Questions and Answers of Forecasting Predictive Analytics

how to write
how to learn
3. How easy is the package to learn and use?
1. Does the package have the facilities you want?computer platforms• 2. What platforms is the package available on?software ease-of-use
12.4 Table 12-2 in this chapter shows Armstrong's Forecasting Audit Checklist. Comment on the value of such checklist and discuss its usefulness in avoiding forecasting biases.
12.3 Where do you see the ¯eld of forecasting going in the future and what can you do to bene¯t from forthcoming changes in the environment?
12.2 Discuss the value of creating meaningful insights about the future when operating in a competitive environment where information is readily and instantly disseminated.
12.1 Describe what can and cannot be predicted while attempting to forecast the future.
10. Ability to monitor e®ectively current events and take corrective action if necessary and possible.
9. Psychological factors, market ine±ciencies, and plain accidents or mere coincidences.
8. Momentum, or inertia, that sustains established patterns and upholds existing relationships, at least in the short-term.
7. Natural events (e.g., a good or bad harvest) and their in°uence on the economic and business environment.
6. Some people's capabilities to control or slow down change (e.g., through monopolies or cartels).
5. Some people's wish to maintain the status quo.
4. Some people's ability to change the future in some favored way(e.g., through new technologies).
3. People's aspirations to change the future in some desired man-ner or direction.
2. People's preferences, tastes, and budget constraints.
1. Economic/market forces (e.g., the law of demand and supply)and biological laws (e.g., the S-shaped increases in growth) that generate and maintain various types of long-term equilibria.
11.3 The above mentioned competition includes methods that use neural networks, machine learning, and expert systems to make their predictions. Contrary to claims that such methods would outperform
11.2 In the above mentioned competition we still have not com-bined the various methods to compare their performance with those of the individual methods being combined. What is your opinion about
11.1 Table 11-12 shows the symmetric Mean Absolute Percent-age Error of 16 methods of the newest forecasting competi- M3-IJF Competition tion (M3-IJF) which compared 3003 time series. (The ¯rst
4. Models that minimize past errors: Available forecasting meth-ods select the best model by a process that depends upon how well a model minimizes one-step-ahead forecasting errors. How-ever, we
3. Unstable or changing patterns or relationships: Statistical mod-els assume that patterns and relationships are constant. This is rarely the case in the real world, however, where special events
2. Measurement of errors: No matter what we try to measure there are always errors of measurement (including clerical and data processing errors), the size of which can be substantial and systematic.
6. Finally, regression is used most often for medium-term, followed regression by long-term, forecasting horizons. This is consistent with theoretical reasoning that postulates that in the medium-
5. The Box-Jenkins method is not used very much for any fore- Box-Jenkins method casting horizons, which is consistent with empirical ¯ndings.
4. It is surprising that the straight-line projection method is used trend-line projection for short-term horizons. Given seasonality and cyclical factors, ups and downs in the short-term make
3. Exponential smoothing and moving average methods are used exponential more for short-term, less for medium-term, and even less for smoothing long-term horizons; this is consistent with empirical
2. Sales force composites and customer expectations are used less for the long-term and more for the medium- and short-terms.Overreliance on these two methods, and on the jury of executive opinion,
1. The jury of executive opinion is the most widely used fore- judgmental forecasts casting method; furthermore, its usage is uniform across all forecasting time horizons. Although this method has
6. Finally the satisfaction of users concerning expert systems and expert systems neural networks
5. It is somewhat surprising that classical decomposition does not classical fare better in Table 11-2. One reason might be that it is as decomposition much a tool for analysis as a forecasting
Table 11-2: Satisfaction with forecasting methods (as a percentage of those re-sponding). Source: Mentzer and Cox (1984).and apply while its post-sample accuracy (see below) is often not better than
4. Users were the least familiar and the most dissatis¯ed with the Box-Jenkins method Box-Jenkins method. This result is consistent with empirical¯ndings indicating that the method is di±cult to
3. The methods of moving average and trend-line analysis also trend-line analysis produced a high level of satisfaction. Furthermore, trend-line analysis had one of the smallest percentages of
2. The method with which users were next most satis¯ed is ex-ponential smoothing. This ¯nding is consistent with empir- exponential ical studies reporting that exponential smoothing is capable
1. Regression is the method users have the highest level of satisfac- regression tion with, despite empirical ¯ndings that time series methods are more accurate than explanatory (regression and
5. Classical decomposition is the second least familiar method. time series Only about one half of the respondents indicated any familiarity decomposition with this method, although it is one of the
4. The Box-Jenkins methodology to ARIMA models is the least ARIMA models familiar of the methods included in Table 11-1. The same is true with other sophisticated methods of which practitioners are
3. Moving average is the most familiar of the objective methods, although, from empirical studies, it is not as accurate as the method of exponential smoothing.
