1. CHAPTER13: FORECASTING

2. 13.1 INTRODUCTION Typical business forecasting situations –A company wishes to forecast the sales of its products – Forecast the returns resulting to the company from the purchase of new equipment. –A local authority forecasts the number of children for the next ten years –The Treasury has a large economic model that allows the investigation of the likely effects on the economy if the Chancellor changes the income tax rate, or alters the interest rate.

3. 13.1.1 Approaches to Forecasting If the company has available the monthly sales figures for its products for the previous twelve months then this information can be used to make a forecast of sales for the next month.

4. To forecast the sales for the next three time points: projecting the sales trend line.

5. projecting the sales trend line is not so simple Intuitively any forecast made from this data would be less reliable.

6. Time-series method –Use historical data collected over time and use this data to project forward to make a forecast Other methods of forecasting –For local authority example, to predict the number of couples within age bands, the birth rate for each age band hence the forecast number of children as required. –The Treasury has a large econometric model that allows the investigation of the likely effects on the economy if the Chancellor changes the income tax rate, or alters the interest rate.

7. 13.1.2 Time-Series A time-series may be formally defined as: –A set of observations made on a particular variable at equidistant time intervals. Some examples of time-series: –The sales data used in the two examples above. –The number of people recorded as unemployed at the end of each month. –The daily closing price for a company shares quoted by London Stock Exchange –The temperature of a hospital patient recorded on an hourly basis. Measure of the accuracy of the forecast

8. 13.1.3Time-Series Graphs Time-series plot: –A visual inspection : useful information about the nature of the time-series. –well-defined trend –seasonal structure. EXAMPLE 1: –well-defined trend having little variability about the trend. –give relatively precise forecasts. –forecasts for time points 13, 14 & 15 –measure of the forecast accuracy for different forecasting methods

9. EXAMPLE 2 –more problematical –forecasts produced from time-series data: less reliable. –forecasts for time points 13, 14 & 15 –The measure of forecast accuracy in this situation would suggest the forecasts were not very reliable. Forecasting method: –calculating the forecast for each required time point – calculating measure of forecast accuracy

10. 13.1.4Exponential Smoothing Methods: Methodology for exponential smoothing is based on intuitive ideas, –a set of 'custom and practice methods' rather than having a well defined underlying theoretical structure. Exponential smoothing model –simple exponential smoothing model –model to deal with time-series that contain a trend –Model to deal with time-series that contain both trend and seasonality.

11. 13.2 THE SIMPLE EXPONENTIAL SMOOTHING MODEL Exchange rate between Pound Sterling and German Mark To forecast the exchange rate for time periods 12, 13 & 14 Quarter1234567891011 Ex.Rate2.952.972.942.963.013.022.982.962.942.972.88

12. No well-defined trend or seasonal variation Using simple exponential smoothing model

13. –This type of time-series data is described as a stationary time-series. – For a stationary time-series the forecast for the next time point is the average value of the 'time-series variable' over the length of the series. –The estimate of the Exchange Rate at time point 12 Simple average Weighted average

14. 13.2.1 A Notation & the simple EWMA relationship A common abbreviated notation for this time-series: –X t, t=1,2,…,n. The estimate of the level made on the basis of the previous t observations is labelled as M t, then a weighted average can be calculated as follows: M t = a t X t + a t-1 X t-1 + a t-2 X t-2 +…+ a 1 X 1 a t + a t-1 + a t-2 +…+ a 1 =1

15. Simple average –a t = a t-1 = a t-2 =…= a 1 =1/t weighted Average –Heavier weighting is given to more current data points a t >a t-1 > a t-2 >…>a 1 a t-j =  (1-  ) j j=1,2,3… 0≤  ≤1 M t =  X t +  (1-  ) X t-1 +  (1-  ) 2 X t-2 +  (1-  ) 3 X t-3 +…

16. –The series ,  (1-  ),  (1-  ) 2,  (1-  ) 3,  (1-  ) 4 : an exponential series (geometric series) –M t =  X t +  (1-  ) X t-1 +  (1-  ) 2 X t-2 +  (1-  ) 3 X t-3 +… –M t =  X t + (1-  )[  X t-1 +  (1-  )X t-2 +  (1-  ) 2 X t-3 +…] – M t =  X t + (1-  )M t-1 –This is the basic exponential smoothing equation The estimate at time t = a proportion of the new information +one minus that proportion of the estimate at time t-1.

17. 13.2.2 Forecasting with the Simple model For X t, t=1,2,…,n, –M t =  X t + (1-  )M t-1 –Calculate M 2 using t=2 –Calculate M 3 using t=3 –Calculate M 4 using t=4 –Calculate M n using t=n – The forecast of the value of X n+1 = M n Two problems for the process –How to start the calculation –How to choose a value for , the smoothing constant.

