Some A Posteriori Probabilities in Stock Market Action Alfred Cowles 3rd, Herbert B, Jones Econometrica, Volume 5, Issue 3 (Jul., 1937), 280-294, Stable URL: bttp//links jstor.org/sici?sict=0012-9682% 28 193707%295%3A3%3C280% 3ASAPPIS%3E2,0,CO%3B2-Q ‘Your use of the ISTOR archive indicates your acceptance of JSTOR’s Terms and Conditions of Use, available at hup://uk,jstor.org/abouvterms.himl, JSTOR’s Terms and Conditions of Use provides, in part, that unless you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you ‘may use content in the JSTOR archive only for your personal, non-commercial use. Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the sereen or printed page of such transmission. Econometrica is published by The Econometric Society. Please contact the publisher for further permissions regarding the use of this work, Publisher contact information may be obtained at hitp://uk,jstor.org/joumalsieconosoc. html Econometrica ©1937 The Econometric Society ISTOR and the ISTOR logo are trademarks of ISTOR, and are Registered in the U.S. Patent and Trademark Office For more information on JSTOR contact [email protected] ©2002 JSTOR hup//uk.jstozorg/ Mon Nov 25 13:04:12 2002SOME A POSTERIORI PROBABILITIES IN STOCK MARKET ACTION By Aurrep Cowues 3x0 ap Huxsenr E. Jones Ty 1983, one of the authors' published an analysis of the results se- cured by 45 representative financial agencies in forecasting the prices of common stocks. This study, embracing the period of 4} years from January, 1928 to June, 1932, and including several thousand individual, forecasts, showed that these were unsuccessful more often than suc- cessful and that, indeed, better results in the aggregate would probably have been secured by investors through following purely random vestment programs. ‘This result naturally suggested the questior stock price action random in nature or, if not, to what extent is it, possible statistically to define the nature of its structure? Herewith is presented at least a partial answer to this question on the basis of evi- dence adduced from internal elements in the stock price series them- selves. Among the different viewpoints from which this problem may be approached are those which consider (1) the element of inertia, and (2) harmonie analysis for the purpose of disclosing evidence as to regu- lar periodicity. With respect to the latter, Professor Harold T, Davis has presented evidence of periodicity in stock prices by means of Schuster’s periodo- gram analysis, and has attempted tests of the significance of these pe- riods by R. A. Fisher’s technique. By this method, however, only aver- ‘age periods can be found and even then their significance ean not be accurately determined because of uncertainty as to the independence of observations. Variable periodicity ean be taken into account by a num- ber of techniques, such as Frisch's changing harmonies,’ or by using varying moving averages,¢ but there seems to be no way to determine the probabilities of significance in connection with any of these ‘methods. For this reason in the following analysis a method more sus- ceptible of interpretation in terms of probability has been used. Carl Snyder,¢ in referring to measurements of the industrial growth ' Alfred Cowles 3rd, “Can Stock Market Forecssters Forecast”, Ecoxo- asmie4, Vol. 1, 1983, pp. 309-324 In a paper presented before the Eoonometrie Society in St. Louis, January, 1936. See abstract in Ecoxomernica, Vol. 4, 1986, pp. 189-190, * Ragnar Friseh, “Changing Harmonies and Other General Types of Compo- nents in Empirical Series," Skondinavish Aktuarietiderift, 1928, pp. 220-238. “Gerhard Tintner, Prices In the Trade Cycle, Vienna, 1935, p. 23. + “Concepts of Momentum and Inertia,” presented at the Atlante City meet- Ing of the Eeonometrie Society, December, 1982, and published in Stabilization of Employment, Charles F. Roos, editor, 1983, p. 77 280ALFRED COWLES AND HERBERT E. JONES 281 of the United States since 1880, says, “The picture that these measures, zzive is that of an amazingly even rate of growth not merely from gen- eration to generation but actually of each separate decennéum through- ‘out the last century. As if there was at work a kind of momentum or inertia that sweeps on in spite of all obstacles.” ‘The term “inertia,” as employed by Mr. Snyder, may be said to be macroscopic, whereas its use in the present study is microscopic. He was concerned with the trend over a period of 100 years or more and concluded that all deviations from that trend were nothing more than inconsequential “jigeles.” ‘The present analysis, on the other hand, is concerned primarily with evidence as to inertia in movements of a few hours, days, weeks, months, or years, which Mr. Snyder, with his longer-range viewpoint, would designate as mere “jiggles.” Evidence of inertia may be disclosed in the following manner. In a penny-tossing series there is a probability of one-half that tails will follow heads and vice versa. If the stock market rises for 1 hour, day, week, month, or year, is there a probability of one-half that it will de- cline in the succeeding comparable unit of time? In an attempt to an- swer this question sequences and reversals were counted, a sequence occurring when a rise follows a rise, or a decline a decline, and a re- versal occurring when a decline follows a rise, or a rise a decline, A study of the ratio of sequences to reversals will disclose structure as defined above, if it exists within the series, and the significance of this structure can be defined by ordinary statistical methods. For stance, the probability can be determined that any ratio occurred by chanee from a random population.