The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
scowaiETRie AtmoACH . ,« - to m w A s n w . ^IV^d^ Eetaorks of m m m j . m i s EX. HeciJxsr fto&rd ot Governor* of the Federal Reserve System at the Eighth Amual Forucats ting Conference of the Aonavicau Statist leal Assoc leticm W w York* Hew tiyck April 29, mm. I was asked to discuss an econometric approach to lorecasting housing trends. three related points? For this purpose* t want to touch briefly (I) an exposItion of the econometric madeI * have used to forecast housing starts', (2) a projection of start* 'or this year based upon the mxSel; and (3) the record of the model ** the recent past. M g W f o r e c a s t i n g Jfeyftig m m Figure I list© two closely related, statistically estimated ^aationo for housing starts. These equations follow the nearly identical form of a model of the housing market which I developed six yaars ago and which was discussed at length in the June S s s x ^ ^ M M * The equations clearly meet the normal statistical tests goodness of fit as well as of significance for the specific independent variables, 13 to 30. ^efficient. the F-rafcios for the individual variables range The equations also show a significant Durban-Watson It is recognised, however, that because of the inclu- sion of lagged values of the dependent variable, this is not a accessary proof of lack of autocorrelation, the coefficient values of the same order of magnitude as in previously published versions this models they differ somewhat, however, because rather massive corrections have occurred in the underlying data and because one t r i a b l e (household formation) now appears in the model only indirectly through its influence on vacancies. -2- Tlte M R article describe® the logic of this caxiel et sosie length, but we can briefly restate it. In effect, this tnodel stress*>s the existence of large scale Inventory swings ©round a more stable level of final demand. starts It hold* that the level of housing determined by four factor®! (1) First is the general level of final housing demand, this depends upon the rate of household formation end the rate at fctiich houses are removed from the existing housing stock. (2) The rate of construction also depends upon the level of available vacancies. When units are completed at a faster rate than the increases in final demend, a backlog of vacancies is built up. xhe available vacancies depress the rate of starts. (3) The rate of inventory accumulation fluctuates widely fcith builders' expectations and dasires to build houses. The desire to build is influenced by current interest rates, as veil as by the *elatlonship between rents and costs. As these cause builders• expectations and profits to change, starts vary. (A) There is a technical relationship between the rate of starts end the number of units in the pipeline under construction. The inventory under construction does not wove in a simple manner with starts, ftather the relationship takes a form which can be expressed as a difference equation. The relationship between changes in the inventory under construction and the rate of change in atarts tends to follow a fluctuating form. Figure 1 ftl»o shows the expected elasticitiea or reactions of starts* to changes in the Independent variables. are given. Two elasticities The first ahovs the iraraediate iapact on starts of clinages whether in vacancies, *nteresfc rates, or relative coots. The second takes into account their total irapacts including their influence cm the inventory under construction. ay the end u£ a year & one-percent increase in vacancies lowers the level of starts by about four-tenths of a percent. A °ae^>ercent increase in costs relative to rents would decrees© the level of starts by about two percent. A one-percent increase in the mortgage interest rate would also cause a similar two-percent in sterts, but its ij$>»ct would take a half-year longer to be fully effective. ffroe* the actual coefficients, we see thet if there are 100,000 additional vacancies, the level of starts in the following year will be reduced by sbout 40,000. fc his reduces future vacancies. Other things being equal, As a result, the excess in vacancies and its depressing iopact on starts virtually disappear aver a twoto three-year period. Alterations between strength and weakness ftom vacancies have been typical of the markets. m x m mortgage interest rotes vary, each ten bests point® movement, according to the uKaJel, causes start a to alter by from 55 to 50 f 000 a year, The interest variable overages rates for nine taonths. On the average, the aodsl estimates that it takes a change in Intercast rates three-quarters of a year to influence starts. From top to bottom of an interest rat© wove, the equations show an impact <m starts of 200 to 300,000 units at an annuel rate. The data underlying the rent-cost indexes are among the most suspect of all. Changes in this relationship over a cyclical period have ranged from two and one-half to five percent. Xhus, their estimated impact on starts is shc*ra to have influenced