Choosing approaches for forecasting issues in forecasting

Choosing approaches for forecasting issues in forecasting - Transcript

ECON50001 Business Forecasting
Exercise 10 Choosing approaches for forecasting
Object Discuss appropriate methods and issues to develop forecasts Data Files None Question 1

You have been employed as a consultant forecaster by the UK branch of a shipping company The shipping company transports parcels and packages around the world They want your guidance on forecasting three aspects of their business They want to be able to produce reliable forecasts for the next 3 years on the expected demand for their services which they can then use to plan the number of trucks labour etc they need to employ They want to be able to forecast what their costs will be in the future compared with today They also want to be able to forecast evaluate the successfulness of a new frequency shipper program they plan to launch whereby customers receive points for each parcel they send with the company and can then use these to send items for free or get other discounts

Discuss the forecasting methods that could be used to meet each objective any problems you may encounter and any data which you may need or data that you are likely to find

Forecast next 3 years on the expected demands for our services

Using Structural model The demands on our services could be affected by a number of reasons the price charged by our competitors the exchange rate and other factors which could affect the demand for sending goods overseas the quality of our services e g how many missing items issues happened in a year how fast we can delivery on average even how much we compensate our customers if any mistakes made by us how much the population within the UK over years etc there are a vast number of elements could affect the demands on our services over time To try to forecast the demands based on such amount of other variables may not seem plausible not only hard to get collect all the data which related to our variables but also hard to generate the appropriate model to do our forecasts Moreover it is not possible to forecast if there is any cyclical variables in the model especially if the cycle changes

Using Non structural model

Using Smoothing method could be one of the ways to forecast when there is limited data however apart from the models will be used in Smoothing models are restrictive the difficulty to choose the best values of parameters in the models would be hard to get the best forecasting So I would not use Smoothing method To forecast by using ARIMA model does not require as much data as using structural model do forecasts for the next 3 years on the expected demands for our services would not rely on any other things exogenous variables at all it might not be well explained by theory academically like the structural model do nevertheless it could provide a better forecasting result comparing with structural model and smoothing method in our case What do we do with ARIMA p d q is to generate a function that shows how current demands will be related by the previous demands on our services AR model or how current demands will be related by the previous demands of random errors MA model furthermore ARIMA model would be generated if needed just by putting them together So all I need from the company is the previous demands on our services recently in the past years and what I am exactly going to do with the adequate data would be Test for stationary and difference if necessary One of the fundamental conditions to conduct ARIMA p d q is to make sure the time series is stationary the values seem to fluctuate with constant variation around a constant mean We have to difference the non stationary series until it turns to be stationary maybe once twice or even three times if necessary We probably have to difference our time series because the demands on our services are more likely to be non stationary We will get what d is in this step Identify the right model s to be estimated In the sense of the smaller values of p and q the less possibility of our errors in our forecasting so we try to find the possible AR p MA q or ARMA p q with smaller values of p and q furthermore a number of plausible models would be selected for further steps to eliminate the inappropriate ones Estimate the model This is normally done by computer the aim is to adjust the values of p and q in our previous plausible models by testing their significance or overfitting The more accurate models would be appeared after this procedure Diagnostic checking In this step we check the residuals in each model that came from the constraints through steps above see whether they look random and well behaved Again this is nothing but eliminating the inappropriate model s that we thought it would be the right one s Using selection criteria There might more than one model left till this step however one final model which is appropriate for us to forecast would be left after this final selection criteria procedure

Pitfalls

1 We are forecasting the next three years so from next year onwards we are actually using our forecasted value to get the next forecast so the next two years forecasting are probably not as accurate as of next year errors roll over years 2 Any unexpected shock could happen such as another competitor joins leaves the market the expected demands would obviously go down and that would affect our forecasting 3 The ARIMA tells us nothing about what factors affect the demand in the future no theory used the company may want to know these things

Forecast for the future costs

There could be two ways to forecast our costs as well as demands above Using Structural model The costs of our shipping company consists of a number of components to forecast our costs in the future would be so complicated if we are considering every cost Such as the labour the renting of our offices the price of fuel for our trucks to delivery the parcel the space renting on flights if necessary etc we may have to generate a model that our total costs is the dependent variable and all the other costs are independent variables and then estimate an appropriate equation to get our forecast Again I am not adopting this method for reasons such as 1 It is difficult to find what each cost is so it is hard to put exactly all of them into our model 2 It is more difficult to get collect all the data of each independent variable in the past more costs 3 It is not possible to forecast if there is any cyclical variable s in the model Especially if the cycle changes 4 Apart from the UK we are still sending parcels overseas so the costs overseas are not easy to forecast The data needed from overseas are probably not feasible to get There are other more reasons to support that it is unwise to forecast our cost in the future using structural model so the method I would like to use is still non structural model

Using Non structural model Smoothing method would still be abandoned because simply it does provide the forecast as good as using ARIMA models By using ARIMA models the only data that I need from the company is the previous cost values as long as I have this information we should be able to forecast our future cost by using non structural model ARIMA p d q what I am trying to achieve is to estimate the relationship between the current value of costs and the previous costs AR model or the previous costs of random errors MA model or both of them ARIMA model The cost in the future comparing with today s would be forecasted by using the model estimated The next step I will do is the same as I forecast next three years demands on our services the methodology

still would be Box Jenkins and all the procedures have to be followed are exactly the same as showed in the previous question above Pitfalls 1 We don t know whether the record of cost values the company provided is trustable or not the wrong information could lead our forecasting to go wrong 2 The accuracy of our forecasted cost values in the near future are always more reliable than the values of a long time in the future So it is more appropriate for short term forecasting 3 Any unexpected shocks could make our forecasts go wrong such as inflation or massive drop in fuel price and so on



Forecast Evaluate the successfulness of the new program

Market survey Customer feedbacks survey could be created for both existing customers and potential market customers I would like to set up the degrees of satisfaction for getting such program up and running by collecting feedback from the market which should give an indication whether it would be a good decision to implement this project or not Analogy and precursor methods If we are a new shipping company just came into the market or less competitive than the others we could easily evaluate our new project by checking whether our strong rivals e g DHL had launched similar project or not why they are not doing so must be reasonable and we could probably take it as the signal of potential failure for launching it and we might consider to launch it if they are doing similar same program

Pitfalls 1 The inappropriate design of the market survey might not reflect the real opinions of the market customers 2 If we are a lot less competitive than the other rivals what they are doing probably for the long term gaining not profitable for shot term Therefore we might not imitate what they do for the time being till we are strong enough