* Also, the consistency of these ratios can be investigated and from their frequeney distributions one can de- termine the probabilities of success in forecasting a rise or decline in stock prices. Samples, of adequate length where available, were ex- on, “On the Comparison of Meteorological Data with Results of translated and abridged by E. W. Woolard, Monthly Weather Review, Vol. 48, 1920, pp. 89-94, pointed out that in a random series, the ratio of se- quences to reversals will be 0.5, that is there will be twice ns many reversals as sequences. It should be noted that, in'the present analysis, the data employed fre not random, hut rather cumulated random serie, that is series in which the first differences, rather than the actual observations, are random. Ina truly ran- ddom series the auto-correlation drops to zero when the series is Ingged against itself by s0 much as one observation. Ina cumulated random series thisis not the case. For example, in the stock price series under consideration, even when the frst differences have been rearranged in a random manner, an auto-correlation with a lag of one observation will yield a very high coefficient. This correlation Coeicient will bea function of the length ofthe series and is approximately equal to 1—log n/(n—1). In such a series, with the frst differences random, the ratio af sequences to reversals will be, if mis large, 10 instend of 0.8 as observed in the case of s series in which the observations themselves are random.282 ECONOMETRICA amined, the intervals between observations being suecessively 20 min- utes, I hour, 1 day, 1, 2, and 3 weeks, 1, 2,3, ~~, 11 months, and 1, 2,3, ++, 10 years. The results of this investigation are presented in Table 1 and are shown graphically in Figure 1. Tt was found that, for every series with intervals between observa- tions of from 20 minutes up to and including 8 years, the sequences out numbered the reversals, For example, in the ease of the monthly series ‘Tate 1 Ramo oF Sequaxcxs ro Revensats ix Srock Paice INpexes 1 [recfazetoee] uur [ waex [reno eae unit | ivoex | pem100 fra [a | we isa |e vow || a ween [ee oon |e as ewos siie| ea ‘The Index of Railroad Stock Prices is composed of several series which were assembled and put together by the Cleveland Trust Company. Canal atock prices ‘were used for the period 183i through 1833. From 1834 through 1879 the index is based on three Harvard soris. The one from 1834 through 1852 includes eight ‘and that from 1853 through 1865 includes 18, The data are from The Review of Economie Statistics for August, 1928. The index from 1866 through 1870 includes 10 stocks, and the data are from The Review of Beonomie Statistice {for 1919. The index from 1880 through 1896 includes 10 stocks, and {rom 1807 to date it includes 15. Theso two latter indexes were compiled by the Cleveland ‘Trust Company. All the earlier indexes were adjusted to form a continuous series terminating with the final Index of 18 stocks. from 1835 to 1935, a total of 1200 observations, there were 748 se- quences and 450 reversals. That is, the probability appeared to be .625 that, if the market had risen in any given month, it would rise in the succeeding month, or, if it had fallen, that it would continue to decline for another month. ‘The standard deviation’ for such a long series + In a random penny-toasing series the probability of a sequence or reversal isALFRED COWLES AND HERBERT . JONES 283 constructed by random penny tossing would be 17.3; therefore the deviation of 149 from the expected value 599 is in excess of eight times the standard deviation. The probability of obtaining such a re- sult in @ penny-tossing series is infinitesimal. If the unit of time be in- creased to 6 months, we find that there are 120 sequences" to 86 reversals or what appears to be a .583 probability that a sequence will e pair of periods. The probal Sar Frouns 1—Ratio of sequences to reversals in direction of stock”price indexes for various time intervals, .01778 that such a ratio of sequences to reversals might occur in a random series such as that of penny tossing previously referred to. For annual series, the probability for chance occurrence, on the same basis as before, is 42052. In fact, the probabilities are inconclusive for all series using units of over 6 months, although this may be due to the limitations of the data. ‘There seems to be @ good chance that with more data it might be possible to demonstrate the existence of struc- 4} and thestandard deviation is npg or v/n/2 where n is