18. 13.2.3 How to start the calculations, A Starting Rule Starting rule –let M 1 = X 1 –Let  =0.25 Calculating

19. MtMt 2.95 2.96 2.95 2.97 2.98 2.97 2.95 t Calculation 12.95M 1 =X 1 22.97M 2 = 0.25X 2 + 0.75M1 32.94M 3 = 0.25X 3 + 0.75M 2 42.96M 4 = 0.25X 4 + 0.75M 3 53.01M 5 = 0.25X 5 + 0.75M 4 63.02M 6 = 0.25X 6 + 0.75M 5 72.98M 7 = 0.25X 7 + 0.75M 6 82.96M 8 = 0.25X 8 + 0.75M 7 92.94M 9 = 0.25X 9 + 0.75M 8 102.97M 10 = 0.25X 10 + 0.75 M 9 112.88M 11 =0.25X 11 +0.75M 10 Forecasting is the last value of M, (M 11 ) is 2.95, this is a weighted moving average based on all the previous 11 time-series data points. X 12 =M 11 =2.95

20. The graph of the data and the M t series is given in below:

21. Define F t (1) to mean the forecast made on the basis of the time-series data values, X 1, X 2, X 3,... X t of the next value of the time- series, X t+1 –For the simple exponential smoothing model the forecast function is F t (1) = M t –F 11 (1) = M 11 =2.95 Define F t (2) to mean the forecast made on the basis of the time-series data values, X 1, X 2, X 3,... X t of the value of X t+2. –F t (2) is called the two step forecast. –F t (2) = M t – F 11 (2) = M 11 =2.95 (forecast value of X t+2 )

22. F t (h) means the h step forecast made on the basis of the previous t time- series points. –F t (h) = M t –F 11 (3) = M 11 =2.95 (forecast value of X 11+3 ) –F 11 (4) = M 11 =2.95 (forecast value of X 11+4 )

23. 13.2.4 Measuring Forecast Precision F t (1) = M t –F 10 (1) = M 10 –F 9 (1) = M 9 –F 8 (1) = M 8 Common Measures of forecast precision: –Mean Absolute Deviation –Mean Square error –Mean Percentage Error

24. Mean Absolute Deviation –MAD =  |E t |/n – Exponentially weighted MAD MAD t =  |E t | +(1-  ) MAD t-1 Mean Square Error –MSE=  (E t ) 2 /n Mean Percentage Error –MPE=  (|E t | /X t )*100/n Example: –MAD = 0.030 –MSE = 0.002 –MPE =1.02%

25. 13.2.5 How to choose a value for , the smoothing constant: M t =  X t + (1-  )M t-1 –IF  =0, M t = M t-1 =M t-1 =…=X 1 For a very small(near 0) value of  we get very heavy smoothing, very little weight is given to the new data, and a heavy weighting given to the history of the series. –IF  =1, M t = X t For large values of  (near 1) a high weighting is given to the current data and very little to the past history of the series.

27. Smoothing constant  determines the level of smoothing. –A small value of  gives heavy smoothing –A large value of  gives less smoothing – In practice the value of  used to make a forecast represents a trade-off between these two extremes.

28. Guidance for smoothing parameter –a)The value of  should be in the range 0.05 to 0.3 (suggestion by C D Lewis) Choose  small if a plot of the series suggests a stable series. Choose  large if a plot of the series suggests a more dynamic series. –b)Choose the value of  to minimise one of the measures of forecast precision.

29. 13.2.6 Building a Spreadsheet Model of the EWMA Model The Conceptual Paper Worksheet:

30. The choice of  : –Using 'Table' command on Excel i. Estimate the value of  that gives the smallest MSE. ii. Enter this estimated value of  into cell Dl. iii.The forecasts for time period 12 13 & 14 are read from cells D15, D16 & D17. –Using Solver command on Excel  =0.667825 Final spreadsheet model(  =0.7) as following table

32. –Mean Square Error by definition is the average squared error, as such is measured in squared units, this does not make for sensible interpretation. The Root Mean Square Error, RMSE, which is the square root of the MSE is in the correct units. For this data RMSE =  0.00149 = 0.0386. –£1 = 2.91 Marks, with a RMSE = 0.0386, the implication being that the forecast error is likely to +/- 0.04 Marks

33. 13.3 EXPONENTIAL SMOOTHING MODEL WITH TREND A product inventory level at the end of week over the last 25 weeks: 1234567891011 1213 140159136157173131177188154179180160182 141516171819202122232425 192224188198206203238228231221259273

34. This time-series exhibits a definite upward trend.

35. 13.3.1 The Exponential Smoothing Model with Trend If assuming the trend is locally linear, at time t the level and rate of change of level, (the slope) is known, X t+1 = Level(t) + Slope(t) + error –Level(t) =M t –Slope or Gradient at time t: R t = M t – M t-1 –The estimate at time t = proportion of the new information + one minus that proportion of the estimate at time t-1,

36. –Estimate of the level at time t =  *new information + (1-  )*Estimate of the level based on time t-1 information –M t =  X t + (1-  )(M t-1 + R t-1 ) –R t =  (M t -M t - 1 ) + (1-  )R t - 1 one step forecast: M t + R t two-step forecast: M t + 2R t –h-step forecast: F t (h) = M t + hR t –Using this model to forecast presents the following problems: a) A starting rule is required, initial values for M 1 and R 1 need to be estimated. b)Values for the two smoothing parameters  and  need to be specified.

