 1, the trend is linear. If > 1, the trend is exponential. If = 0, there is no trend (same as simple smoothing).
Forecasting

15

Starting up the general trend model

Initial values
Initial trend = Average growth in first four data Initial level = First data observation Initial trend

Search for parameters in the following ranges

Level weight (h1) Trend weight (h2) Phi () 0.10 to 0.90 0.05 to 0.30 0.60 to 1.00

Forecasting

16

Multiplicative seasonality
The seasonal index is the expected ratio of actual data to the average for the year. Actual data / Index = Seasonally adjusted data Seasonally adjusted data x Index = Actual data

Forecasting

Multimon.xls

17

Multiplicative seasonal adjustment

1. Compute moving average based on length of seasonality (4 quarters or 12 months).

2.
3. 4. 5. 6. 7. 8.

Divide actual data by corresponding moving average.

Average ratios to eliminate randomness. Compute normalization factor to adjust mean ratios so they sum to 4 (quarterly data) or 12 (monthly data). Multiply mean ratios by normalization factor to get final seasonal indexes. Deseasonalize data by dividing by the seasonal index. Forecast deseasonalized data. Seasonalize forecasts from step 7 to get final forecasts.
Forecasting

18

Additive seasonality
The seasonal index is the expected difference between actual data and the average for the year. Actual data - Index = Seasonally adjusted data Seasonally adjusted data + Index = Actual data

Forecasting

Additmon.xls

19

Additive seasonal adjustment

1. Compute moving average based on length of seasonality (4 quarters or 12 months).

2.
3. 4. 5. 6.

Compute differences: Actual data - moving average.

Average differences to eliminate randomness. Compute normalization factor to adjust mean differences so they sum to zero. Compute final indexes: Mean difference normalization factor. Deseasonalize data: Actual data seasonal index.

7.
8.

Forecast deseasonalized data.

Seasonalize forecasts from step 7 to get final forecasts.
Forecasting

20

Forecasting simulations
Dynamic simulation
Short-range (one-step-ahead) forecasting test Use data in fit periods to select model During forecast periods: 1. Make one forecast. 2. Observe error. 3. Adjust model. 4. Go to 1.

Static simulation

Long-range forecasting test Use data in fit periods to select model Make all forecasts at once
Forecasting

21

Data transformations for forecasting

Deseasonalize
Isolates trend

% Change
Isolates trend

Natural log
Converts exponential trend to linear

Square root
Reduce variance

Aggregate
Quarterly or monthly data to annual
Forecasting

22

Forecasting management
Organize for forecasting
Pinpoint responsibility Only one corporate forecast Separate forecasting and planning Choose a standard measure Keep a track record Benchmark Hold performance reviews

Monitor accuracy

Scrub the data

Adjust outliers Throw out unique data

Forecasting

23

Forecasting management (cont.)

Compare alternative forecasts
Top-down vs. bottom up Monthly vs. quarterly data Deseasonalized vs. raw data Percent change data Time series forecasts Regression forecasts

Simulate forecasting
One-step-ahead Long-range

Estimate confidence limits

What is the range of forecast errors in past?
Forecasting

24