2. Users are also very familiar with the simpler quantitative meth-simple methods ods of moving averages, straight-line projection, exponential smoothing, and the more statistically sophisticated
1. Forecasting users are very familiar with the subjective (judg-judgmental forecasts mental) methods of jury of executive opinion, sales force com-posite, and customer expectations.
(d) the number of predictions required.
(c) the time horizon of forecasting;
(b) the characteristics of the time series;
(a) whether users want simply to forecast, or they also want to understand and in°uence the course of future events;
10.4 If you were to count, only once, from 0 to 100, how many times will you encounter the number eight?
Can we say how long it will be before it is interrupted by a recession?
10.3 The U.S. economy has been growing without interruption since May 1991. The longest previous expansion in the U.S.economy lasted for 105 months while the average post World War II expansion has
who are paid to choose the best stocks could not beat the broad market averages."\This year's failure|the ninth in the past thirteen years, according to Lipper Analytical Services Inc.|brings to mind
10.2 Glassman (1997) describes the reasons why various investment funds do not outperform the average of the market. Comment on the following quotes from Glassman's article:\For the third year in a
10.1 In 1996 the giant ¯rm of Philips Electronics lost $350 million on revenues of $41 billion. For the last ¯ve years Philips has been having great trouble modernizing itself and turning to
4. The bottom line is that very few newsletters can \beat"the S&P 500. In addition, few can beat the market fore-casts derived from a statistical representation of publicly available information.
3. Most of our tests focus on the ability of newsletters to call the direction of the market, that is, market timing.
2. Consistent with mutual funds studies, we ¯nd that poor performance is far more persistent than good performance.
1. Only 22.8% of newsletters have average returns higher than a passive portfolio of equity and cash with the same volatility. Indeed, some recommendations are remarkably poor. For example, the
9.4 In Johnson (1991), the following quote is said to have been made by Samuel Taylor in 1801. Comment on this quote.Samuel Taylor, of Fox & Taylor, °atly refused to believe that machinery could
9.3 Consider the price trends in computers and o±ce equipment.A computer whose speed was a fraction of a Mhz and which had only 8,000 words of RAM memory cost $10,000 in 1968($40,000 in 1997
9.2 Today one can buy a color photocopier which can also be used as a color computer printer, as a plain paper color fax machine and as a scanner. The price of this all-inclusive machine is under
(c) Today one can buy a PC running at 266Mhz and having 32 megabytes of RAM memory and 3 gigabytes of disk memory, plus many other powerful characteristics, for under $1,500. (In 1968 an electrical
(b) In 1997, an IBM computer program (Deep Blue) beat the world chess champion (Kasparov).
(a) There have been several computer programs developed recently that recognize continuous speech and turn it into written text in a computer word processing program.
9.1 Build a scenario about the future implications of the following new inventions:
(a) Cycles in copper prices and in cumulative random numbers; (b)Cycles in cumulative random numbers \0-200"; (c) Cycles in the S&P 500 and in cumulative random numbers; (d) Cycles in cumulative
(d) Give two reasons why you might want to use the state space form of these models rather than the usual form.
(c) Show that Holt's method (Section 4/3/3) can be written in the following \error feedback form":Lt = Lt¡1 + bt¡1 + ®et bt = bt¡1 + ¯®et where et = Yt¡Lt¡1¡bt¡1. Use this result to ¯nd a
8.10 (a) Write an AR(3) model in state space form.(b) Write an MA(1) model in state space form. (Hint: Set F = 0.)
(c) In no more than half a page, discuss the di®erences between this model and that considered in Exercise 8.7.You should include mention of the assumptions in each model and explain which approach
(a) Let Yt denote the log of the average room rate and Xt denote the log of the CPI. Suppose these form a bivariate time series. Both series were di®erenced and a bivariate AR(12) model was ¯tted
8.9 Consider the regression model ¯tted in Exercise 8.7 concern-ing the cost of tourist accommodation and the CPI.
(f) If the level of drug given varied from day-to-day, how could you modify your model to allow for this?