the number of observa- tions and p and q are respectively the probabilities of success and failure, Se quences oF reversals are determined from the first differences of the original price series. The total numberof sequences and reversal is one less than the number of ‘The standard deviation of the number of sequences, therefore, will be /n—Z/2 where n is the number of observations of the original price (n—2)/2 where Sis the observed number )/21/[(n—2)8/2] isthe observed deviation in terms of the standard deviation. The prot of as large a deviation oc- fourring by chance ean be found in the ordinary table of the normal probability function284 ECONOMETRICA ture for all units of time up to and including 3 years. It is difficult, how- ever, to explain the very small excess of sequences over reversals for series using intervals of 2 weeks and 8 weeks in view of the fact that for the slightly shorter unit of 1 week we have signifieant indication of structure and also for series using intervals of 1 month. In fact, a8 will be shown later, the series represented by units of 1 month proves to be the most significant from a praetieal point of view. ‘The above analysis was based upon a study of the stock market as a whole, It may now be of interest for us to examine the evidence with regard to industrial groups such as motors, oils, stecls, and so forth. ‘To this end the indexes of common stock prices of 61 industrial groups, prepared by the Standard Statistics Company, were analyzed. Their monthly deviations from the median were noted, the median being uused beeause more groups are normally below, than above, the ari metic means, in view of the fact that extremely large gains are larger in percentage than extremely large declines. Sequences and reversals ‘were counted for the purpose of determining whether there was a tend- eney for such groups to persist in exceeding, or falling below, the me- dian, In other words, ifthe oil stocks, as a group, were in the 30 of the 61 industrial groups which advanced more than the median group in January, what is the probability that the oils will also be found in the strongest 30 groups in February? For the 16 months from January, 1934 to April, 1935, when the general market movement was approxi- mately horizontal, in 917 observations there were 570 sequences and 345 reversals. That is, if the oil stocks were among the strongest 30 groups in January the probabilities would appear to be .623 that they ‘would also be found among the strongest 30 groups in February. The application of the theory of probability to interpret the significance of this result is hampered, as in many other analyses of economic time series, by uncertainty as to what is the number of independent observa- tions in the sample. The action of the oils, for example, may be eorre- lated with that of the motors or some other group, s0 that when one is stronger than the median, the other also tends to be stronger. Modern statistical technique appears to offer no ready solution for this problem of the independenee of observations, and we must, therefore, eontent ourselves with obtaining unusually favorable probabilities. In the period, May, 1935 to February, 1936, a rising market of 551 observations, there were 379 sequences and 170 reversals, indicating an ‘apparent probability of .690 in favor of a sequence. Here, however, an- other factor has intruded itself. The stocks of certain industries ean be shown to be more cyclical in nature than those of other industries. For instance, in severe depressions the building of new houses is almost ‘completely stopped, and yet people under such conditions go on eatingALFRED COWLES AND HERBERT E, JONES 285, almost as much food as in periods of prosperity. These tendencies are reflected by wider fluctuations in the earnings of producers of building ‘materials than in the ease of purveyors of food, and also by wider eyeli- cal fluctuations in the stocks of the former corporations than of the latter. This tendeney results, in the ease of a rising market, in the building-stock prices being persistently stronger than the average, and the food stocks persistently weaker. A eount of sequences and reversals under such conditions measures, therefore, to some extent, the differ- ‘enees in cyclical behavior among the various groups rather than the tendency toward inertia which is measured in the case of a period where the market as a whole is moving horizontally. Sinee there are not many periods of great length in which the market as a whole has moved horizontally, we are limited in the data available for this par- ticular analysis. Considering only the data for the horizontal 16 months from January, 1934 to April, 1935, the excess of sequences over re- versals is 7.5 times the standard error for a random series, Even as- suming that half of the observations are not independent, there is still the exceedingly small probability of .00168 of ocourrenee in a random ‘The action of individual stocks was also investigated. Instead of the oils as a group, it was considered, for example, whether the Standard Oil Company of New