37. 13.3.2 The Starting Rule and smoothing parameters a)The starting Rule. The simplest starting rule is to fit a straight line to the first few data points. This can be done by fitting a straight line by eye to the time-series graph and measuring the intercept and slope.

38. –When t = 1 the value of inventory is 146 (as estimated from the graph) –When t = 11 the value of inventory is 176 (as estimated from the graph) –  Y/  X = (176-146)/(11-1)=30 –M 1 =146 –R 1 =30

39. Choice of smoothing parameter –a)Choose the values of  and  according to advice offered by experienced users: Woodward & Goldsmith 4, suggest values of  = 0.1 and  = 0.01. –b) Choose the values of  and  to minimise one of the measures of forecast precision. M 1 =146, R 1 =3,  = 0.5 and  = 0.5, to give the following spreadsheet calculations

41. 13.3.3 Forecasting with the trend model The one step forecast, the forecast for time point 26 is:263.37 + 12.92 The two step forecast, the forecast for time point 27 is:263.37 + 2*12.92 the forecast function for the time point h steps ahead is: F t (h) = M t + hR t at all time points –F 1 (1) = M 1 + R 1 –F 2 (1) = M 2 + 2R 2

42. The forecast for the time periods 26, 27 & 28: –F 25 (1) = M 25 + R 25 =263.37 + 1*12.92 =276.29 –F 25 (2) = M 25 + 2R 25 = 263.37 + 2*12.92 = 289.22 –F 25 (3) = M 25 + 3R 25 = 263.37 + 3*12.92 = 302.14

43. 13.3.4 A forecasting process –a)Choose the values of  and  according to advice offered by experienced users: Woodward & Goldsmith 3, suggest values of  = 0.1 and  = 0.01. –F 25 (I) = M 25 + R 25 =238.72 –F 25 (2) = M 25 + 2R 25 = 241.89 –F 25 (3) = M 25 + 3R 25 = 245.06 –MSE.=367.01.

44. –b)Choose the values of  and  to minimise one of the measures of forecast precision.  = 0.07 and  = 0.99 to minimise MSE –F25(l) = M 25 + R 25 =259.24 – F25(2) = M 25 + 2R 25 = 268.09 – F25(3) = M 25 + 3R 25 = 276.93 –MSE=367.01

45. –Summary of the forecasts using differing smoothing parameters: ( ,  ) F 25 (1) F 25 (2)F 25 (3)M.S.E. (0.5, 0.5)276.29289.22302.14455.59 (0.1,0.01) 238.72241.89245.06367.01 (0.07, 0.99)259.24268.93276.93270.00

46. Conceptual Paper Worksheet

48. 13.4 Summary - The simple exponential smoothing model applies to time-series where the time- series plot shows no evidenceof trend or seasonal factors. - The basic recurrence relationship is: -M t =  X t + (1-  )M t-1 -The forecast function is: F t (h) = M t - The exponential smoothing model with trend applies to time-series where the time-series plot shows a marked underlying trend - The basic recurrence relationships are: - M t =  X t + (l-  )(M t-1 + R t-1 ) - R t =  (M t -M t-1 ) + (l-  )R t-1 - The forecast function is: F t (h) = M t + hR t

49. - All exponential smoothing models require:- a) A starting Rule b) A choice of smoothing parameter a) For the simple model the simplest starting rule is: M 1 =X 1 a)For the trend model the simplest starting rule is to fit a line to the first few data points, and from this line estimate the value of Mt and Rt b) Choice of smoothing parameter i) By experience ii) By minimising a measure of precision.

50. -Forecasting -For h=1,2,3,…. F t (h) = M t - Forecasting -For h=1,2,3,…. F t (h) = M t + hR t - The measures of forecast precision i) Mean Absolute Deviation –MAD =  |E t |/n ii) Mean Square Error –MSE=  (E t ) 2 /n iii) Mean Percentage Error –MPE=  (|E t | /X t )*100/n

51. Group Work Collect the daily closing price for anyone company shares quoted by Shenzhen Stock Exchange or Shanghai Stock Exchange, and Use Exponential Smoothing Model(Simple or With Tread) to forecast the closing price of Next Day. Compare your forecasting closing price and the actual closing price of Next Day. Remarks 1) Work in the group ( Total 10 groups) 2) Preparing PPT document and Excel spreadsheet model, and selecting 2-3 representatives of the group by yourself, and making the presentation of your results in the next class (Monday, 17 March, 2008). 3) The presentation time for each group is 15 minutes.