the level of starts by from 50 to 100,000 with the changes occurring «ver a two-year period. figures 2 and 3 are charts which show the estimated impact of each of the independent variables on housing starts since 1950. They alao show the relationship of the estimated totals to the actual. She charts indicate that the total impact of movements in each of the individual variables has normally ranged between plus and minus 150,000 starts. It usually takes several years for an individual variable to alter by that much. On the other hand, as several variables have moved together, the annual rate of starts has varied rapidly. There have been at least four major fluctuations in starts since 1950. A r«viev of the charts give® ua some pause in making too literal an interpretation of the coefficients. Quarterly movement* in the r e p o r t s series. W e note the erratic There Is some indica- tion that the calculated total® ©ay be more reliable than the reported la many periods. On the other hand, the sharp differences between the cuvemttts of vacancies end the interest rate between the early mid late 1950*a causes concern. Some of the estimated Glesticitiea may reflect a movement in trend® and not in the shorttim variations. L ^ x m ^ fry Ms*d on the model and current forecasts of the independent Variables, I estimate that private starts in 1966 will be approximately 1.4 million. This is roughly 100,000 less than last year. Thin projection is baaed on an averaging of several different for*a of the model. Each individual equation &ivea a slightly different teault. What leads to this expected drop? Basic demand from house- hold formation and net removals would be expected to increase this year's starts over last by approximately 40,000 units. On the othsr hand, a fairly rapid rise in coats related to rents is expected to decrease starts by a similar 40,000. Increasing interest rates also will, according to the model, depress starts by about 40,000. higher level of vacancies carrying over as a result of the large number of starts in W 6 3 and «arly '$4 has a negative impact of The about 3O,0tH>* These are the initial decreases. Sine® fulling starts require fewer units la the construction pipeline, the inventory under construction is expected to fall by roughly 30,000* The total »f these diverse movements leads to the e x a c t e d decrease of approximately 100,000 units. UUcjtt this model was run at the and of October, it projected 1*460,000 start* for the year. Xhus, the projected level of start® ha* fallen by 60,000 over the past five months. This is true even though the actual level of starts is as yat running close to the Noveetoer projection. The expected deterioration later this year occurs because it is now assumed that interest rates sod relative costs will both reach higher levels then were predicted for our initial run last fail. ftecord of ScmuMfrrta Since tills ia a faceting of the American Statistical M s o c l stlon dedicated to the improvement of our statistical techniques, it appears proper for us to consider the record and p r o b l e m of economic Projections of this type* Since developing this model, I have as a matter of principle attempted to go on record with m y forecasts. X believe that the best nay to test a forecasting model ia to see how it works. figure 4 contains Information cm forecasts made from this modal in four previous yesrs. In two of these years, the Initial forecast was revised after a month as additional information became available. -7- l have also included for each of these years published cstitaates froe* the Business and Defense Service® Administration of the Department of C o a m r c e . I have selected their projections simply because they ere eesily available in published font. I have no idea 'whether their record is better or verse than that of other fore* casters. X do note* however, that in the last four years their pro* tactions have called for minimum changes of 1/2 to 1-1/2 percent a year. looking at the record* I come to no firm conclusion as to Whether or not the econometric model is doing well. In two years the projections were very close; one was moderately off; and one very bad. correct* Xn three of the four years, the direction of change this is a considerably better record than Commerce's fchoae direction was wrong in three of the four years. The cooperi- *one show a standoff caaparing percentage errors by years—each was better twice* the average percentage error for the four years was percent for €<amree and 5*4 per cent for this model, or a plus in favor of the non-econometrie model. Frankly, I have no way of determining whether an average ®eaa error of 3.4 percent for a fairly volatile series such as housing starts is good or not. considerably better. The median error et 3.5 per cent is % would Judge that if the median performance stayed et this level, the record of this forecasting procedure could he considered more than satisfactory. The reasons for the variance ere fairly clear. In the first place, it should be recognised that some variance