Jersey, if it were stronger than the median of all stocks in January, would more likely than not be stronger in February. ‘Taking 190 representative stocks for the years 1934 and 1935 and the first three months of 1936, and using the same technique as that em- ployed in the ease of the industrial groups, a total of 4659 observations, there was found, when the market moved horizontally, from January, 1934 to May, 1935, a ratio of sequenees to reversals of 1.07 to I which is 9 times the standard error. In the last 12 months, a rising period, the ratio of sequences to reversals was 1.29 to 1. It follows from the pre- vious discussion of cyclical behavior that the difference between the two periods is to be expected and that the evidence of the second period must be brought in question, ‘Taking 1 year as the unit of measurement for the period from 1920 to 1935, the tendency is very pronouneed for stocks whieh have ex- ceeded the median in one year to exceed it also in the year following, In 1837 observations were 1200 sequences and 635 reversals. The excess of sequences was about 13 times the standard error for a random series constructed on the basis of equal probabilities. During the period un- der consideration the market as a whole manifested 8 sequenees and 7 reversals. The evidence therefore seems to indicate that, when 1 year is the unit of time, individual stocks will persist in doing better, or worse, than the median of all stocks.286 ECONOMETRICA ‘This evidence of strueture in stock prices suggests alluring possi ties in the way of forecasting. In fact, many professional speculators, including in particular exponents of the so-called “Dow Theory” widely publicized by popular financial journals, have adopted systems based in the main on the principle that it is advantageous to swim with the tide, ‘The practicability of such forecasts, however, will depend, ‘Taoe 2 Ansotors Prncensacz Cuanans 1s Stock Puce INDEXES wor | er | rs Se wows a | en | m= | =e ee oc zs iltesilast om Set forsee [e Shee sie not only on the ratio of sequences to reversals, but also on the broker- ‘age costs and the average change in stock prices during the unit of time selected. The brokerage costs, of course, are known. To determine the average percentage change in the stock prices for various units of time fan extensive study was made. The difference between the index at the beginning of one unit and the beginning of the next, given in pereentage of the former value, was computed. The results are shown in Table 2. ‘These averages are absolute values, that is, the directions of the moves, were not considered. The values, plotted on # time scale, are shown in Figure 2. The increase is very regular, approximating # sinooth expo- nential curve as shown in the diagrat With these data it is possible to compute the net gain whieh would have resulted from an application of this type of foreeasting to the stock market averages.ALFRED COWLES AND HERBERT B. JONES 287 1(0) = Expected annual net profit, in per cent, Rd) = Ratio of sequences to reversals for time interval ¢, C(®) = Average change per time interval ¢ in per cent, Y(@) = Number of time intervals, t, in one year, B = Brokerage cost for one complete trade, in per cent. In the long run there will be an average of R(t) sequences for each roversal, that is, a speculator will, on the average, be on the right side of the market for (i) units of time for each unit that he is on the Frouns 2—Absolute percentage changes in stock price indexes ‘wrong side, if his market position is changed only after the occurrence of each reversal. The average net time in the right direction between changes of position will then be [R(‘)—1] time units. The average move per unit of time is C({); therefore, the gross gain per position will then be [R()—1]C(®). The net gain per position will be [R@ —1]-c@ — B. Since one will be in the market in the right direction R(®) units of time and in the wrong direction 1 unit, the total time per position will be [R(® +1]. ‘The number of positions taken per year will be Y(/{R@ +1). The net annual gain, in per cent, will then be given by Irconmcoen 10 = of [a@ 1Jew - B} is + ‘| = 100.288 ECONOMETRICA When the values of (®) and C(#) are substituted in this equation values of the expected annual net gain for various units of time and brokerage costs are obtained as shown in Table 3. Tame 3 Ratio of | Average Sequences | Pereeat | ————————_—____ ‘Time Unit ge co 1 day. 1 week 1 month 2 months 3 months ‘As might be expected, the daily and weekly units are too short. The probability of success is not sufficient to compensate for the fact that the changes per unit of time are small relative to brokerage costs. The average net gain per trade is largest for units of 2 months but, because with this unit so few trades are completed in a year, the annual net gain is less than when units of one month are used. It appears, indeed, that, for the period under consideration, one month is the optimum unit of time. At this point a study was undertaken to determine whether it would be possible to select a group of stocks which