Real Estate Modelling and Forecasting Hardcover_15

Real estate forecasting in practice 427 an expert makes an adjustment to the forecast driven by future employ-ment growth, this adjustment is based on a less efficient use of the his-torical relationship between rent and employment growth. The expert should direct his/her efforts towards influences that will genuinely add to the forecast. When the forecasts from a model and expert opinion bring different kinds of information and when the forecasts are not cor-related, it is beneficial to combine them (Sanders and Ritzman, 2001). (2) Track record assessment. Purely judgemental forecasts or adjusted model forecasts should be evaluated in a similar manner to forecasts from econometric models. The literature on this subject strongly suggests that track record is important. It is the only way to show whether expert opinion is really beneficial and whether judgement leads to persistent outperformance. It provides trust in the capabilities of the expert and helps the integration and mutual appreciation of knowledge between the quantitative team and market experts. Clements and Hendry (1998) assert that the secret to the successful use of econometric and time series models is to learn from past errors. The same approach should be followed for expert opinions. By documenting the reasons for the forecasts, Goodwin (2000a) argues that this makes experts learn from their past mistakes and control their level of unwarranted intervention in the future. It enables the expert to learn why some adjustments improve forecasts while others do not. As Franses (2006) notes, the best way to do this is to assess the forecasts based on a track record. Dotheexpertslookathowaccuratetheirforecastsare,though?Fildes and Goodwin (2007) find that experts are apparently not too bothered about whether their adjustments actually improve the forecasts. This doesnothelpcredibility,andhenceitisimportanttokeeptrackrecords. (3) Transparency. The way that the forecast is adjusted and the judgement is produced must be transparent. If it is unknown how the expert has modified the model, the forecast process is unclear and subjective. 13.6 Integration of econometric and judgemental forecasts The discussion in section 13.2 has made clear that there are benefits from bringing judgement into the forecast process. As Makridakis, Wheelwright and Hyndman (1998, p. 503) put it: ‘The big challenge in arriving at accu-rate forecasts is to utilize the best aspects of statistical predictions while exploiting the value of knowledge and judgmental information, while also capitalizing on the experience of top and other managers.’ The potential benefits of combining the forecasts are acknowledged by forecasters, and 428 Real Estate Modelling and Forecasting thisleadstothesubjectofhowbesttointegratemodel-basedandjudgemen-tal forecasts. The integration of econometric and judgemental forecasts is a well-researched topic in business economics and finance. In summary, this literature points to different approaches to integrating econometric forecasts and judgemental views. A useful account of how the forecasts are combined is given by Timmermann (2006). (1) Mechanicaladjustmentstothestatisticalforecast.Theforecastteammayinves-tigate whether gains can be made by mechanical adjustments to the model’s forecasts in the light of recent errors. For example, one such procedure is to take part of the error in forecasting the latest period (usually a half of the error) and add that to the forecast for the next period. Consider that a model of retail rents based on consumer spend-ing has over-predicted rent growth in the last few periods (fitted above actual values). This could be due to intense competition between retail-ers, affecting their turnover, that is not captured by the model. We mechanically adjust the first forecast point by deducting half the error of the previous period or the average of the previous two periods and perhaps a quarter of the error of the following period (so that we lower the predicted rental growth). A prerequisite for this mechanical adjust-ment is, of course, our belief that the source of the error in the last few observations will remain in the forecast period. Vere and Griffith (1995) have found supportive evidence for this method but McNees (1986) has challenged it. (2) Combining judgemental and statistical forecasts produced independently. Aside from mechanical adjustment, another approach is to combine experts’ judgemental forecasts with the estimates of a statistical method pro-duced separately. It is assumed that these forecasts are produced inde-pendently; if the parties are aware of each other’s views, they might anchor their forecasts. This approach appears to work best when the errors of these forecasts take opposite signs or they are negatively cor-related (note that a historical record may not be available), although it is not unlikely that a consensus will be observed in the direction of the two sets of forecasts. A way to combine these forecasts is to take a straightforward average of the judgemental and econometric forecasts (see Armstrong, 2001). More sophisticated methods can be used. If a record of judgemental forecasts is kept then the combination can be produced on the basis of past accuracy; for example, a higher weight is attached to the method that recently ledtomore accurate forecasts. As Goodwin(2005) remarks, a large amount of data is required to perform this exercise, which the real estate market definitely lacks. Real estate forecasting in practice 429 Goodwin also puts forward Theil’s correction to control judgemental forecasts for bias. This also requires a long series of forecast evaluation data.Theil’sproposalistotakeanexpert’sforecastsandtheactualvalues and fit a regression line to these data. Such a regression may be yield = 2 +0.7 ×judgemental yield forecast In this regression, yield is the actual yield series over a sufficiently long period of time to run a regression. Assume that the target variable yield refers to the yield at the end of the year. Judgemental yield forecast is the forecast that was made at, say, the beginning of each year. When making the out-of-sample forecast, we can utilise the above regression. If the expert predicts a yield of 6 per cent, then the forecast yield is 2% +0.7 ×6% = 6.2% Goodwin (2000b) has found evidence suggesting that Theil’s method works.Itrequiresalongrecordofdatatocarryoutthisanalysis,however, and, as such, its application to real estate is restricted. Goodwin (2005) also raises the issue of who should combine the forecasts. He suggests that the process is more effective if the user combines the forecasts. For example, if the expert combines the forecasts and he/she is aware of the econometric forecasts, then the statistical forecast can be used as an anchor. Of course, the expert might also be the user. For further reading on this subject, Franses (2006) proposes a tool to formalise the so-called ‘conjunct’ forecasts – that is, forecasts resulting from an adjustment by the expert once he/she has seen the forecast. (3) The ‘house view’. This is a widely used forum to mediate forecasts and agree the organisation’s final forecasts. The statistical forecasts and the judgementalinputarecombined,butthisintegrationisnotmechanical or rule-based. In the so-called ‘house view’ meetings to decide on the final forecasts, forecasters and experts sit together, bringing their views to the table. There is not really a formula as to how the final output will be reached. Again, in these meetings, intervention can be made based on the practices we described earlier, including added factors, but the process is more interactive. Makridakis, Wheelwright and Hyndman (1998) provide an example of a house view meeting. The following description of the process draws upon this study but is adapted to the real estate case. The house view process can be broken down into three steps. Step 1 The first step involves the preparation of the statistical (model-based) forecast. This forecast is then presented to those attending the house view meeting, who can represent different business units and seniority. 430 Real Estate Modelling and Forecasting Participants are given the statistical forecasts for, say, yields (in a partic-ularmarketoracross markets).Thisshouldbeaccompanied byanexpla-nation of what the drivers of the forecast are, including the forecaster’s confidence in the model, recent errors and other relevant information. Step 2 The participants are asked to use their knowledge and market experi-ence to estimate the extent to which the objective forecast for the yield ought to be changed and to write down the factors involved. That is, the participants are not asked to make a forecast from scratch but to anchor it to the objective statistical forecast. If the team would like to remove anchoring to the statistical forecast, however, individuals are asked to construct their forecast independently of the model-based one. In their example, Makridakis, Wheelwright and Hyndman refer to a form that can be completed to facilitate the process. For yield forecasts, this form would contain a wide range of influences on yields. The statistical model makes use of fundamentals such as rent growth and interest rates to explain real estate yields, whereas the form contains fields pointing to non-quantifiable factors, such as the momentum and mood in the market, investment demand, liquidity, confidence in real estate, views as to whether the market is mis-priced and other factors that the participants may wish to put forward as currently important influences on yields. This form is prepared in advance containing all these influences but, of course, the house view participants can add more. If a form is used and the statistical forecast for yields is 6 per cent for next year, for example, the participants can specify a fixed percentage per factor (strong momentum, hence yields will fall to 5 per cent; or, due to strong momentum, yields will be lower than 6 per cent, or between 5.5 per cent and 6 per cent, or between 5 per cent and 5.5 per cent). This depends on how the team would wish to record the forecasts by the participants. All forecasts have similar weight and are recorded. Step 3 The individual forecasts are summarised, tabulated and presented to participants, and the discussion begins. Some consensus is expected on the drivers of the forecast of the target variable over the next year or years. In the discussions assessing the weight of the influences, the participants’ ranks and functional positions can still play a role and bias the final outcome. All in all, this process will result in agreeing the organisation’s final forecast. At the same time, from step 2, there is a record of what each individual said, so the participants get feedback that will help them improve their judgemental forecasts. Real estate forecasting in practice 431 Figure 13.1 Forecasting model intervention 2007 2008 2009 Under the category of ‘house views’, we should include any other interactive process that is not as formal as the three steps described above. Indeed, this formal process is rare in real estate; rather, there is a simpler interaction in the house view process. This informal arrangement makes it more difficult to record judgemental forecasts, however, as the discussion can kick off and participants may make up their minds only during the course of the meeting. The outcome of the house view meeting may be point forecasts over the forecast horizon. It may also be a range of forecasts – e.g. a yield between 5.5 per cent and 6 per cent. The statistical forecast can be taken as the base forecast around which the house view forecast is made. For example, assume a statistical forecast for total returns over the next five years that averages 8 per cent per annum. The house view meeting can alter the pattern of the model forecasts but, on average, be very close to the statistical forecasts. Furthermore, point forecasts can be complemented with a weighted probability of being lower or higher. This weighted probability will reflect judgement. Giventhedifferentwaystointerveneinandadjustmodel-basedforecasts, a way forward is illustrated in figure 13.1. The value for 2007 is the actual rent growth value. The model-based forecasts for 2008 and 2009 are given by the plain triangle. In all probability these forecasts will not be entirely accurate, as the error will incorporate the impact of random events, and the actual rent growth values for 2008 and 2009 could be either of the two shaded triangles – that is, the actual rent growth will be higher or lower than predicted by the model. Expert judgement can come in two ways to modify this forecast. (1) By weighting additional market information, a probability can be given as to which direction the actual value will go. In the figure, such a judge-ment may suggest that, based on market developments not captured by the model, there is a greater probability that rent growth will be lower ... - tailieumienphi.vn