is to be axpeefced. The standard error of forecast for tile current model is l*/,900 starts per quarter. Assuming that it has been roughly the *aoe in the past periods, the actual forecast fell within the standard error on an annual basis in two years and was very close 1ft one* Another major source of error in the forecasts is the Poor underlying data. They have undergone several major revisions. The first two forecasts were made from data prior to their revisions. It ie not clear how much of the error came from this fact. Perhaps one should be surprised that the model did as well as it did, given the basic changes in the data. Another point to recognise ie that in use, this particular requires that several items be projected. Thus for this year, we have to project the mortgage Interest rate and the cost elements through the third quarter. Errors in projecting these exogenous variables end up as part of the total variance in the final fore- cast. My general feeling about the model, outside of a nonaal pride of authorship, is that it has been a useful esercise in modelbuilding and forecasting. As a result of this particular model, we have a much better concept of how the housing market works than we had before its development. tion. Others are building upon tills founda- can expect that w e r time considerable improvements will be saaito* ftvea in its present sfcete3 however, X find it worth white to go through and recalculate the model each time n forecast L& required. Given the else of the standard error of forecast the other difficulties with our information, I believe that the model au*t still (k* used as part of « coordinated, over~sli analysis o£ the housing market rather than a® c unique estimate. For ayanpis, if for policy purposes I required a specific estimate at this time, 1 might adjust this year's projection up somewhat. On the other hand, because a specific model exists, do have some idea® of the orders of magnitude of impects on starts thet can be expected m fates, or vacancies. a result of changes in cost®, interest Without a model of this sort, it is extremely difficult to give any auaoricel content to the idea that these ttoveraenta in these variables will alter the rat© of housing starts, tfhile I do not place a high reliability on the specific numbers involved in estimating the effects of changes in these variables, I do feel that as long as the model continues to forecast fairly Veil, we must place some confidence in the specific numbers that it throws out. Figure 1 ALTERNATIVE HOUSING STARTS EQUATIONS 1,7 '-7698 , -.0690 V_ x +.5363 St -.2754 St_ 3 +.6100 -4 + 2.467 T ( ~^ (•1542) (.0126) (.1001) R 2 - .869 (.0715) (.1196) (.4465) S u = 17.46 •k -k -k -k -k -k -2 1.°027 -.0858 Vs^+,5336 S t ^ -.3148 St_ 3 +.5956 ^ ^ <•2039) (.0153) (.1025) R 2 = .866 (.0732) (.1201) + 3.166 Rem Q (.6534) S u = 17.65 •k -k -k -k k -k -ELASTICITIES OF STARTS IN RELATION TO INDEPENDENT VARIABLES First Period coeff. elasticity Mean Long Run coeff. elasticity 1383.3 - .086 - .322 - .110 - .412 980.3 .596 1.582 .763 2.026 552.0 -1.003 -1.499 -1.284 -1.919 •k -k -k -k k k 0 K •s B y housing s t a r t s , q u a r t e r l y totals in t h o u s a n d s , s e a s o n a l l y a d j u s t e d , U ^ e a u of the C e n s u s , C o n s t r u c t i o n R e p o r t s - - H o u s i n g S t a r t s , S e r i e s C - 2 0 . >Uab V ^ c* * V a c a n c i e s d e r i v e d e n d o g e n o u s l y from b a s e of a v a i l a b l e v a c a n c i e s in U S e r SCerU S < - S . B u r e a u of the C e n s u s . For V s , total a v a i l a b l e v a c a n c i e s are r e d u c e d *t of the accumulated change in households since 1950. Bu V r e n t component of the consumer price index (1957-59 = 100). of Labor Statistics. y 8f (] 5 ^f B „ 5®»ldential construction--GNP implicit price deflator residential construction St n ° n ^ C h l n t e r e s t on c o n v e n t i o n a l f i r s t m o r t g a g e s for the p u r c h a s e of n e w , o n e - f a m i l y , Vj. °Uses. F e d e r a l H o m e L o a n B a n k B o a r d . Mh* *« 1 7 > 5 0 0 p l u g . 1 8 0 p e r c e n t of dwelling units available at beginning of FIGURE 2 400 200 + 0 200 1953 1955 1957 1959 1961 1963 1965 200 100 100 200 400 200 200 400 1953 1955 1957 1959 1961 1963 1965 FIGURE 3 Figure 4 SUMMARY OF PRIOR FORECASTS (In thousands at annual rates) Projection Amount 62 D e c.62 62 Error Actual change Change* Amount Percent Amount Percent Amount Percent Directic 1300 1350 + 78 + 97 + 6.0 + 7.7 + 82 + 51 + 6.3 + 3.9 - 4 +46 - 0.3 + 3.5 correcl correci 1963 1963 1963 1296 1359 1459 -178 - 97 - 15 -12.1 + 7.2 + 8.4 + 7.2 -293 -230 -130 -18.4 -14.5 - 1.0 + 115 +133 + 115 wrong wrong wrong 1964 1964 1450 1605 -165 -10.2 + 1.0 69 45 4.5 2.9 1965 1965 1965 1420 1480 1552 -117 - 27 - 7.6 + 11 + 0.7 34 4 38 2.3 0.3 2.5 - 32 -144 - - 6.7 - 8.2 Hi N °V.63 + 16 + 61 - 5.6 + 4.0 correci wrong - 83 - 23 + 49 - 5.5 - 1.5 + 3.3 correci correci wrong - 86 •n it 64 c.64 Nov 64 in v 1460 1398 1515 «65 s + 22 - 1.8 2.1 - 9.3 + 1.5 i x months' data available at time of forecast.
@FedFRASER