would continue to be more volatile than the average and whether, if this could be done, the re- sults of speculation, employing such a list, would be more successful than where average stocks were used. An investigation was made cov- ering the period from 1900 to 1919. The number of stocks examined ‘was 44 at the beginning of the period, increasing to 81 by 1918. The 10 per cent manifesting the greatest movement for 1 year, in the direction of the market, was chosen as the volatile group. It was found that such 1 group was about 1.7 times as volatile as the market for 6 months fol- lowing its selection. TThe list of stocks making up the volatile group was adjusted at the end of each 6-month period. ‘The list of rapidly moving stocks, selected as deseribed above, was then subjected to a comparison with the results secured employing average stocks for the period under consideration. The average stocks would have yielded an average annual net gain of 8.9 per cent while the volatile stocks would have shown an average net gain of but 5.4 per cent. Though the volatile group manifested an average move consid~ erably more than that of the market, this was more than offset by theALFRED COWLES AND HERBERT F. JONES 289 fact that its ratio of sequences to reversals was less favorable. The sub- stitution of volatile for average stocks was therefore abandoned. In the ease of the stock market averages a study was undertaken to determine the degree of consisteney in the data by considering distribu- tions of the variables over finite periods of time. Figure 3 represents the frequeney distributions of the ratios of sequences to reversals for units of 1 day, 1 week, 1 month, and 3 months. Periods of 12 and 24 units of time were considered in all eases except the last, where insuffi- cient data made it necessary to consider periods of only 4 units. In the upper part of each diagram are the actual histograms, or frequency dis- tributions. A better way to illustrate the probability of variations is by summation, or cumulative frequency distribution. ‘The histograms, therefore, were smoothed, as shown by the solid continuous curve, and ‘the cumulated curves computed from the smoothed distributions. The dotted curves represent the distributions that would be expected from series for which the probability of a sequence or reversal is }. They were ‘computed from the formula Cx a? PS) = where P(S) is the probability of obtaining S sequences in a sample of n observations.® For units of 1 day the periods of 12 units have a distribution very nearly random with a slight tendency towards a skewness to the right. ‘This skewness is more pronounced in the diagram for periods of 24 units, ‘The cumulative frequency curves have a scale on the left giving the probabilities of a smaller ratio, and a scale on the right, the probabili- ties of a larger ratio. From the figure it will be seen that the probability of obtaining a ratio of sequences to reversals less than 1 is about .43. Or again, the probability of obtaining a ratio larger than 2 is about .20 in the ease of periods of 12 days and about .09 for periods of 24 days. From such a diagram it is possible to determine the limiting ratio for any probability. For instance, the limits of a probability band of .50 cean be determined by finding the ratios for which the probabilities are 75 and .25, respectively, of obtaining a larger ratio. In this ease, for + In n observations there are (n—t) first differences. In this set of (m1) first differences there ean be S sequences and (n—2)~S reversals. The number of different orders in which (~2) things can be arranged in two sets, 8 and (n-2)-S, ia (n—2)1/SI(n—2~S)l=,-10s, But a sequence ean ocour when cither « rite follows a rise or a decline follows a decline, The total number of ‘samples containing S sequences, therefore, will be 2-_-1Cs. Since possible samples, the probability of obtaining S sequences will be given by the above equation290 RCONOMETRICA UNITS OF TAY TRS ET WEEK SERS OF STEARATE aTiO OF SEQUENCES To REVERSALS pee fe Bo ‘naYiO OF SEQUENCES 10 REVERSALS TNT OF WONT UNITE OF 3 MONTHS. FROST SRN ATS ants oF eadences 1H AEVERGALS oe a PS El sufi of seubences 10 REVERSHLS Froune 3—Frequeney distributions of ratio of sequences to reversals in direction of stock priee indexes.ALFRED COWLES AND HERBERT B, JONDS 201 periods of 24 units, 0.9 is the lower limiting ratio and 1.4 the upper ratio, The median ratio is 1.1. Therefore, the ratio of sequences to re- versals, for periods of 24 days, is 1.1; and 50 per cent of the time the ratios will lie between .9 and 1.4. ‘The distribution for units of 1 week shows a greater tendeney to- wards structure than in the ease of daily units, In this ease the eumu- lative curve indicates that the probabilities are about 2 to 1 in favor of a ratio greater than unity. The tendeney towards structure is even more Sitges RF eC [tan see iia |i — er | meee E i Fr Cl |: THAN 4 UNIT rf ere ig sh atc a | Ee a 4 28. 