(e) Construct an ARIMA model ignoring the intervention and compare the forecasts with those obtained from your preferred intervention model. How much does the intervention a®ect the forecasts?
(d) Fit a new intervention model with a delayed response to the drug. Which model ¯ts the data better? Are the forecasts from the two models very di®erent?
(c) What does the model say about the e®ect of the drug?
(b) Fit an intervention model with a step function interven-tion to the series. Write out the model including the ARIMA model for the errors.
(a) Produce a time plot of the data showing where the intervention occurred.
8.8 The data in Table 8-8 are the daily scores achieved by a schizophrenic patient on a test of perceptual speed. The pa-tient began receiving a powerful tranquilizer (chlorpromazine)on the
(e) Forecast the average price per room for the next twelve months using your ¯tted model. (Hint: You will need to¯nd forecasts of the CPI ¯gures ¯rst.)
(d) Follow the modeling procedure in Section 8/2 to ¯t a dynamic regression model. Explain your reasoning in arriving at the ¯nal model.
(c) Produce time series plots of both variables and explain why logarithms of both variables need to be taken before¯tting any models.
(b) Estimate the monthly CPI using the data in Table 8-7.
(a) Use the data in Table 8-6 to calculate the average cost of a night's accommodation in Victoria each month.
8.7 Table 8-6 gives the total monthly takings from accommodation and the total room nights occupied at hotels, motels, and guest houses in Victoria, Australia, between January 1980 and June 1995.
(b) Now generate data with the same input data from Table 8-5 and the following transfer functions.4. r = 1, s = 0, b = 1 with !0 = 2:0 and ±1 = 0:7.5. r = 0, s = 2, b = 0 with !0 = 1:2, !1 =
8.6 An input (explanatory) time series Xt is shown in Table 8-5.(a) Using equation (8.5), generate three output time series Yt corresponding to the three sets of transfer function weights below.v1 v2
8.5 Sketch the graph of the impulse response weights for the following transfer functions:(a) Yt = 2(1 ¡ 0:5B)B2Xt(b) Yt =3B 1 ¡ 0:7B Xt(c) Yt =1 ¡ 0:5B 1:2 ¡ 0:8B Xt(d) Yt =1 1 ¡ 1:1B + 0:5B2Xt.
(d) If the methane input was increased, how long would it take before the carbon dioxide emission is a®ected?
(c) What are the values of the coe±cients !0, !1, !2, ±1, ±2,µ1, µ2, Á1, and Á2?
(b) What sort of ARIMA model is used for the errors?
(a) What are the values ofb, r, and s for the transfer function?
8.4 Box, Jenkins, and Reinsell (1994) ¯t a dynamic regression model to data from a gas combustion chamber. The two variables of interest are the volume of methane entering the chamber (Xt in cubic
(d) Explain why the Nt term should be modeled with an ARIMA model rather than modeling the data using a standard regression package. In your discussion, com-ment on the properties of the estimates,
c) Describe how this model could be used to forecast elec-tricity demand for the next 12 months.
Explain what the estimates of b1 and b2 tell us about electricity consumption.
(b) The estimated coe±cients are Parameter Estimate s.e. Z P-value b1 0.0077 0.0015 4.98 0.000 b2 0.0208 0.0023 9.23 0.000µ1 0.5830 0.0720 8.10 0.000Á12 ¡0:5373 0.0856 -6.27 0.000Á24 ¡0:4667
(a) What sort of ARIMA model is identi¯ed for Nt? Explain how the statistician would have arrived at this model.
8.3 Electricity consumption is often modeled as a function of temperature. Temperature is measured by daily heating de-grees and cooling degrees. Heating degrees is 65±F minus the average daily
(b) Using the ACF and PACF of the errors, identify and estimate an appropriate ARMA model for the error.Write down the full regression model and explain how you arrived at this model.
(b) Plot the ACF and PACF of the errors to verify that an AR(1) model for the errors is appropriate.8.2 (a) Fit a linear regression model with an AR(1) proxy model for error to the Lake Huron data
(a) Re¯t the regression model with an AR(1) model for the errors. How much di®erence does the error model make to the estimated parameters?

Showing 1 - 100 of 506