استفاده از الگوریتم های ژنتیکی و تجزیه و تحلیل رگرسیون خطی برای پیش بینی تقاضای مسکن خصوصی

An accurate prediction of prospective construction supply and demand, especially the private residential market, is paramount important to policy makers, as it could help formulate strategies to cultivate/stabilize the economy and satisfy the social needs (at macro level). Despite that, a realistic prediction of future private residential demand is never an easy task, as it is governed by a number of social and economic factors. In this paper, four leading indicator models are developed and compared for directly forecasting Hong Kong private sector residential demand. These comprise a (i) Linear Regression Analysis (LRA) model, (ii) Genetic Algorithms (GA) model, (iii) GA-LRA model, where LRA is used to select the indicator variables; and (iv) GA-LRA model with Adaptive Mutation Rate (AMR) to reduce the likelihood of local optima. The findings indicate that the GA-LRA model with AMR provides the most accurate forecasts and over a longer time horizon. In providing a range of possible forecasts, the model also provides an opportunity for the decision-maker to exercise judgment in selecting the most appropriate forecasts.

Precise estimation of demand for new residential properties is never a simple task, as it could be influenced by a number of dynamic factors, viz., demographic change, economic pattern, government policy and external environment [1]. While many major cities are confronted with a shortage of public housing and a soaring private property price [2], it is usually the government's responsibility to formulate suitable long-term housing strategies and policies to regulate and accommodate the housing needs of different sectors such that a sufficient amount of land and housing units are available to satisfy the demand. In order to make housing policy decisions, it is first necessary to estimate both short-term and long-term future housing demand. As the housing stock is relatively inelastic in the short run, an overly conservative prediction in the housing demand could result in a shortage of residential supply. However, no one would ever imagine an overly optimistic housing forecast could also lead to profound effects to the locality especially on the overall economy. Recent example in Hong Kong (HK) has illustrated that a surplus supply of residential units had an inverse relation to the price of real estates (the property price in HK plummeted by almost 60% between 1998 and 2003). Reliable estimation of new residential property not only concerns policy makers, planners and home purchasers/tenants, but could also determine the survival of many companies related to the construction sector [3]. Despite its strategic significance, little research has been carried out to enhance the methods for predicting the residential demand. In some cases, estimations are made according to a projection of flats required for new households (e.g. new marriage, divorce, new immigrant, etc.) and existing families (e.g. those affected by redevelopment programs). Surely, demographic change would have significant implication to the housing demand, yet one should not ignore the impacts of economic change on the desire of property purchase [4]. According to Hillebrandt [5], the effects of economy on construction occur at all level and in all aspects of economic life, hinting that the economy (e.g. income, interest rate, etc.) may somehow influence the demand for residential properties, especially on private housing. This paper reports on a comparison of four leading indicator models for forecasting HK private sector housing supply (as a proxy for demand) directly. These comprise a (i) Linear Regression Analysis (LRA) model, (ii) Genetic Algorithms (GA) model, (iii) GA-LRA model, where LRA is used to select the indicator variables; and (iv) GA-LRA model with Adaptive Mutation Rate (AMR) to reduce the possibility of local optima. The findings suggest that the GA-LRA model with AMR provides the most accurate forecasts and over a longer time horizon. In providing a range of possible forecasts, the model also provides opportunity for the decision-maker to exercise some judgment in selecting the most appropriate forecasts.

By comparing four models for forecasting private sector housing demand, it was found that: (i) the LRA method is easier to operate and the amount of time required to build a model is shorter; (ii) the GA method involves some tedious calculations and the aid of a computer is necessary when the problem is complex or the search space is large; (iii) the GA method allows the decision-maker a larger involvement, such as in assigning the GA parameters and choosing the appropriate results from the pool of solutions; (iv) the GA method can generate more than one solution each time, e.g., in this case 50 solutions were generated each time since the population size is 50; (v) the GA method has a better accuracy; and (vi) the GA method has a longer prediction period. An important point is that the GA program needs to be properly parameterized to avoid reaching local optima and converge to a global optimum with a high degree of consistency, regardless of the specification of the initial population. On the other hand, the program can spend a considerable amount of time without showing improvement, and then suddenly produce a jump. It still not yet clear, however, quite how to do this parameterization, what kind of problems the GA is most suited for, what controls its convergence rate, and what precisely are the roles of crossover, mutation, etc., in the overall search in progress. There is growing evidence that the “optimum” parameters values may be problem-specific—no general methodology being presently available to optimize the selection of these parameters. Only general experience shows that the value of the crossover ratio (Pc) is usually 0.6–0.8; while for the mutation rate (Pm) the expected number of bits mutated per chromosome should be kept less than one. Similarly, setting the convergent ratio c at 0.6 has been found to avoid either reaching a local optima or taking too long to converge. In addition, as with any form of prediction or forecast, many uncertainties and errors exist. In this case, they may be due to: (i) Lack of significance of economic indicators: For this study, a total of 10 economic indicators were available for constructing the model but the GA-LRA model benefited from the selection of only 6 of these. This is presumably because the omitted indicators fail to, or spuriously, represent the economic factors involved. For example, the reason the real wage index was not selected by the LRA for use in the GA-LRA model may be that this index does not really reflect the income and purchasing power of HK people, as the actual income for many high-income people is mainly from returns on their investments rather than their salary. (ii) Interdependence of variables: In constructing the forecasting model, it is assumed that the economic indicators, which serve as the variables, are independent. In fact, all these indicators are related to different aspect of the economic conditions of HK and are therefore, by their very nature, likely to be highly interdependent. (iii) Change in economic indicators: In making predictions using the leading characteristics of the chosen economic indicators, it is assumed that these indicators will follow a similar pattern or trend in the whole period under consideration. If there is an abrupt change in the indicators, the prediction may fail. Abrupt changes to the indicators can easily happen due to: • Policy: Changes in economic policy often have a significance effect on economic conditions, even in the construction industry. • Housing habits: An increasing number of HK people are now buying houses in Mainland China—a trend that is very difficult to be shown by an economic indicator and therefore reflected in the models constructed. • Economic structure: The HK economic structure has changed very rapidly in recent years. This has resulted in significant changes in land use and redistribution of property and therefore the general economic cycle. There is no guarantee, therefore, that the economic indicators will follow the same cycle. Finally, it should be noted that only some basic aspects have been explored here, however, and there is considerable potential for future study by • carrying out more systematic tests to optimize the parameters—particularly those of population size, the fitness evaluation function, the crossover and mutation rate; • using ‘best subset’ instead of stepwise regression to identify suitable indicator variables; • further tests on the AMR; • using the niche formation and modified AWA instead of Roulette wheel selection; • further investigation on the search space beyond that of stimulation from the ‘all possible repressor’ method; and • using further stopping criteria, such as when the program does not show significant improvement for certain number of generations, or when the results generated have achieved a certain satisfactory level, instead of just 300 iterations or 1000-AMR iterations. Developing the GA method beyond that of merely replacing the R2 method in constructing the mathematical model—it could be used in other aspects of the problem as well, such as to determine which set of economic indicators would produce the best result or to study the leading characteristics of the candidate economic indicators. Although there is no theory yet to support such a replacement, some empirical tests could be made to gauge the usefulness of this approach.