5 Bs F ECC ‘gaint bane mencon | lashlte baat w paca Fiovne 4—Frequency distributions of absolute pereentage ‘changes in stock price indexes. pronounced for units of 1 month where, for periods of 12 months, the chanees are about 4 to 1 in favor of a ratio greater than unity. ‘The cumulative curves are relatively steep, also, which indicates a high coneentration around the average. This is especially true for the periods of 24 months. For units of 3 months the distributions are not nearly s0 favorable as in the previous case, They are flatter, indicating large variations, and the average ratio is muzh smaller. Figure 4 shows the results of the same type of analysis applied to the data for the percentage changes. In this ease only three units of time ‘were investigated as the daily units already had proved to be impracti-No 40 surnevaoug ()o—saoxvng aovanaowng aovUaAy Jo sonmaeqoad oar3 9 paavy, ozo 080" onset tee te 090 2100" a0" 5000" soo s8t0 01000 2 wo sees ror asve ooze twsr | aeor | os ost sr00" $100" ss10" asso goto sto" ett" 3 ? avsy Orono orte | tee Oot het eT we g 2 200 soo. sero" 3800" 5800" aso tego 5 5 sos sree eer) OE ws E i 2 oss e000" erro" 8880" ato” apo satu eet a g wor ws a9 wor oe 0 ser- | ort 3 & 2 2 om fro e100" set” _—atso” gogo soso" soto g so- see- ag- geo- oe see- see- | oor § oso | 100" 200" otoo_——onto_—stzso" area” 010" 000 5 wie soe aes ger orst- —ceer- o-oo | ot ozo | to00° 9000 otoo_— seu" ston tama a1 s wore sa'ce- oeor— vee eee bro SOT ATI | of" ro oof 900 oe) OREOALYRED COWLES AND HERBERT E. JONES 293, cable. For units of 3 months the data were insufficient for grouping in periods of more than 4 units. For units of 1 week and 1 month the difference between the periods of 12 units and 24 units is so slight as to be without statistical significance, The next step was to combine these distributions in order to deter- mine the distribution of the expected net gain. To do this mathe ‘matically gives rise to hopeless complications unless certain simplifying assumptions are made which themselves cast doubt on the value of the solutions, Therefore, it was necessary to compute the distributions by empirical means. Juin ean “ens esp | Sumac Ae Prose | 2| TT Feuey ele | E ear He i EAI : YT Te f | Et le =| BAU rele iB E HH 334 be F f ele I PROT ee | PA aa nh Fiaune 6.—Frequeney distributions of expected annual net profits based on ‘method of sequence probabilities. Brokerage computed at I per cent. ‘To determine these empirical distributions, values of R(®) and C(t) were taken which adequately covered their range of variation, and I(d) was computed for all possible combinations of these values. The average brokerage charge was assumed to be 1 per eent per trade. To facilitate computation tables were employed of the type illustrated in ‘Table 4. ‘The roman type figures in the body of the table give the values of 1(@ for the various values of R() in the first column and C(() in the294, ECONOMETRICA first row. The probabilities for each value of R(t) and C(#) are shown in the last column and the bottom row respectively. The probabilities for the various values of I() are the produets of the probabilities of R(O and C(0) and are shown by the figures in italies in the table.* The values of 1(#) have been classified and the sum of the probabilities for each value of Z(t) in any class taken as the probability of that class. ‘The results of this analysis are presented in Figure 5. Here again are shown the histograms and cumulative frequency curves representing units of 1 week, 1 month, and 3 months. For units of 1 week the cumulative curves indicate an average annual loss of about 10 per cent, with but 1 chance in 3 of obtaining a net profit over any period of 24 weeks. The same is true to a lesser extent when we use units of 8 months, In this ease we see there is about an even chance of a loss or gain. For units of 1 month, however, an average net gain of about 7 per cent is indicated. But, even here, no great consistency is evident. In fact the cumulative curve indieates that the chance of loss for any one year is about 1 in 3, Furthermore the results should be interpreted with eaution of the fact that various units of time, other than 1 month, were con- sidered and rejected. In all, 26 such units, ranging from 20 minutes up to 10 years, were examined. The series represented by units of one month, therefore, was selected by hindsight as the most favorable one in 26 trials ‘This type of foreeasting could not be employed by speculators with any assurance of consistent or large profits. On the other hand, the significant exeess of sequenees over reversals for all units from 20 min- utes up to 6 months, with the exception of units of 2 weeks and 3 weeks mentioned previously, represents conclusive evidenee of structure in stock prices. Aurnen Cows 3RD Hernzer E, Jonps Cowles Commission for Research in Economics Colorado Springs, Colorado "The values of R(t) and C(t) for the periods studied are randomly distributed and a correlation analysis between (0) and C(t) gave a coefficient of ~0.2 with ‘standard error of 0.1, This coefficient is on the borderline of significance but is very small, Therefore, the probabilities of (2) may be computed in this manner without great error.